feat(analysis): add model similarity nodes

Separate CWB-derived diagnostics from merge weighting while streaming each logical layer once.

.comfyignore

@@ -0,0 +1 @@
+tests/

__init__.py

@@ -1,7 +1,7 @@
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
 
-from .nodes.lora_merger import LoRAMultiMerge, LoRAMultiMergeDARE
+from .nodes.lora_merger import LoRAMultiMerge, LoRAMultiMergeDARE, LoRAMultiMergeDAREEnhanced
 from .nodes.metakeys import (
     ModelMetaKeys, TextEncoderMetaKeys, LoRAMetaKeys,
     CheckpointMetaKeys, EmbeddingMetaKeys
@@ -20,10 +20,27 @@ from .nodes.merger import (
     ModelThreeMerger, TextEncoderThreeMerger, LoRAThreeMerger,
     CheckpointThreeMerger, EmbeddingThreeMerger
 )
+from .nodes.consensus_merger import CWB_MERGER_NODES
+from .nodes.model_analysis import MODEL_ANALYSIS_NODES
 from .nodes.lora_extract_svd import (
     LoRAExtractFixed, LoRAExtractRatio, LoRAExtractQuantile,
     LoRAExtractKnee, LoRAExtractFrobenius
 )
+from .nodes.dora_extract_wd import (
+    DoRAExtractFixed, DoRAExtractRatio, DoRAExtractQuantile,
+    DoRAExtractKnee, DoRAExtractFrobenius
+)
+from .nodes.dora_learned_wd import (
+    DoRALearnedExtractFixed, DoRALearnedExtractRatio, DoRALearnedExtractQuantile,
+    DoRALearnedExtractKnee, DoRALearnedExtractFrobenius
+)
+from .nodes.text_encoder_extract import (
+    TextEncoderLoRAExtractFixed, TextEncoderLoRAExtractRatio,
+    TextEncoderLoRAExtractQuantile, TextEncoderLoRAExtractKnee,
+    TextEncoderLoRAExtractFrobenius, TextEncoderDoRAExtractFixed,
+    TextEncoderDoRAExtractRatio, TextEncoderDoRAExtractQuantile,
+    TextEncoderDoRAExtractKnee, TextEncoderDoRAExtractFrobenius,
+)
 from .nodes.lora_resize import (
     LoRAResizeFixed, LoRAResizeRatio,
     LoRAResizeFrobenius, LoRAResizeCumulative,
@@ -39,6 +56,8 @@ from .nodes.model_info_nodes import (
 )
 
 
+from .nodes.lora_rename import AnimaLoraRename
+
 class ModelUtilsExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
@@ -58,16 +77,34 @@ class ModelUtilsExtension(ComfyExtension):
             # Three-Model Mergers
             ModelThreeMerger, TextEncoderThreeMerger, LoRAThreeMerger,
             CheckpointThreeMerger, EmbeddingThreeMerger,
+            # Dedicated CWB Mergers
+            *CWB_MERGER_NODES,
+            # Two-Model Analysis
+            *MODEL_ANALYSIS_NODES,
             # LoRA Extraction
             LoRAExtractFixed, LoRAExtractRatio, LoRAExtractQuantile,
             LoRAExtractKnee, LoRAExtractFrobenius,
+            # DoRA Extraction
+            DoRAExtractFixed, DoRAExtractRatio, DoRAExtractQuantile,
+            DoRAExtractKnee, DoRAExtractFrobenius,
+            # Learned DoRA Extraction
+            DoRALearnedExtractFixed, DoRALearnedExtractRatio, DoRALearnedExtractQuantile,
+            DoRALearnedExtractKnee, DoRALearnedExtractFrobenius,
+            # Text Encoder LoRA/DoRA Extraction
+            TextEncoderLoRAExtractFixed, TextEncoderLoRAExtractRatio,
+            TextEncoderLoRAExtractQuantile, TextEncoderLoRAExtractKnee,
+            TextEncoderLoRAExtractFrobenius, TextEncoderDoRAExtractFixed,
+            TextEncoderDoRAExtractRatio, TextEncoderDoRAExtractQuantile,
+            TextEncoderDoRAExtractKnee, TextEncoderDoRAExtractFrobenius,
             # LoRA Resize
             LoRAResizeFixed, LoRAResizeRatio,
             LoRAResizeFrobenius, LoRAResizeCumulative,
             # LoRA Multi-Merge
-            LoRAMultiMerge, LoRAMultiMergeDARE,
+            LoRAMultiMerge, LoRAMultiMergeDARE, LoRAMultiMergeDAREEnhanced,
             # LoRA Merge To Model
             LoRAMergeToModel,
+            # LoRA Utilities
+            AnimaLoraRename,
             # Downloaders
             CheckpointInfoMetaDownloader, DiffusionModelInfoMetaDownloader, LoRAInfoMetaDownloader, EmbeddingInfoMetaDownloader,
             VAEInfoMetaDownloader, ControlNetInfoMetaDownloader, ManualPathInfoMetaDownloader,

docs/consensus_mergers.md

@@ -0,0 +1,111 @@
+# Consensus-Weighted Blending Mergers
+
+The dedicated CWB nodes merge two or three safetensors files using
+Consensus-Weighted Blending. They calculate an element-wise mean or median
+consensus, compare each aligned source vector to that consensus, and normalize
+similarity-derived weights before producing the output vector.
+
+## Node families
+
+Two-input and three-input variants are provided for checkpoints, standalone
+diffusion models, text encoders, LoRAs, and embeddings.
+
+Model A supplies output metadata and anchors shared-layer shape, naming, and
+mismatch behavior. The output key set is the union of all inputs. A valid
+tensor found only in a secondary input is copied unchanged. When Model A lacks
+a tensor that is present in multiple secondary inputs, those available sources
+are CWB-merged without inserting a zero contribution for Model A. Embedding
+mergers additionally use the longest compatible first dimension as their
+alignment reference.
+
+## Presets and custom controls
+
+`baseline` is the default. The standard clarity, smooth, varied, and diversity
+presets are joined by `power_blend` and the DSC presets demonstrated by the CWB
+conditioning implementation. Select `custom` to make every manual CWB control
+authoritative. Every named merger preset keeps `global_scale` at `1.0`; scaling
+persisted model or LoRA weights is only performed when explicitly selected in
+the custom controls.
+
+Similarity alignment greedily matches source rows to the reference by cosine
+similarity. Index alignment keeps absolute row positions. Position weight can
+bias similarity matches toward nearby normalized positions. Preserve Common
+Prefix copies the longest numerically identical leading row span directly from
+the first input.
+
+Inputs have equal prior weight. The similarity calculation supplies the
+effective contribution for each aligned vector. If every candidate is rejected
+or receives zero weight, CWB falls back to equal weighting.
+
+## Missing tensors and filtering
+
+- `skip` preserves Model A when an anchored input is missing or incompatible.
+  In embedding-union mode it omits that source for the affected key.
+- `zeros` supplies a compatible zero tensor for the missing source.
+- `error` stops the merge.
+- Exclude patterns preserve the anchored source tensor.
+- Discard patterns remove the matching output tensor or logical LoRA layer.
+
+The missing-input modes above apply to Model-A-owned layers. Model A being
+absent from a secondary-only layer is not itself a mismatch. A shape conflict
+between multiple available secondary providers does use the selected mismatch
+mode, with the earliest available input as its preservation anchor.
+
+Patterns use regular expressions by default. Enable Glob Patterns to use glob
+syntax instead.
+
+## LoRA behavior
+
+LoRA inputs use the same parser as the resize and multi-merge nodes. It accepts
+all seven low-rank A/B spellings recognized by ComfyUI's `LoRAAdapter`, plus
+direct `.diff`, `.diff_b`, `.w_norm`, `.b_norm`, and `.set_weight` layers.
+
+Recognized diffusion-model outputs use the canonical
+`diffusion_model.<layer>.lora_A.weight` / `.lora_B.weight` convention.
+Secondary formats are matched by normalized logical layer name. Differing
+ranks are zero-padded like the DARE/TIES mergers: A/down tensors on dimension 0
+and B/up tensors on dimension 1. An existing anchor-source alpha key is updated
+to the resulting maximum rank; no alpha key is invented when the anchor has
+none.
+Similarity alignment operates on paired latent-rank components. A rows and B
+columns share one mapping; fixed input/output feature axes are never reordered.
+Input alpha scaling is absorbed into the B factor before blending, and global
+scale is applied once to the resulting pair rather than once per factor.
+Incomplete pairs and unrecognized tensors are copied from the earliest
+available source unchanged.
+Companion-bearing groups (`lora_mid`, reshape, DoRA, or set-weight) are kept
+atomically from their earliest available source rather than partially
+transformed. Secondary-only logical groups are included automatically: a group
+from one source is copied unchanged, while compatible groups from multiple
+secondary sources are merged using only those sources. Recognized roles are
+written with the same canonical output convention as shared layers.
+
+`include_1d_diffs` is disabled by default. Disabled shared 1D direct layers
+preserve Model A, while secondary-only layers preserve their earliest source.
+Enabled 1D direct layers participate as complete tensors and are always saved
+in FP32.
+
+## Precision and quantized inputs
+
+CWB computes in FP32. Unless Override Dtype is enabled, any participating FP32
+source keeps the generated logical layer in FP32. Guarded low-bit tensors and
+enabled 1D direct diffs are never downcast by the override.
+
+ComfyUI/comfy-kitchen quantized inputs are rejected before an output writer is
+opened. Files with three or more bare INT8, UINT8, FP8, or FP4 tensors are also
+rejected. One or two isolated low-bit tensors produce one warning per input and
+are preserved without CWB arithmetic; an affected LoRA causes its complete
+logical layer to be preserved.
+
+## Streaming and failure safety
+
+Model parameters and direct patches use fixed-coordinate blending; similarity
+row matching is limited to embeddings and LoRA rank components. Each logical
+layer is loaded, transferred to the processing device, completed, copied back
+to CPU, queued for writing, and released before the next layer. Outputs are
+written to a temporary sibling and atomically published only after every layer
+and asynchronous write succeeds. Failed merges do not replace or finalize the
+requested output. A CUDA OOM retries only the affected layer on CPU after
+releasing the failed GPU working set; failure of that retry aborts atomically.
+Generated files are published under the matching category inside
+`folder_paths.models_dir`, never ComfyUI's general output directory.

docs/merger_2_model_modes.md

@@ -33,6 +33,30 @@
 
 ---
 
+## Power-Up (DARE+TIES)
+> Combines DARE (Drop and Rescale) with TIES (Trim, Elect Sign) — the current best-practice algorithm for merging fine-tuned models, particularly LoRA-tuned ones. DARE sparsifies the task vector via random dropout and rescaling; TIES then removes residual low-magnitude noise and enforces sign consistency. The result is a cleaner, more targeted capability transfer from B to A with less interference.
+
+**Models Used:** A, B
+**Parameters:**
+- **Alpha:** DARE drop rate. Fraction of delta parameters randomly zeroed. Higher values produce a sparser, more targeted delta. Typical range: `0.5`–`0.9`.
+- **Beta:** TIES trim quantile. Fraction of the smallest-magnitude delta parameters zeroed after DARE, removing residual noise. `0.0` disables trimming. Typical range: `0.1`–`0.3`.
+- **Gamma:** Lambda scale. Final multiplier applied to the merged delta before adding to Model A. Equivalent to the task arithmetic scaling coefficient. Typical range: `0.5`–`1.5`.
+- **Seed:** Random seed for the DARE dropout mask. Controls reproducibility.
+
+**Algorithm:**
+1. Compute task vector `δ = B − A`
+2. **DARE:** apply random binary mask with keep probability `1 − alpha`; rescale survivors by `1/(1−alpha)`
+3. **TIES trim:** zero out parameters below the `beta`-quantile magnitude threshold
+4. **TIES elect:** determine dominant sign per position; zero out disagreeing parameters
+5. Return `A + gamma × δ_filtered`
+
+**Guidance:**
+- Start with `alpha=0.9, beta=0.2, gamma=1.0` for LoRA-derived models
+- Lower `alpha` (e.g. `0.5`) for full fine-tunes where more delta mass should be retained
+- Increase `gamma` beyond `1.0` to amplify B's influence; decrease below `1.0` to soften it
+
+---
+
 ## Enhanced Man Interp
 > Sophisticated interpolation between values from A and B depending on their difference relative to other values, with manual threshold control.
 
@@ -97,6 +121,15 @@ When merging models with different layer structures (e.g., fine-tuned models wit
 
 ---
 
+## Dtype Preservation & Override
+
+By default, the merger preserves the highest precision `dtype` for each key found across all active source models (Model A and Model B) and the requested `save_dtype`.
+
+- **Preserve Higher Precision**: If a key is `float32` in any of the source models, but you requested `save_dtype="bf16"` or `"fp16"`, that key will be saved as `float32` anyway. This prevents degrading critical high-precision layers (such as normalization layers or layers with restored outliers).
+- **Explicit Override**: If you prefer to force the entire model to be saved strictly as the selected `save_dtype` (e.g. for complete downcasting to `bf16`), toggle `override_dtype` to `True`.
+
+---
+
 ## Layer Filtering (Regex Patterns)
 
 Use regex patterns to control which layers are merged:

docs/merger_3_model_modes.md

@@ -67,6 +67,15 @@ When merging models with different layer structures, the `mismatch_mode` paramet
 
 ---
 
+## Dtype Preservation & Override
+
+By default, the merger preserves the highest precision `dtype` for each key found across all active source models (Model A, B, and C) and the requested `save_dtype`.
+
+- **Preserve Higher Precision**: If a key is `float32` in any of the active source models, but you requested `save_dtype="bf16"` or `"fp16"`, that key will be saved as `float32` anyway. This prevents degrading critical high-precision layers (such as normalization layers or layers with restored outliers).
+- **Explicit Override**: If you prefer to force the entire model to be saved strictly as the selected `save_dtype` (e.g. for complete downcasting to `bf16`), toggle `override_dtype` to `True`.
+
+---
+
 ## Layer Filtering (Regex Patterns)
 
 Use regex patterns to control which layers are merged:

docs/model_analysis.md

@@ -0,0 +1,75 @@
+# Two-Model Similarity Analysis
+
+These nodes compare two safetensors models without merging them or writing an
+output file. Separate nodes are provided for checkpoints, diffusion models,
+text encoders, LoRAs, and embeddings.
+
+## Outputs
+
+`comparison_report` contains conventional numerical differences and topology
+findings. `cwb_report` separately contains diagnostics derived from the
+similarity, alignment, and consensus stages of Consensus-Weighted Blending.
+`documentation` contains this reference.
+
+The reports begin with a global summary, followed by inferred blockwise values
+and one entry for every comparable tensor. Block names are approximated from
+the tensor-key hierarchy through its first numeric path component; they are not
+an architecture-specific interpretation.
+
+## Standard metrics
+
+- MAE, MSE, and RMSE describe absolute error at the original weight scale.
+- Maximum absolute difference identifies the largest individual change.
+- Relative L2 is `||A-B|| / ((||A||+||B||)/2)` and is scale-aware.
+- Cosine similarity compares direction; Pearson correlation compares centered
+  variation.
+- Exact equality, sign agreement, norm ratio, and finite-value coverage expose
+  differences hidden by averages.
+
+The global section reports both parameter-weighted metrics and equal-layer
+averages. No arbitrary combined similarity score is assigned. Undefined
+zero-norm or constant-tensor statistics are shown as `N/A`.
+
+Only shared, same-shape floating tensors enter numerical aggregates. Missing
+keys and shape mismatches are listed separately and are never padded, cropped,
+or replaced with zeros. Non-floating tensors receive an exact-equality check.
+Quantized models carrying ComfyUI quantization metadata are rejected because
+encoded storage values do not represent directly comparable model weights.
+
+## CWB diagnostics
+
+The CWB report provides pairwise row-cosine statistics and each input's affinity
+to both mean and median consensus. It stops before CWB's merge-weighting stage:
+no thresholds, contribution weights, diversity weighting, vector
+reconstitution, norm rescaling, or output writer are used.
+
+Checkpoint, diffusion-model, and text-encoder tensors always retain their fixed
+coordinate alignment. Embedding and LoRA nodes expose one optional
+`cwb_similarity_alignment` control. When enabled, it performs
+threshold-independent greedy one-to-one matching derived from CWB and reports
+coverage, matched-score statistics, and improvement over index alignment. This
+operation is quadratic and may be slow for large embeddings.
+
+LoRA down/up tensors are treated as one logical layer. Rank-component matching
+uses the product of down and up cosine similarity, including CWB's paired-sign
+equivalence for comparison. No source tensor is modified.
+
+## Memory and failure handling
+
+Inputs always use Unified Efficient Loader's low-memory streaming mode. Matching
+A/B tensors are loaded concurrently through bounded `async_stream()` pipelines
+with `batch_size=1` and `prefetch_batches=1`, but no subsequent logical layer is
+requested early. A complete LoRA down/up pair is loaded once and used for both
+reports. CUDA inputs use pinned transfer while retaining the current CPU source
+tensors for fallback.
+
+Calculations use FP32. When CUDA processing runs out of memory, CUDA memory is
+cleared and the current logical layer is retried once on CPU. After success or
+failure, all tensor references for that layer are released and every loaded key
+is marked processed exactly once before advancing. Both reports list every CPU
+fallback. A failed CPU retry or any non-OOM error stops analysis with the layer
+name in the error.
+
+`top_weight_differences` bounds the scalar-coordinate list; every comparable
+tensor is still present in the layerwise report. `DOCUMENTATION ONLY` loads no
+model data.

nodes/dora_extract_wd.py

@@ -0,0 +1,989 @@
+"""
+DoRA Extraction via SVD and Weight Decomposition.
+
+Native implementation of Weight-Decomposed Low-Rank Adapter extraction
+with multiple rank selection modes. Extracted models use the DoRA structure.
+"""
+import fnmatch
+import os
+import re
+import torch
+import torch.linalg as linalg
+import folder_paths
+import comfy.utils
+from tqdm import tqdm
+from comfy_api.latest import io
+from .device_utils import (
+    estimate_model_size, prepare_for_large_operation,
+    cleanup_after_operation, get_device_capabilities
+)
+
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned, IncrementalSafetensorsWriter
+
+
+
+
+# =============================================================================
+# Rank Index Functions - Determine LoRA rank from singular values
+# =============================================================================
+
+def _index_sv_fixed(S: torch.Tensor, dim: int) -> int:
+    """Fixed rank mode - use specified dimension."""
+    return max(1, min(dim, len(S) - 1))
+
+
+def _index_sv_threshold(S: torch.Tensor, threshold: float) -> int:
+    """Threshold mode - keep singular values above absolute threshold."""
+    rank = int(torch.sum(S > threshold).item())
+    return max(1, min(rank, len(S) - 1))
+
+
+def _index_sv_ratio(S: torch.Tensor, ratio: float) -> int:
+    """Ratio mode - keep singular values > max(S) / ratio (kohya convention)."""
+    if ratio <= 0:
+        return len(S) - 1
+    min_sv = S[0] / ratio
+    rank = int(torch.sum(S > min_sv).item())
+    return max(1, min(rank, len(S) - 1))
+
+
+def _index_sv_cumulative(S: torch.Tensor, target: float, max_rank: int = None) -> int:
+    """Cumulative mode - keep enough SVs to reach target % of total.
+
+    Calculates relative to max_rank if provided, otherwise relative to full.
+    """
+    if max_rank is not None and max_rank < len(S):
+        total = torch.sum(S[:max_rank])
+    else:
+        total = torch.sum(S)
+
+    if total < 1e-8:
+        return 1
+    cumsum = torch.cumsum(S, dim=0) / total
+    rank = int(torch.searchsorted(cumsum, target).item()) + 1
+    return max(1, min(rank, len(S) - 1))
+
+
+def _index_sv_fro(S: torch.Tensor, target: float, max_rank: int = None) -> int:
+    """Frobenius norm mode - preserve target fraction of Frobenius norm.
+
+    Calculates relative to max_rank if provided, otherwise relative to full.
+    This means "retain target% of what's achievable within max_rank".
+    """
+    if max_rank is not None and max_rank < len(S):
+        # Calculate relative to what's achievable within max_rank
+        S_capped = S[:max_rank]
+        S_sq = S_capped.pow(2)
+        total_sq = torch.sum(S_sq)
+    else:
+        S_sq = S.pow(2)
+        total_sq = torch.sum(S_sq)
+
+    if total_sq < 1e-8:
+        return 1
+
+    # Cumsum of all S (not capped) to find where we reach target
+    cumsum = torch.cumsum(S.pow(2), dim=0) / total_sq
+    rank = int(torch.searchsorted(cumsum, target ** 2).item()) + 1
+    return max(1, min(rank, len(S) - 1))
+
+
+def _index_sv_knee(S: torch.Tensor, min_sv: float = 1e-8) -> int:
+    """Knee detection - find elbow point on singular value curve."""
+    n = len(S)
+    if n < 3:
+        return 1
+    s_max, s_min = S[0].item(), S[-1].item()
+    if s_max - s_min < min_sv:
+        return 1
+    # Normalize to [0, 1]
+    s_norm = (S - s_min) / (s_max - s_min)
+    x_norm = torch.linspace(0, 1, n, device=S.device, dtype=S.dtype)
+    # Distance from diagonal
+    distances = (x_norm + s_norm - 1).abs()
+    rank = torch.argmax(distances).item() + 1
+    return max(1, min(rank, n - 1))
+
+
+def _index_sv_cumulative_knee(S: torch.Tensor, min_sv: float = 1e-8) -> int:
+    """Knee detection on cumulative singular value curve."""
+    n = len(S)
+    if n < 3:
+        return 1
+    total = torch.sum(S)
+    if total < min_sv:
+        return 1
+    cumsum = torch.cumsum(S, dim=0) / total
+    y_min, y_max = cumsum[0].item(), cumsum[-1].item()
+    if y_max - y_min < min_sv:
+        return 1
+    y_norm = (cumsum - y_min) / (y_max - y_min)
+    x_norm = torch.linspace(0, 1, n, device=S.device, dtype=S.dtype)
+    distances = (y_norm - x_norm).abs()
+    rank = torch.argmax(distances).item() + 1
+    return max(1, min(rank, n - 1))
+
+
+def _index_sv_rel_decrease(S: torch.Tensor, tau: float = 0.1) -> int:
+    """Relative decrease mode - stop when consecutive SV ratio drops below tau."""
+    if len(S) < 2:
+        return 1
+    ratios = S[1:] / (S[:-1] + 1e-8)
+    for k in range(len(ratios)):
+        if ratios[k] < tau:
+            return k + 1
+    return len(S)
+
+
+def _compute_rank(S: torch.Tensor, mode: str, mode_param: float,
+                  max_rank: int = None) -> int:
+    """Compute rank based on mode and parameters."""
+    if mode == "fixed":
+        rank = _index_sv_fixed(S, int(mode_param))
+    elif mode == "threshold":
+        rank = _index_sv_threshold(S, mode_param)
+    elif mode == "ratio":
+        rank = _index_sv_ratio(S, mode_param)
+    elif mode == "quantile" or mode == "sv_cumulative":
+        rank = _index_sv_cumulative(S, mode_param, max_rank)
+    elif mode == "sv_fro":
+        rank = _index_sv_fro(S, mode_param, max_rank)
+    elif mode == "sv_knee":
+        rank = _index_sv_knee(S)
+    elif mode == "sv_cumulative_knee":
+        rank = _index_sv_cumulative_knee(S)
+    elif mode == "sv_rel_decrease":
+        rank = _index_sv_rel_decrease(S, mode_param)
+    else:
+        rank = _index_sv_fixed(S, int(mode_param) if mode_param else 64)
+
+    if max_rank is not None:
+        rank = min(rank, max_rank)
+    return rank
+
+
+# =============================================================================
+# SVD Extraction Functions
+# =============================================================================
+
+def _svd_extract_linear_lowrank(
+    weight: torch.Tensor,
+    rank: int,
+    device: str,
+    clamp_quantile: float = 0.99,
+    niter: int = 2,
+) -> tuple:
+    """
+    Low-rank SVD decomposition using svd_lowrank for fixed rank extraction.
+
+    Much faster than full SVD when rank << min(m, n) because it only computes
+    the top-k singular values using randomized algorithms.
+
+    Args:
+        niter: Power iterations for SVD accuracy (higher = more accurate but slower)
+
+    Returns:
+        (lora_down, lora_up, diff), "low rank" or (weight, "full")
+    """
+    weight = weight.to(device=device, dtype=torch.float32)
+    out_dim, in_dim = weight.shape
+
+    # Clamp rank to valid range
+    max_possible_rank = min(out_dim, in_dim)
+    rank = max(1, min(rank, max_possible_rank))
+
+    # Check if decomposition is worthwhile
+    if rank >= max_possible_rank:
+        return weight, "full"
+
+    # Use svd_lowrank for faster low-rank approximation
+    try:
+        U, S, V = torch.svd_lowrank(weight, q=rank, niter=niter)
+        # U: [out_dim, rank], S: [rank], V: [in_dim, rank]
+        Vh = V.T  # [rank, in_dim]
+    except Exception as e:
+        raise RuntimeError(f"svd_lowrank failed: {e}")
+
+    # Clamp outliers
+    if clamp_quantile < 1.0 and len(S) > 0:
+        try:
+            max_val = torch.quantile(S, clamp_quantile)
+            S = S.clamp(max=max_val)
+        except RuntimeError:
+            pass
+
+    # Construct LoRA matrices
+    lora_up = U @ torch.diag(S)  # [out_dim, rank]
+    lora_down = Vh               # [rank, in_dim]
+
+    # Compute reconstruction diff
+    diff = weight - (lora_up @ lora_down)
+
+    return (lora_down, lora_up, diff), "low rank"
+
+
+def _svd_extract_linear(
+    weight: torch.Tensor,
+    mode: str,
+    mode_param: float,
+    device: str,
+    max_rank: int = None,
+    clamp_quantile: float = 0.99,
+    niter: int = 2,
+) -> tuple:
+    """
+    SVD decomposition for linear layers.
+
+    For fixed rank mode, uses svd_lowrank which is much faster.
+    For adaptive modes, uses full SVD to analyze all singular values.
+
+    Returns:
+        (lora_down, lora_up, rank), "low rank" or (weight, "full")
+    """
+    weight = weight.to(device=device, dtype=torch.float32)
+    out_dim, in_dim = weight.shape
+
+    # For fixed rank, use the faster svd_lowrank
+    if mode == "fixed" and mode_param > 0:
+        target_rank = int(mode_param)
+        if max_rank is not None:
+            target_rank = min(target_rank, max_rank)
+        return _svd_extract_linear_lowrank(weight, target_rank, device, clamp_quantile, niter)
+
+    # For adaptive modes, use full SVD to analyze singular values
+    try:
+        U, S, Vh = linalg.svd(weight, full_matrices=False)
+    except Exception as e:
+        raise RuntimeError(f"SVD failed: {e}")
+
+    # Compute rank based on mode
+    rank = _compute_rank(S, mode, mode_param, max_rank)
+
+    # Check if decomposition is worthwhile
+    if rank >= min(out_dim, in_dim):
+        return weight, "full"
+
+    # Truncate to rank
+    U = U[:, :rank]
+    S = S[:rank]
+    Vh = Vh[:rank, :]
+
+    # Clamp outliers
+    if clamp_quantile < 1.0 and len(S) > 0:
+        try:
+            max_val = torch.quantile(S, clamp_quantile)
+            S = S.clamp(max=max_val)
+        except RuntimeError:
+            pass  # Skip clamping if quantile fails
+
+    # Construct LoRA matrices
+    # lora_up = U @ diag(S), lora_down = Vh
+    lora_up = U @ torch.diag(S)  # [out_dim, rank]
+    lora_down = Vh               # [rank, in_dim]
+
+    # Compute reconstruction diff for optional sparse bias
+    diff = weight - (lora_up @ lora_down)
+
+    return (lora_down, lora_up, diff), "low rank"
+
+
+def _svd_extract_conv(
+    weight: torch.Tensor,
+    mode: str,
+    mode_param: float,
+    device: str,
+    max_rank: int = None,
+    clamp_quantile: float = 0.99,
+) -> tuple:
+    """
+    SVD decomposition for conv2d layers.
+
+    Returns:
+        (lora_down, lora_up, rank), "low rank" or (weight, "full")
+    """
+    out_ch, in_ch, kh, kw = weight.shape
+    is_1x1 = (kh == 1 and kw == 1)
+
+    # Flatten for SVD
+    if is_1x1:
+        mat = weight.view(out_ch, in_ch)  # [out_ch, in_ch]
+    else:
+        mat = weight.reshape(out_ch, -1)  # [out_ch, in_ch*k*k]
+
+    result, mode_str = _svd_extract_linear(mat, mode, mode_param, device, max_rank, clamp_quantile)
+
+    if mode_str == "full":
+        return weight, "full"
+
+    lora_down, lora_up, diff = result
+    rank = lora_down.shape[0]
+
+    # Reshape for conv
+    if is_1x1:
+        lora_down = lora_down.unsqueeze(-1).unsqueeze(-1)  # [rank, in_ch, 1, 1]
+        lora_up = lora_up.unsqueeze(-1).unsqueeze(-1)      # [out_ch, rank, 1, 1]
+    else:
+        lora_down = lora_down.reshape(rank, in_ch, kh, kw)
+        lora_up = lora_up.reshape(out_ch, rank, 1, 1)
+
+    return (lora_down, lora_up, diff), "low rank"
+
+
+# =============================================================================
+# Chunked Extraction for Large Layers
+# =============================================================================
+
+def _detect_fused_layer(key: str, shape: tuple) -> int:
+    """Detect fused QKV/MLP layers and return number of chunks."""
+    if len(shape) != 2:
+        return 1
+
+    out_dim, in_dim = shape
+    key_lower = key.lower()
+
+    # QKV layers (3 chunks)
+    if "qkv" in key_lower and out_dim % 3 == 0 and out_dim // 3 >= in_dim // 4:
+        return 3
+
+    # KV layers (2 chunks)
+    if "kv" in key_lower and "qkv" not in key_lower and out_dim % 2 == 0:
+        return 2
+
+    # Fused MLP (2 chunks)
+    if ("linear1" in key_lower or "gate_up" in key_lower) and out_dim % 2 == 0 and out_dim >= in_dim * 2:
+        return 2
+
+    return 1
+
+
+def _extract_chunked_layer(
+    weight_diff: torch.Tensor,
+    num_chunks: int,
+    mode: str,
+    mode_param: float,
+    device: str,
+    max_rank: int = None,
+) -> tuple:
+    """Extract LoRA from fused layer by chunking."""
+    out_dim, in_dim = weight_diff.shape
+    chunk_size = out_dim // num_chunks
+
+    all_lora_up = []
+    all_lora_down = []
+
+    for i in range(num_chunks):
+        chunk = weight_diff[i * chunk_size:(i + 1) * chunk_size, :]
+
+        try:
+            result, mode_str = _svd_extract_linear(chunk, mode, mode_param, device, max_rank)
+            if mode_str == "full":
+                return None, None, 0
+            lora_down, lora_up, _ = result
+            all_lora_down.append(lora_down)
+            all_lora_up.append(lora_up)
+        except Exception:
+            return None, None, 0
+
+    # Combine chunks
+    # Note: Using the first chunk's down matrix as the basis for all chunks
+    # is a better approximation than a simple mean if the chunks share the same input space (like QKV).
+    # Ideally, we would use a more sophisticated joint SVD, but this is a reasonable fallback.
+    combined_lora_up = torch.cat(all_lora_up, dim=0)
+    combined_lora_down = all_lora_down[0]
+    rank = combined_lora_down.shape[0]
+
+    return combined_lora_up, combined_lora_down, rank
+
+
+# =============================================================================
+# Pattern Matching
+# =============================================================================
+
+def _compile_patterns(pattern_string: str, glob_mode: bool = False) -> list:
+    """Compile patterns from whitespace-separated string.
+
+    In regex mode (default) returns compiled re.Pattern objects.
+    In glob mode returns plain strings; fnmatch handles matching.
+    """
+    if not pattern_string or not pattern_string.strip():
+        return []
+    patterns = []
+    for p in pattern_string.split():
+        p = p.strip()
+        if not p:
+            continue
+        if glob_mode:
+            patterns.append(p)
+        else:
+            try:
+                patterns.append(re.compile(p))
+            except re.error:
+                pass
+    return patterns
+
+
+def _matches_any_pattern(key: str, patterns: list, glob_mode: bool = False) -> bool:
+    """Check if key matches any pattern.
+
+    Glob mode: each pattern is matched as a substring glob using fnmatch
+    (dots are literal, * matches any sequence of characters).
+    Regex mode: each pattern is a compiled re.Pattern, matched via search().
+    """
+    if glob_mode:
+        return any(fnmatch.fnmatch(key, f"*{p}*") for p in patterns)
+    return any(p.search(key) for p in patterns)
+
+
+# =============================================================================
+# Main Extraction Function
+# =============================================================================
+
+def extract_dora_from_files(
+    model_a_path: str,
+    model_b_path: str,
+    mode: str,
+    linear_param: float,
+    conv_param: float,
+    device: str,
+    save_dtype: str,
+    output_path: str,
+    linear_max_rank: int = None,
+    conv_max_rank: int = None,
+    clamp_quantile: float = 0.99,
+    min_diff: float = 0.0,
+    skip_patterns_str: str = "",
+    mismatch_mode: str = "skip",
+    chunk_large_layers: bool = True,
+    svd_niter: int = 2,
+    lazy_load: bool = True,
+    force_clear_cache: bool = True,
+    glob_skip_patterns: bool = False,
+) -> None:
+    """
+    Extract LoRA from difference between two models, writing incrementally to disk.
+
+    Args:
+        model_a_path: Finetuned model path
+        model_b_path: Base model path
+        mode: Rank selection mode
+        linear_param: Mode parameter for linear layers
+        conv_param: Mode parameter for conv layers
+        device: Computation device
+        save_dtype: Output dtype
+        output_path: Full path to output safetensors file
+        linear_max_rank: Max rank for linear layers
+        conv_max_rank: Max rank for conv layers
+        clamp_quantile: Quantile for weight clamping
+        min_diff: Minimum difference threshold
+        skip_patterns_str: Patterns for layers to skip
+        mismatch_mode: How to handle mismatches
+        chunk_large_layers: Enable chunked extraction
+        glob_skip_patterns: When True, treat skip_patterns as glob (* wildcard).
+                            When False (default), treat as Python regex.
+    """
+    save_torch_dtype = {
+        "fp32": torch.float32,
+        "fp16": torch.float16,
+        "bf16": torch.bfloat16,
+    }.get(save_dtype, torch.float16)
+
+    skip_patterns = _compile_patterns(skip_patterns_str, glob_mode=glob_skip_patterns)
+
+    # Prepare memory before heavy operation
+    total_size_gb = estimate_model_size(model_a_path) + estimate_model_size(model_b_path)
+    print(f"[DoRA Extract] Preparing memory for {total_size_gb:.2f}GB operation...")
+    prepare_for_large_operation(total_size_gb * 1.5, torch.device(device))  # 1.5x for SVD headroom
+
+    handler_a = MemoryEfficientSafeOpen(model_a_path, low_memory=lazy_load)
+    handler_b = MemoryEfficientSafeOpen(model_b_path, low_memory=lazy_load)
+
+    try:
+        keys_a = set(handler_a.keys())
+        keys_b = set(handler_b.keys())
+        weight_keys = [k for k in keys_a if k.endswith(".weight")]
+        pbar = comfy.utils.ProgressBar(len(weight_keys))
+        stats = {"extracted": 0, "full": 0, "skipped": 0, "chunked": 0}
+
+        def _process_layer(key):
+            lora_name = _format_lora_key(key)
+
+            if _matches_any_pattern(key, skip_patterns, glob_mode=glob_skip_patterns):
+                return "skipped", None
+
+            # Load tensors with pinned memory for CUDA
+            use_pinned = device == 'cuda'
+
+            if key not in keys_b:
+                if mismatch_mode == "skip":
+                    return "skipped", None
+                if mismatch_mode == "error":
+                    raise ValueError(f"Key {key} not found in model B")
+
+                cpu_a = handler_a.get_tensor(key)
+                if use_pinned:
+                    weight_diff = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
+                else:
+                    weight_diff = cpu_a.to(device=device, dtype=torch.float32)
+                del cpu_a
+
+                # Without base, DoRA fallback just treats tensor_a as diff
+                dora_scale = None
+            else:
+                cpu_a = handler_a.get_tensor(key)
+                cpu_b = handler_b.get_tensor(key)
+
+                if use_pinned:
+                    tensor_a = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
+                    tensor_b = transfer_to_gpu_pinned(cpu_b, device, torch.float32)
+                else:
+                    tensor_a = cpu_a.to(device=device, dtype=torch.float32)
+                    tensor_b = cpu_b.to(device=device, dtype=torch.float32)
+                del cpu_a, cpu_b
+
+                if tensor_a.shape != tensor_b.shape:
+                    if mismatch_mode == "skip":
+                        del tensor_a, tensor_b
+                        return "skipped", None
+                    if mismatch_mode == "error":
+                        raise ValueError(f"Shape mismatch for {key}: {tensor_a.shape} vs {tensor_b.shape}")
+
+                    weight_diff = tensor_a
+                    dora_scale = None
+                    del tensor_b
+                else:
+                    # 1-D weights (e.g. x_pad_token, norm .weight) have no
+                    # weight-decomposition direction and are skipped below
+                    # anyway; bail out before the norm, which would index
+                    # dim=1 and raise on a 1-D tensor.
+                    if tensor_a.ndim < 2:
+                        del tensor_a, tensor_b
+                        return "skipped", None
+
+                    # Weight Decomposition calculation
+                    is_conv_dim = tensor_a.ndim == 4
+                    
+                    if is_conv_dim:
+                        out_ch = tensor_a.shape[0]
+                        flat_a = tensor_a.reshape(out_ch, -1)
+                        flat_b = tensor_b.reshape(out_ch, -1)
+                        norm_a = torch.linalg.norm(flat_a, dim=1, keepdim=True)
+                        norm_b = torch.linalg.norm(flat_b, dim=1, keepdim=True)
+                    else:
+                        norm_a = torch.linalg.norm(tensor_a, dim=1, keepdim=True)
+                        norm_b = torch.linalg.norm(tensor_b, dim=1, keepdim=True)
+                    
+                    # Compute directional diff: W_dir = W_ft * (norm_base / norm_ft)
+                    # Use a small epsilon to avoid division by zero
+                    eps = 1e-8
+                    norm_ratio = norm_b / (norm_a + eps)
+                    if is_conv_dim:
+                        norm_ratio = norm_ratio.view(-1, 1, 1, 1)
+                        dora_scale = norm_a.view(-1, 1, 1, 1)
+                    else:
+                        dora_scale = norm_a
+                        
+                    w_dir = tensor_a * norm_ratio
+                    weight_diff = w_dir - tensor_b
+                    del tensor_a, tensor_b, w_dir
+
+            # Skip small differences
+            if min_diff > 0 and weight_diff.abs().max() < min_diff:
+                del weight_diff
+                return "skipped", None
+
+            # Skip 1D tensors
+            if weight_diff.ndim < 2:
+                del weight_diff
+                return "skipped", None
+
+            is_conv = weight_diff.ndim == 4
+            layer_results = {}
+
+            try:
+                if is_conv:
+                    result, mode_str = _svd_extract_conv(
+                        weight_diff, mode, conv_param, device, conv_max_rank, clamp_quantile
+                    )
+                else:
+                    result, mode_str = _svd_extract_linear(
+                        weight_diff, mode, linear_param, device, linear_max_rank, clamp_quantile, svd_niter
+                    )
+            except Exception as e:
+                # Try chunked extraction for large tensors
+                if chunk_large_layers and not is_conv:
+                    num_chunks = _detect_fused_layer(key, weight_diff.shape)
+                    if num_chunks > 1:
+                        print(f"[DoRA Extract] Chunked: {key} ({num_chunks} chunks)")
+                        lora_up, lora_down, rank = _extract_chunked_layer(
+                            weight_diff, num_chunks, mode, linear_param, device, linear_max_rank
+                        )
+                        if lora_up is not None:
+                            layer_results[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.alpha"] = torch.tensor(rank, dtype=save_torch_dtype)
+                            if dora_scale is not None:
+                                layer_results[f"{lora_name}.dora_scale"] = dora_scale.to(save_torch_dtype).cpu().contiguous()
+                            del weight_diff
+                            return "chunked", layer_results
+
+                print(f"[DoRA Extract] Failed: {key}: {e}")
+                del weight_diff
+                return "skipped", None
+
+            # Store result
+            if mode_str == "full":
+                layer_results[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
+                status = "full"
+            else:
+                lora_down, lora_up, rank_or_diff = result
+                # rank is the size of lora_down's first dimension
+                rank = lora_down.shape[0]
+                layer_results[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.alpha"] = torch.tensor(rank, dtype=save_torch_dtype)
+                if dora_scale is not None:
+                    layer_results[f"{lora_name}.dora_scale"] = dora_scale.to(save_torch_dtype).cpu().contiguous()
+                status = "extracted"
+
+            del weight_diff
+            return status, layer_results
+
+        writer = IncrementalSafetensorsWriter(output_path)
+        writer.__enter__()
+        try:
+            for key in tqdm(weight_keys, desc="Extracting DoRA", unit="layers"):
+                status, layer_sd = _process_layer(key)
+                stats[status] += 1
+                if layer_sd:
+                    writer.write_dict(layer_sd)
+
+                if force_clear_cache:
+                    import gc
+                    gc.collect()
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+
+                pbar.update(1)
+        finally:
+            writer.__exit__(None, None, None)
+
+        print(f"[DoRA Extract] Done: {stats['extracted']} extracted, {stats['chunked']} chunked, "
+              f"{stats['full']} full, {stats['skipped']} skipped")
+
+    finally:
+        handler_a.__exit__(None, None, None)
+        handler_b.__exit__(None, None, None)
+        cleanup_after_operation()
+
+
+def _build_lora_output_path(output_filename: str) -> str:
+    """Build output path for LoRA file."""
+    output_dir = folder_paths.get_folder_paths("loras")[0]
+    os.makedirs(output_dir, exist_ok=True)
+    return os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
+
+
+# =============================================================================
+# Node Definitions
+# =============================================================================
+
+def _get_model_inputs():
+    return [
+        io.Combo.Input("model_a", options=folder_paths.get_filename_list("diffusion_models"),
+                      tooltip="Finetuned model (A - B = LoRA)"),
+        io.Combo.Input("model_b", options=folder_paths.get_filename_list("diffusion_models"),
+                      tooltip="Base model (A - B = LoRA)"),
+    ]
+
+
+def _format_lora_key(key: str) -> str:
+    """
+    Format the key for LoRA saving.
+    Always standardizes to 'diffusion_model.' prefix for maximum compatibility.
+
+    ComfyUI maps 'diffusion_model.<layer>' generically for all model types, making
+    it universally compatible. The 'transformer.' prefix only works for a subset of
+    model types via model-specific mapping code and is intentionally not used here.
+    """
+    if key.endswith(".weight"):
+        key = key[:-7]
+
+    # Handle ComfyUI Checkpoint wrapper (ldm/sgm format: model.diffusion_model.*)
+    if key.startswith("model.diffusion_model."):
+        inner_key = key[22:]  # strip len("model.diffusion_model.")
+        return f"diffusion_model.{inner_key}"
+
+    # Handle net.* wrapper
+    if key.startswith("net."):
+        inner_key = key[4:]  # strip len("net.")
+        return f"diffusion_model.{inner_key}"
+
+    # Handle direct Diffusers keys (transformer_blocks.* / single_transformer_blocks.* without prefix)
+    if key.startswith("transformer_blocks") or key.startswith("single_transformer_blocks"):
+        return f"diffusion_model.{key}"
+
+    # Handle legacy lora_unet_ prefix (for resizing without base)
+    if key.startswith("lora_unet_"):
+        return key
+
+    # Handle already-prefixed keys
+    if key.startswith("diffusion_model."):
+        return key
+
+    # Absorb now-invalid transformer. prefix into diffusion_model. for compatibility
+    if key.startswith("transformer."):
+        return f"diffusion_model.{key[12:]}"
+
+    # Handle known Diffusers UNet blocks
+    if any(key.startswith(p) for p in ["down_blocks", "up_blocks", "mid_block", "conv_in", "conv_out", "time_embedding", "class_embedding"]):
+        return f"diffusion_model.{key}"
+
+    # Default fallback
+    return f"diffusion_model.{key}"
+
+
+def _get_common_inputs():
+    return [
+        io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+        io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+        io.Boolean.Input("chunk_large_layers", default=False,
+                        tooltip="Split large fused layers (QKV, MLP) into chunks"),
+        io.Float.Input("clamp_quantile", default=0.99, min=0.5, max=1.0, step=0.01,
+                      tooltip="Clamp outlier singular values"),
+        io.Float.Input("min_diff", default=0.0, min=0.0, max=1.0, step=0.001,
+                      tooltip="Skip layers with max difference below this"),
+        io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+        io.String.Input("output_filename", default="extracted_lora"),
+        io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
+        io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+        io.String.Input("skip_patterns", default="", multiline=True,
+                       tooltip="Patterns for layers to skip (regex or glob depending on glob_skip_patterns)"),
+        io.Boolean.Input("glob_skip_patterns", default=False,
+                        tooltip="When True, skip_patterns use glob syntax (* = any sequence, ? = any char, dots are literal). "
+                                "When False (default), patterns are Python regex matched as substrings."),
+    ]
+
+
+class DoRAExtractFixed(io.ComfyNode):
+    """Extract DoRA with fixed rank."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRAExtractFixed",
+            display_name="DoRA Extract (Fixed Rank)",
+            category="ModelUtils/DoRA",
+            description="Extract DoRA with specified fixed rank for each layer type.",
+            inputs=[
+                *_get_model_inputs(),
+                io.Int.Input("linear_dim", default=64, min=1, max=16384,
+                            tooltip="Rank for linear/attention layers"),
+                io.Int.Input("conv_dim", default=32, min=1, max=16384,
+                            tooltip="Rank for conv layers"),
+                io.Int.Input("svd_niter", default=2, min=0, max=10,
+                            tooltip="SVD power iterations (higher = more accurate but slower)"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_dim, conv_dim, svd_niter, chunk_large_layers,
+                clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_from_files(
+            model_a_path, model_b_path, "fixed", linear_dim, conv_dim,
+            device, save_dtype, output_path, linear_dim, conv_dim,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers, svd_niter,
+            lazy_load, force_clear_cache, glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)
+
+
+class DoRAExtractRatio(io.ComfyNode):
+    """Extract DoRA with singular value ratio threshold."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRAExtractRatio",
+            display_name="DoRA Extract (Ratio)",
+            category="ModelUtils/DoRA",
+            description="Keep singular values > max(S) / ratio.",
+            inputs=[
+                *_get_model_inputs(),
+                io.Float.Input("linear_ratio", default=2.0, min=1.0, max=100.0, step=0.1,
+                              tooltip="Ratio threshold for linear layers (higher = more SVs kept)"),
+                io.Float.Input("conv_ratio", default=2.0, min=1.0, max=100.0, step=0.1,
+                              tooltip="Ratio threshold for conv layers (higher = more SVs kept)"),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
+                            tooltip="Maximum rank for linear layers"),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
+                            tooltip="Maximum rank for conv layers"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_ratio, conv_ratio, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_from_files(
+            model_a_path, model_b_path, "ratio", linear_ratio, conv_ratio,
+            device, save_dtype, output_path, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache,
+            glob_skip_patterns=glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)
+
+
+class DoRAExtractQuantile(io.ComfyNode):
+    """Extract DoRA with cumulative singular value quantile."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRAExtractQuantile",
+            display_name="DoRA Extract (Quantile)",
+            category="ModelUtils/DoRA",
+            description="Keep enough singular values to reach target percentage.",
+            inputs=[
+                *_get_model_inputs(),
+                io.Float.Input("linear_quantile", default=0.9, min=0.0, max=1.0, step=0.01,
+                              tooltip="Target cumulative % for linear layers"),
+                io.Float.Input("conv_quantile", default=0.9, min=0.0, max=1.0, step=0.01,
+                              tooltip="Target cumulative % for conv layers"),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
+                            tooltip="Maximum rank for linear layers"),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
+                            tooltip="Maximum rank for conv layers"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_quantile, conv_quantile, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_from_files(
+            model_a_path, model_b_path, "quantile", linear_quantile, conv_quantile,
+            device, save_dtype, output_path, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache,
+            glob_skip_patterns=glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)
+
+
+class DoRAExtractKnee(io.ComfyNode):
+    """Extract DoRA with automatic knee detection."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRAExtractKnee",
+            display_name="DoRA Extract (Knee Detection)",
+            category="ModelUtils/DoRA",
+            description="Automatically find optimal rank using knee detection on singular value curve.",
+            inputs=[
+                *_get_model_inputs(),
+                io.Combo.Input("knee_method", options=["sv_knee", "sv_cumulative_knee"],
+                              default="sv_knee", tooltip="Knee detection method"),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
+                            tooltip="Maximum rank for linear layers"),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
+                            tooltip="Maximum rank for conv layers"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, knee_method, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_from_files(
+            model_a_path, model_b_path, knee_method, 0, 0,
+            device, save_dtype, output_path, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache,
+            glob_skip_patterns=glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)
+
+
+class DoRAExtractFrobenius(io.ComfyNode):
+    """Extract DoRA preserving target fraction of Frobenius norm."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRAExtractFrobenius",
+            display_name="DoRA Extract (Frobenius)",
+            category="ModelUtils/DoRA",
+            description="Preserve target fraction of Frobenius norm.",
+            inputs=[
+                *_get_model_inputs(),
+                io.Float.Input("linear_target", default=0.9, min=0.0, max=1.0, step=0.01,
+                              tooltip="Target Frobenius norm fraction for linear"),
+                io.Float.Input("conv_target", default=0.9, min=0.0, max=1.0, step=0.01,
+                              tooltip="Target Frobenius norm fraction for conv"),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
+                            tooltip="Maximum rank for linear layers"),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
+                            tooltip="Maximum rank for conv layers"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_target, conv_target, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_from_files(
+            model_a_path, model_b_path, "sv_fro", linear_target, conv_target,
+            device, save_dtype, output_path, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache,
+            glob_skip_patterns=glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)

nodes/dora_learned_wd.py

@@ -0,0 +1,790 @@
+"""
+Learned DoRA Extraction via Optimization.
+
+Extracts Weight-Decomposed Low-Rank Adapters using gradient descent optimization
+to minimize the reconstruction MSE loss, matching the true non-linear DoRA formulation.
+Uses analytical SVD purely as the initialization seed.
+"""
+import fnmatch
+import os
+import re
+import math
+import torch
+import torch.linalg as linalg
+import folder_paths
+import comfy.utils
+from tqdm import tqdm
+from comfy_api.latest import io
+from .device_utils import (
+    estimate_model_size, prepare_for_large_operation,
+    cleanup_after_operation
+)
+
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned, IncrementalSafetensorsWriter
+
+# Import the existing SV ranking and chunking logic so we don't have to duplicate it
+from .dora_extract_wd import (
+    _compute_rank, _svd_extract_linear_lowrank, _svd_extract_linear,
+    _svd_extract_conv, _detect_fused_layer, _compile_patterns,
+    _matches_any_pattern, _build_lora_output_path, _get_model_inputs,
+    _format_lora_key, _get_common_inputs
+)
+
+
+def _adaptive_lr_update_cosine(curr_lr: float, improved: bool, worse_loss_counter: int, iteration: int, tensor_shape: tuple, min_lr: float, early_stop_stall: int, lr_factor: float = 0.95, lr_cooldown: int = 1) -> tuple:
+    M, N = tensor_shape
+    shape_ratio = abs(M - N) / max(M, N) if max(M, N) > 0 else 0
+
+    t = min(worse_loss_counter / max(early_stop_stall, 1), 1.0)
+    u_factor = (1 + math.cos(2 * math.pi * t)) / 2
+
+    if improved:
+        base_boost = 1.25
+        distance = base_boost - 1.0
+        scaled_distance = distance * (1.0 + 0.5 * shape_ratio)
+        boost_mult = 1.0 + scaled_distance * (1.0 - u_factor)
+        return min(curr_lr * boost_mult, 100.0), True
+    else:
+        cooldown = max(lr_cooldown, 1)
+        if iteration % cooldown != 0:
+            return curr_lr, False
+
+        min_decay = lr_factor - 0.03 * shape_ratio
+        max_decay = 0.995
+        decay_mult = min_decay + (max_decay - min_decay) * u_factor
+        return max(curr_lr * decay_mult, min_lr), True
+
+
+def _compute_shape_aware_plateau_params(M: int, N: int, lr_patience: int, lr_factor: float, lr_cooldown: int, lr_shape_influence: float = 1.0) -> tuple:
+    if lr_shape_influence > 0:
+        aspect_ratio = max(M, N) / min(M, N) if min(M, N) > 0 else 1.0
+        ar_factor = math.sqrt(aspect_ratio)
+        blend = lr_shape_influence
+
+        effective_patience = lr_patience
+        raw_factor = lr_factor
+        aggressive_factor = raw_factor**ar_factor
+        effective_factor = raw_factor + (aggressive_factor - raw_factor) * blend
+        effective_cooldown = lr_cooldown
+    else:
+        effective_patience = lr_patience
+        effective_factor = lr_factor
+        effective_cooldown = lr_cooldown
+
+    return effective_patience, effective_factor, effective_cooldown
+
+
+def _optimize_dora(
+    tensor_base: torch.Tensor,
+    tensor_ft: torch.Tensor,
+    init_down: torch.Tensor,
+    init_up: torch.Tensor,
+    optimize_iters: int,
+    learning_rate: float,
+    optimizer_type: str = "prodigy",
+    lr_schedule: str = "plateau",
+    lr_patience: int = 2,
+    lr_factor: float = 0.9,
+    lr_cooldown: int = 2,
+    early_stop_loss: float = 1e-6,
+    early_stop_stall: int = 2000,
+    early_stop_lr: float = 9.01e-9,
+    layer_name: str = "Layer",
+) -> tuple:
+    """
+    Run gradient descent to optimize lora_down and lora_up against the actual DoRA formula.
+
+    Args:
+        tensor_base: Base model weight
+        tensor_ft: Finetuned model weight
+        init_down: SVD-initialized lora_down matrix
+        init_up: SVD-initialized lora_up matrix
+        optimize_iters: Number of iterations for AdamW
+        learning_rate: Learning rate for AdamW
+        optimizer_type: Optimizer choice (adamw, radam, prodigy)
+        lr_schedule: LR Schedule choice (exponential, plateau, adaptive)
+        early_stop_loss: Target loss to stop optimizing
+        early_stop_stall: Number of non-improving iterations before stopping
+
+    Returns:
+        (optimized_down, optimized_up)
+    """
+    if optimize_iters <= 0:
+        return init_down, init_up
+
+    is_conv = tensor_base.ndim == 4
+    if is_conv:
+        out_ch, in_ch, kh, kw = tensor_base.shape
+        flat_base = tensor_base.reshape(out_ch, -1)
+        flat_ft = tensor_ft.reshape(out_ch, -1)
+
+        # Flatten init matrices for matmul: (out, rank) @ (rank, in*k*k)
+        rank = init_down.shape[0]
+        flat_up = init_up.reshape(out_ch, rank)
+        flat_down = init_down.reshape(rank, -1)
+    else:
+        flat_base = tensor_base
+        flat_ft = tensor_ft
+        flat_up = init_up
+        flat_down = init_down
+
+    # Magnitude of the finetuned model (fixed DoRA scale)
+    m = torch.linalg.norm(flat_ft, dim=1, keepdim=True).detach()
+    eps = 1e-8
+
+    # Set up learnable parameters
+    with torch.inference_mode(False), torch.enable_grad():
+        # Clone inference tensors to normal tensors for autograd
+        flat_base_opt = flat_base.clone()
+        flat_ft_opt = flat_ft.clone()
+        m_opt = m.clone()
+
+        # Free original tensors to save memory
+        del flat_base, flat_ft, m
+
+        B = torch.nn.Parameter(flat_up.clone())
+        A = torch.nn.Parameter(flat_down.clone())
+
+        opt_type = optimizer_type.lower()
+        if opt_type == "radam":
+            optimizer = torch.optim.RAdam([B, A], lr=learning_rate)
+        elif opt_type == "prodigy":
+            try:
+                from prodigyplus.prodigy_plus_schedulefree import ProdigyPlusScheduleFree
+                optimizer = ProdigyPlusScheduleFree([B, A], lr=learning_rate, use_schedulefree=False)
+            except ImportError:
+                print("ProdigyPlusScheduleFree optimizer not found. Falling back to AdamW. To install: pip install prodigyplus")
+                optimizer = torch.optim.AdamW([B, A], lr=learning_rate)
+        else:
+            optimizer = torch.optim.AdamW([B, A], lr=learning_rate)
+
+        # Learning rate scheduling and early stopping state
+        best_loss = float('inf')
+        worse_loss_counter = 0
+        plateau_counter = 0
+        cooldown_counter = 0
+        current_lr = learning_rate
+        schedule_type = lr_schedule.lower()
+
+        M_dim, N_dim = flat_base_opt.shape[0], flat_base_opt.shape[1]
+        eff_patience, eff_factor, eff_cooldown = _compute_shape_aware_plateau_params(M_dim, N_dim, lr_patience, lr_factor, lr_cooldown)
+
+        # Optimization loop
+        opt_pbar = tqdm(range(optimize_iters), desc=f"  └─ Optimizing {layer_name}", leave=False, dynamic_ncols=True, mininterval=0.5)
+        for i in opt_pbar:
+            optimizer.zero_grad()
+
+            # 1. Reconstruct directional weight: W_dir = W_base + B @ A
+            w_approx = flat_base_opt + (B @ A)
+
+            # 2. Normalize directional weight
+            approx_norm = torch.linalg.norm(w_approx, dim=1, keepdim=True) + eps
+
+            # 3. Scale by target magnitude
+            w_reconstructed = m_opt * (w_approx / approx_norm)
+
+            # 4. Compute MSE loss against actual finetuned weight
+            loss = torch.nn.functional.mse_loss(w_reconstructed, flat_ft_opt)
+            loss_val = loss.item()
+
+            if i == 0:
+                initial_loss = loss_val
+                best_loss = loss_val
+
+            loss.backward()
+            optimizer.step()
+
+            # --- Early Stopping & Schedulers ---
+            improved = False
+            if loss_val < best_loss * 0.9999: # 0.01% improvement threshold
+                best_loss = loss_val
+                worse_loss_counter = 0
+                plateau_counter = 0
+                improved = True
+            else:
+                worse_loss_counter += 1
+                plateau_counter += 1
+
+            if loss_val <= early_stop_loss:
+                break
+
+            if worse_loss_counter > early_stop_stall:
+                break
+
+            if current_lr <= early_stop_lr:
+                break
+
+            # LR Schedulers
+            if schedule_type == "adaptive":
+                current_lr, updated = _adaptive_lr_update_cosine(
+                    current_lr, improved, worse_loss_counter, i, (M_dim, N_dim), early_stop_lr, early_stop_stall, lr_factor, lr_cooldown
+                )
+                if updated:
+                    for param_group in optimizer.param_groups:
+                        param_group['lr'] = current_lr
+            elif schedule_type == "plateau":
+                if cooldown_counter > 0:
+                    cooldown_counter -= 1
+                elif plateau_counter >= max(eff_patience, 1): # Decay if patience threshold reached
+                    current_lr = max(current_lr * eff_factor, early_stop_lr)
+                    for param_group in optimizer.param_groups:
+                        param_group['lr'] = current_lr
+                    plateau_counter = 0 # reset counter after decay
+                    cooldown_counter = eff_cooldown
+            elif schedule_type == "exponential":
+                current_lr = max(current_lr * 0.99, early_stop_lr)
+                for param_group in optimizer.param_groups:
+                    param_group['lr'] = current_lr
+
+            if i % 10 == 0 or i == optimize_iters - 1:
+                if schedule_type == "plateau":
+                    opt_pbar.set_postfix({"loss": f"{loss_val:.4e}", "lr": f"{current_lr:.2e}", "plat": f"{plateau_counter}/{eff_patience}"}, refresh=False)
+                else:
+                    opt_pbar.set_postfix({"loss": f"{loss_val:.4e}", "lr": f"{current_lr:.2e}", "stall": f"{worse_loss_counter}"}, refresh=False)
+
+        final_loss = loss.item()
+        opt_pbar.close()
+        print(f"  └─ {layer_name} | Loss: {initial_loss:.6f} -> {final_loss:.6f} (stopped at iter {i+1})")
+
+        final_up = B.detach()
+        final_down = A.detach()
+
+    # Reshape back to original dimensions
+    if is_conv:
+        final_up = final_up.reshape(out_ch, rank, 1, 1)
+        final_down = final_down.reshape(rank, in_ch, kh, kw)
+
+    return final_down, final_up
+
+
+def _extract_chunked_learned_layer(
+    tensor_base: torch.Tensor,
+    tensor_ft: torch.Tensor,
+    num_chunks: int,
+    mode: str,
+    mode_param: float,
+    device: str,
+    max_rank: int = None,
+    optimize_iters: int = 500,
+    learning_rate: float = 1.0,
+    optimizer_type: str = "prodigy",
+    lr_schedule: str = "plateau",
+    lr_patience: int = 2,
+    lr_factor: float = 0.9,
+    lr_cooldown: int = 2,
+    early_stop_loss: float = 1e-6,
+    early_stop_stall: int = 2000,
+    early_stop_lr: float = 9.01e-9,
+    layer_name: str = "Layer",
+) -> tuple:
+    """Extract Learned DoRA from fused layer by chunking."""
+    out_dim, in_dim = tensor_base.shape
+    chunk_size = out_dim // num_chunks
+
+    all_lora_up = []
+    all_lora_down = []
+
+    for i in range(num_chunks):
+        chunk_base = tensor_base[i * chunk_size:(i + 1) * chunk_size, :]
+        chunk_ft = tensor_ft[i * chunk_size:(i + 1) * chunk_size, :]
+
+        norm_base = torch.linalg.norm(chunk_base, dim=1, keepdim=True)
+        norm_ft = torch.linalg.norm(chunk_ft, dim=1, keepdim=True)
+        eps = 1e-8
+
+        w_dir = chunk_ft * (norm_base / (norm_ft + eps))
+        weight_diff = w_dir - chunk_base
+
+        try:
+            result, mode_str = _svd_extract_linear(weight_diff, mode, mode_param, device, max_rank)
+            if mode_str == "full":
+                return None, None, 0
+            lora_down, lora_up, _ = result
+
+            # Apply learned optimization
+            lora_down, lora_up = _optimize_dora(
+                chunk_base, chunk_ft, lora_down, lora_up, optimize_iters, learning_rate,
+                optimizer_type, lr_schedule, lr_patience, lr_factor, lr_cooldown,
+                early_stop_loss, early_stop_stall, early_stop_lr,
+                layer_name=f"{layer_name} (Chunk {i+1}/{num_chunks})"
+            )
+
+            all_lora_down.append(lora_down)
+            all_lora_up.append(lora_up)
+        except Exception:
+            return None, None, 0
+
+    combined_lora_up = torch.cat(all_lora_up, dim=0)
+    combined_lora_down = all_lora_down[0]
+    rank = combined_lora_down.shape[0]
+
+    return combined_lora_up, combined_lora_down, rank
+
+
+def extract_dora_learned_from_files(
+    model_a_path: str,
+    model_b_path: str,
+    mode: str,
+    linear_param: float,
+    conv_param: float,
+    device: str,
+    save_dtype: str,
+    output_path: str,
+    optimize_iters: int = 500,
+    learning_rate: float = 1.0,
+    optimizer_type: str = "prodigy",
+    lr_schedule: str = "plateau",
+    lr_patience: int = 2,
+    lr_factor: float = 0.9,
+    lr_cooldown: int = 2,
+    early_stop_loss: float = 1e-6,
+    early_stop_stall: int = 2000,
+    early_stop_lr: float = 9.01e-9,
+    linear_max_rank: int = None,
+    conv_max_rank: int = None,
+    clamp_quantile: float = 0.99,
+    min_diff: float = 0.0,
+    skip_patterns_str: str = "",
+    mismatch_mode: str = "skip",
+    chunk_large_layers: bool = True,
+    svd_niter: int = 2,
+    lazy_load: bool = True,
+    force_clear_cache: bool = True,
+    glob_skip_patterns: bool = False,
+) -> None:
+    """
+    Extract Learned DoRA from difference between two models.
+    """
+    save_torch_dtype = {
+        "fp32": torch.float32,
+        "fp16": torch.float16,
+        "bf16": torch.bfloat16,
+    }.get(save_dtype, torch.float16)
+
+    skip_patterns = _compile_patterns(skip_patterns_str, glob_mode=glob_skip_patterns)
+
+    total_size_gb = estimate_model_size(model_a_path) + estimate_model_size(model_b_path)
+    print(f"[Learned DoRA Extract] Preparing memory for {total_size_gb:.2f}GB operation...")
+    prepare_for_large_operation(total_size_gb * 1.5, torch.device(device))
+
+    handler_a = MemoryEfficientSafeOpen(model_a_path, low_memory=lazy_load)
+    handler_b = MemoryEfficientSafeOpen(model_b_path, low_memory=lazy_load)
+
+    try:
+        keys_a = set(handler_a.keys())
+        keys_b = set(handler_b.keys())
+        weight_keys = [k for k in keys_a if k.endswith(".weight")]
+        pbar = comfy.utils.ProgressBar(len(weight_keys))
+        stats = {"extracted": 0, "full": 0, "skipped": 0, "chunked": 0}
+
+        def _process_layer(key):
+            lora_name = _format_lora_key(key)
+
+            if _matches_any_pattern(key, skip_patterns, glob_mode=glob_skip_patterns):
+                return "skipped", None
+
+            use_pinned = device == 'cuda'
+
+            if key not in keys_b:
+                if mismatch_mode == "skip":
+                    return "skipped", None
+                if mismatch_mode == "error":
+                    raise ValueError(f"Key {key} not found in model B")
+
+                # Cannot extract true DoRA without base. Fallback to diff.
+                cpu_a = handler_a.get_tensor(key)
+                tensor_a = transfer_to_gpu_pinned(cpu_a, device, torch.float32) if use_pinned else cpu_a.to(device=device, dtype=torch.float32)
+                del cpu_a
+                weight_diff = tensor_a
+                tensor_base = torch.zeros_like(tensor_a)
+                tensor_ft = tensor_a
+                dora_scale = None
+            else:
+                cpu_a = handler_a.get_tensor(key)
+                cpu_b = handler_b.get_tensor(key)
+
+                if use_pinned:
+                    tensor_a = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
+                    tensor_b = transfer_to_gpu_pinned(cpu_b, device, torch.float32)
+                else:
+                    tensor_a = cpu_a.to(device=device, dtype=torch.float32)
+                    tensor_b = cpu_b.to(device=device, dtype=torch.float32)
+                del cpu_a, cpu_b
+
+                if tensor_a.shape != tensor_b.shape:
+                    if mismatch_mode == "skip":
+                        del tensor_a, tensor_b
+                        return "skipped", None
+                    if mismatch_mode == "error":
+                        raise ValueError(f"Shape mismatch for {key}: {tensor_a.shape} vs {tensor_b.shape}")
+
+                    weight_diff = tensor_a
+                    tensor_base = torch.zeros_like(tensor_a)
+                    tensor_ft = tensor_a
+                    dora_scale = None
+                    del tensor_b
+                else:
+                    tensor_base = tensor_b
+                    tensor_ft = tensor_a
+
+                    is_conv_dim = tensor_a.ndim == 4
+                    if is_conv_dim:
+                        out_ch = tensor_a.shape[0]
+                        flat_a = tensor_a.reshape(out_ch, -1)
+                        flat_b = tensor_b.reshape(out_ch, -1)
+                        norm_a = torch.linalg.norm(flat_a, dim=1, keepdim=True)
+                        norm_b = torch.linalg.norm(flat_b, dim=1, keepdim=True)
+                    else:
+                        norm_a = torch.linalg.norm(tensor_a, dim=1, keepdim=True)
+                        norm_b = torch.linalg.norm(tensor_b, dim=1, keepdim=True)
+
+                    eps = 1e-8
+                    norm_ratio = norm_b / (norm_a + eps)
+                    if is_conv_dim:
+                        norm_ratio = norm_ratio.view(-1, 1, 1, 1)
+                        dora_scale = norm_a.view(-1, 1, 1, 1)
+                    else:
+                        dora_scale = norm_a
+
+                    w_dir = tensor_a * norm_ratio
+                    weight_diff = w_dir - tensor_b
+                    del w_dir
+
+            if min_diff > 0 and weight_diff.abs().max() < min_diff:
+                del weight_diff, tensor_base, tensor_ft
+                return "skipped", None
+
+            if weight_diff.ndim < 2:
+                del weight_diff, tensor_base, tensor_ft
+                return "skipped", None
+
+            is_conv = weight_diff.ndim == 4
+            layer_results = {}
+
+            try:
+                # 1. Run Analytical SVD (Initialization)
+                if is_conv:
+                    result, mode_str = _svd_extract_conv(
+                        weight_diff, mode, conv_param, device, conv_max_rank, clamp_quantile
+                    )
+                else:
+                    result, mode_str = _svd_extract_linear(
+                        weight_diff, mode, linear_param, device, linear_max_rank, clamp_quantile, svd_niter
+                    )
+
+                # 2. Run Optimization (Learned Rounding)
+                if mode_str != "full" and dora_scale is not None:
+                    lora_down, lora_up, _ = result
+                    lora_down, lora_up = _optimize_dora(
+                        tensor_base, tensor_ft, lora_down, lora_up, optimize_iters, learning_rate,
+                        optimizer_type, lr_schedule, lr_patience, lr_factor, lr_cooldown,
+                        early_stop_loss, early_stop_stall, early_stop_lr,
+                        layer_name=lora_name
+                    )
+                    result = (lora_down, lora_up, _)
+
+            except Exception as e:
+                if chunk_large_layers and not is_conv:
+                    num_chunks = _detect_fused_layer(key, weight_diff.shape)
+                    if num_chunks > 1:
+                        print(f"[Learned DoRA] Chunked: {key} ({num_chunks} chunks)")
+                        lora_up, lora_down, rank = _extract_chunked_learned_layer(
+                            tensor_base, tensor_ft, num_chunks, mode, linear_param, device, linear_max_rank, optimize_iters, learning_rate,
+                            optimizer_type, lr_schedule, lr_patience, lr_factor, lr_cooldown,
+                            early_stop_loss, early_stop_stall, early_stop_lr,
+                            layer_name=lora_name
+                        )
+                        if lora_up is not None:
+                            layer_results[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.alpha"] = torch.tensor(rank, dtype=save_torch_dtype)
+                            if dora_scale is not None:
+                                layer_results[f"{lora_name}.dora_scale"] = dora_scale.to(save_torch_dtype).cpu().contiguous()
+                            del weight_diff, tensor_base, tensor_ft
+                            return "chunked", layer_results
+
+                print(f"[Learned DoRA] Failed: {key}: {e}")
+                del weight_diff, tensor_base, tensor_ft
+                return "skipped", None
+
+            if mode_str == "full":
+                layer_results[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
+                status = "full"
+            else:
+                lora_down, lora_up, _ = result
+                rank = lora_down.shape[0]
+                layer_results[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.alpha"] = torch.tensor(rank, dtype=save_torch_dtype)
+                if dora_scale is not None:
+                    layer_results[f"{lora_name}.dora_scale"] = dora_scale.to(save_torch_dtype).cpu().contiguous()
+                status = "extracted"
+
+            del weight_diff, tensor_base, tensor_ft
+            return status, layer_results
+
+        writer = IncrementalSafetensorsWriter(output_path)
+        writer.__enter__()
+        try:
+            for key in tqdm(weight_keys, desc="Extracting Learned DoRA", unit="layers"):
+                status, layer_sd = _process_layer(key)
+                stats[status] += 1
+                if layer_sd:
+                    writer.write_dict(layer_sd)
+
+                if force_clear_cache:
+                    import gc
+                    gc.collect()
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+
+                pbar.update(1)
+        finally:
+            writer.__exit__(None, None, None)
+
+        print(f"[Learned DoRA] Done: {stats['extracted']} extracted, {stats['chunked']} chunked, "
+              f"{stats['full']} full, {stats['skipped']} skipped")
+
+    finally:
+        handler_a.__exit__(None, None, None)
+        handler_b.__exit__(None, None, None)
+        cleanup_after_operation()
+
+
+# =============================================================================
+# Node Definitions
+# =============================================================================
+
+def _get_learned_inputs():
+    return [
+        io.Int.Input("optimize_iters", default=500, min=0, max=10000,
+                    tooltip="Number of gradient descent iterations (0 = skip optimization)"),
+        io.Float.Input("learning_rate", default=1.0, min=0.0001, max=100.0, step=0.001,
+                      tooltip="Base learning rate. AdamW prefers ~0.01, Prodigy prefers ~1.0"),
+        io.Combo.Input("optimizer", options=["prodigy", "adamw", "radam"], default="prodigy",
+                      tooltip="Optimization algorithm"),
+        io.Combo.Input("lr_schedule", options=["plateau", "adaptive", "exponential", "constant"], default="plateau",
+                      tooltip="Learning rate scheduling strategy"),
+        io.Int.Input("lr_patience", default=2, min=0, max=1000, tooltip="Steps to wait before decaying LR (Plateau)"),
+        io.Float.Input("lr_factor", default=0.9, min=0.01, max=1.0, step=0.01, tooltip="Factor to decay LR by (Plateau)"),
+        io.Int.Input("lr_cooldown", default=2, min=0, max=1000, tooltip="Steps to wait after decay before checking again (Plateau)"),
+        io.Float.Input("early_stop_loss", default=1e-6, min=0.0, max=1.0, step=1e-9,
+                      tooltip="Stop early if MSE loss drops below this value"),
+        io.Int.Input("early_stop_stall", default=2000, min=0, max=10000,
+                      tooltip="Stop early if loss doesn't improve for this many iterations"),
+        io.Float.Input("early_stop_lr", default=9.01e-9, min=0.0, max=1.0, step=1e-10,
+                      tooltip="Stop early if learning rate drops below this value"),
+    ]
+
+
+class DoRALearnedExtractFixed(io.ComfyNode):
+    """Extract Learned DoRA with fixed rank."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRALearnedExtractFixed",
+            display_name="DoRA Learned Extract (Fixed Rank) (WIP)",
+            category="ModelUtils/DoRA",
+            description="Extract DoRA via MSE Loss minimization with fixed rank.",
+            is_experimental=True,
+            inputs=[
+                *_get_model_inputs(),
+                io.Int.Input("linear_dim", default=64, min=1, max=16384, tooltip="Rank for linear/attention layers"),
+                io.Int.Input("conv_dim", default=32, min=1, max=16384, tooltip="Rank for conv layers"),
+                *_get_learned_inputs(),
+                io.Int.Input("svd_niter", default=2, min=0, max=10, tooltip="SVD iterations for initialization"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_dim, conv_dim, optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, svd_niter, chunk_large_layers,
+                clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_learned_from_files(
+            model_a_path, model_b_path, "fixed", linear_dim, conv_dim,
+            device, save_dtype, output_path, optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, linear_dim, conv_dim,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers, svd_niter,
+            lazy_load, force_clear_cache, glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)
+
+
+class DoRALearnedExtractRatio(io.ComfyNode):
+    """Extract Learned DoRA with singular value ratio threshold."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRALearnedExtractRatio",
+            display_name="DoRA Learned Extract (Ratio) (WIP)",
+            category="ModelUtils/DoRA",
+            description="Learned extraction keeping SVs > max(S) / ratio.",
+            is_experimental=True,
+            inputs=[
+                *_get_model_inputs(),
+                io.Float.Input("linear_ratio", default=2.0, min=1.0, max=100.0, step=0.1),
+                io.Float.Input("conv_ratio", default=2.0, min=1.0, max=100.0, step=0.1),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_learned_inputs(),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_ratio, conv_ratio, linear_max_rank, conv_max_rank,
+                optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, chunk_large_layers, clamp_quantile, min_diff, mismatch_mode,
+                output_filename, save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_learned_from_files(
+            model_a_path, model_b_path, "ratio", linear_ratio, conv_ratio,
+            device, save_dtype, output_path, optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)
+
+
+class DoRALearnedExtractQuantile(io.ComfyNode):
+    """Extract Learned DoRA with cumulative singular value quantile."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRALearnedExtractQuantile",
+            display_name="DoRA Learned Extract (Quantile) (WIP)",
+            category="ModelUtils/DoRA",
+            description="Learned extraction keeping enough SVs to reach target percentage.",
+            is_experimental=True,
+            inputs=[
+                *_get_model_inputs(),
+                io.Float.Input("linear_quantile", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Float.Input("conv_quantile", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_learned_inputs(),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_quantile, conv_quantile, linear_max_rank, conv_max_rank,
+                optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, chunk_large_layers, clamp_quantile, min_diff, mismatch_mode,
+                output_filename, save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_learned_from_files(
+            model_a_path, model_b_path, "quantile", linear_quantile, conv_quantile,
+            device, save_dtype, output_path, optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)
+
+
+class DoRALearnedExtractKnee(io.ComfyNode):
+    """Extract Learned DoRA with automatic knee detection."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRALearnedExtractKnee",
+            display_name="DoRA Learned Extract (Knee Detection) (WIP)",
+            category="ModelUtils/DoRA",
+            description="Learned extraction automatically finding optimal rank using knee detection.",
+            is_experimental=True,
+            inputs=[
+                *_get_model_inputs(),
+                io.Combo.Input("knee_method", options=["sv_knee", "sv_cumulative_knee"], default="sv_knee"),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_learned_inputs(),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, knee_method, linear_max_rank, conv_max_rank,
+                optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, chunk_large_layers, clamp_quantile, min_diff, mismatch_mode,
+                output_filename, save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_learned_from_files(
+            model_a_path, model_b_path, knee_method, 0, 0,
+            device, save_dtype, output_path, optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)
+
+
+class DoRALearnedExtractFrobenius(io.ComfyNode):
+    """Extract Learned DoRA preserving target fraction of Frobenius norm."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DoRALearnedExtractFrobenius",
+            display_name="DoRA Learned Extract (Frobenius) (WIP)",
+            category="ModelUtils/DoRA",
+            description="Learned extraction preserving target fraction of Frobenius norm.",
+            is_experimental=True,
+            inputs=[
+                *_get_model_inputs(),
+                io.Float.Input("linear_target", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Float.Input("conv_target", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_learned_inputs(),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_target, conv_target, linear_max_rank, conv_max_rank,
+                optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, chunk_large_layers, clamp_quantile, min_diff, mismatch_mode,
+                output_filename, save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_dora_learned_from_files(
+            model_a_path, model_b_path, "sv_fro", linear_target, conv_target,
+            device, save_dtype, output_path, optimize_iters, learning_rate, optimizer, lr_schedule, lr_patience, lr_factor, lr_cooldown, early_stop_loss, early_stop_stall, early_stop_lr, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+
+        return io.NodeOutput(output_path)

nodes/downloader_utils.py

@@ -36,12 +36,7 @@ except ImportError:
 # Assuming downloader_utils.py is inside nodes/, the root is one level up
 root_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 sys.path.append(os.path.join(root_dir, 'unifiedefficientloader-reference'))
-try:
-    from unifiedefficientloader import UnifiedSafetensorsLoader
-    HAS_UNIFIED_LOADER = True
-except ImportError:
-    HAS_UNIFIED_LOADER = False
-    print("Warning: unifiedefficientloader not found. Safetensors metadata extraction disabled.")
+from unifiedefficientloader import UnifiedSafetensorsLoader
 
 # --- Database Path Logic ---
 try:
@@ -416,8 +411,6 @@ def filter_images_by_nsfw(images, threshold_level):
     return valid_images
 
 def extract_safetensors_metadata(filepath):
-    if not HAS_UNIFIED_LOADER:
-        return None
     try:
         # Use low_memory=True to only read the header and metadata without loading tensors
         loader = UnifiedSafetensorsLoader(filepath, low_memory=True)

nodes/lora_extract_svd.py

@@ -4,6 +4,7 @@ LoRA Extraction via SVD decomposition.
 Native implementation of Low-Rank Adapter extraction with multiple rank selection modes.
 No external dependencies required.
 """
+import fnmatch
 import os
 import re
 import torch
@@ -12,13 +13,16 @@ import folder_paths
 import comfy.utils
 from tqdm import tqdm
 from comfy_api.latest import io
-from safetensors.torch import save_file
-from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
 from .device_utils import (
     estimate_model_size, prepare_for_large_operation,
     cleanup_after_operation, get_device_capabilities
 )
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned, IncrementalSafetensorsWriter
+from .quantization_guard import inspect_low_bit_input
+
+
+
 
 # =============================================================================
 # Rank Index Functions - Determine LoRA rank from singular values
@@ -396,14 +400,22 @@ def _extract_chunked_layer(
 # Pattern Matching
 # =============================================================================
 
-def _compile_patterns(pattern_string: str) -> list:
-    """Compile regex patterns from whitespace-separated string."""
+def _compile_patterns(pattern_string: str, glob_mode: bool = False) -> list:
+    """Compile patterns from whitespace-separated string.
+
+    In regex mode (default) returns compiled re.Pattern objects.
+    In glob mode returns plain strings; fnmatch handles matching.
+    """
     if not pattern_string or not pattern_string.strip():
         return []
     patterns = []
     for p in pattern_string.split():
         p = p.strip()
-        if p:
+        if not p:
+            continue
+        if glob_mode:
+            patterns.append(p)
+        else:
             try:
                 patterns.append(re.compile(p))
             except re.error:
@@ -411,8 +423,15 @@ def _compile_patterns(pattern_string: str) -> list:
     return patterns
 
 
-def _matches_any_pattern(key: str, patterns: list) -> bool:
-    """Check if key matches any pattern."""
+def _matches_any_pattern(key: str, patterns: list, glob_mode: bool = False) -> bool:
+    """Check if key matches any pattern.
+
+    Glob mode: each pattern is matched as a substring glob using fnmatch
+    (dots are literal, * matches any sequence of characters).
+    Regex mode: each pattern is a compiled re.Pattern, matched via search().
+    """
+    if glob_mode:
+        return any(fnmatch.fnmatch(key, f"*{p}*") for p in patterns)
     return any(p.search(key) for p in patterns)
 
 
@@ -428,6 +447,7 @@ def extract_lora_from_files(
     conv_param: float,
     device: str,
     save_dtype: str,
+    output_path: str,
     linear_max_rank: int = None,
     conv_max_rank: int = None,
     clamp_quantile: float = 0.99,
@@ -438,9 +458,11 @@ def extract_lora_from_files(
     svd_niter: int = 2,
     lazy_load: bool = True,
     force_clear_cache: bool = True,
-) -> dict[str, torch.Tensor]:
+    glob_skip_patterns: bool = False,
+    include_1d_diffs: bool = False,
+) -> None:
     """
-    Extract LoRA from difference between two models.
+    Extract LoRA from difference between two models, writing incrementally to disk.
 
     Args:
         model_a_path: Finetuned model path
@@ -449,7 +471,8 @@ def extract_lora_from_files(
         linear_param: Mode parameter for linear layers
         conv_param: Mode parameter for conv layers
         device: Computation device
-        save_dtype: Output dtype
+        save_dtype: Output dtype for layers whose sources are not FP32
+        output_path: Full path to output safetensors file
         linear_max_rank: Max rank for linear layers
         conv_max_rank: Max rank for conv layers
         clamp_quantile: Quantile for weight clamping
@@ -457,16 +480,18 @@ def extract_lora_from_files(
         skip_patterns_str: Patterns for layers to skip
         mismatch_mode: How to handle mismatches
         chunk_large_layers: Enable chunked extraction
+        glob_skip_patterns: When True, treat skip_patterns as glob (* wildcard).
+                            When False (default), treat as Python regex.
+        include_1d_diffs: Save 1D tensors in FP32 and keys ending in .scale or
+                          .lin as full .diff patches instead of skipping them.
     """
-    output_sd = {}
-
     save_torch_dtype = {
         "fp32": torch.float32,
         "fp16": torch.float16,
         "bf16": torch.bfloat16,
     }.get(save_dtype, torch.float16)
 
-    skip_patterns = _compile_patterns(skip_patterns_str)
+    skip_patterns = _compile_patterns(skip_patterns_str, glob_mode=glob_skip_patterns)
 
     # Prepare memory before heavy operation
     total_size_gb = estimate_model_size(model_a_path) + estimate_model_size(model_b_path)
@@ -477,18 +502,37 @@ def extract_lora_from_files(
     handler_b = MemoryEfficientSafeOpen(model_b_path, low_memory=lazy_load)
 
     try:
+        low_bit_a = inspect_low_bit_input(handler_a, f"Model A ({model_a_path})", "LoRA Extract")
+        low_bit_b = inspect_low_bit_input(handler_b, f"Model B ({model_b_path})", "LoRA Extract")
         keys_a = set(handler_a.keys())
         keys_b = set(handler_b.keys())
-        weight_keys = [k for k in keys_a if k.endswith(".weight")]
-        pbar = comfy.utils.ProgressBar(len(weight_keys))
+        direct_diff_suffixes = (".scale", ".lin")
+        extraction_keys = [
+            k for k in keys_a
+            if k not in low_bit_a
+            and k not in low_bit_b
+            and (k.endswith(".weight")
+            or (
+                include_1d_diffs
+                and (len(handler_a.get_shape(k)) == 1 or k.endswith(direct_diff_suffixes))
+            ))
+        ]
+        pbar = comfy.utils.ProgressBar(len(extraction_keys))
         stats = {"extracted": 0, "full": 0, "skipped": 0, "chunked": 0}
 
         def _process_layer(key):
             lora_name = _format_lora_key(key)
 
-            if _matches_any_pattern(key, skip_patterns):
+            if _matches_any_pattern(key, skip_patterns, glob_mode=glob_skip_patterns):
                 return "skipped", None
 
+            source_dtypes = [handler_a.get_dtype(key)]
+            if key in keys_b:
+                source_dtypes.append(handler_b.get_dtype(key))
+            layer_save_dtype = (
+                torch.float32 if torch.float32 in source_dtypes else save_torch_dtype
+            )
+
             # Load tensors with pinned memory for CUDA
             use_pinned = device == 'cuda'
 
@@ -540,7 +584,18 @@ def extract_lora_from_files(
                 del weight_diff
                 return "skipped", None
 
-            # Skip 1D tensors
+            # SVD requires matrix-shaped tensors. Optionally preserve 1D weights
+            # and scale/lin parameters as exact runtime-applicable diff patches.
+            is_direct_diff = weight_diff.ndim == 1 or key.endswith(direct_diff_suffixes)
+            if include_1d_diffs and is_direct_diff:
+                direct_diff_dtype = torch.float32 if weight_diff.ndim == 1 else layer_save_dtype
+                layer_results = {
+                    _format_direct_diff_key(key, lora_name):
+                        weight_diff.to(direct_diff_dtype).cpu().contiguous()
+                }
+                del weight_diff
+                return "full", layer_results
+
             if weight_diff.ndim < 2:
                 del weight_diff
                 return "skipped", None
@@ -567,9 +622,8 @@ def extract_lora_from_files(
                             weight_diff, num_chunks, mode, linear_param, device, linear_max_rank
                         )
                         if lora_up is not None:
-                            layer_results[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
-                            layer_results[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
-                            layer_results[f"{lora_name}.alpha"] = torch.tensor(rank).to(save_torch_dtype)
+                            layer_results[f"{lora_name}.lora_B.weight"] = lora_up.to(layer_save_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.lora_A.weight"] = lora_down.to(layer_save_dtype).cpu().contiguous()
                             del weight_diff
                             return "chunked", layer_results
 
@@ -579,31 +633,35 @@ def extract_lora_from_files(
 
             # Store result
             if mode_str == "full":
-                layer_results[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.diff"] = weight_diff.to(layer_save_dtype).cpu().contiguous()
                 status = "full"
             else:
                 lora_down, lora_up, _ = result
-                layer_results[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
-                layer_results[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
-                layer_results[f"{lora_name}.alpha"] = torch.tensor(lora_down.shape[0]).to(save_torch_dtype)
+                layer_results[f"{lora_name}.lora_A.weight"] = lora_down.to(layer_save_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.lora_B.weight"] = lora_up.to(layer_save_dtype).cpu().contiguous()
                 status = "extracted"
 
             del weight_diff
             return status, layer_results
 
-        for key in tqdm(weight_keys, desc="Extracting LoRA", unit="layers"):
-            status, layer_sd = _process_layer(key)
-            stats[status] += 1
-            if layer_sd:
-                output_sd.update(layer_sd)
+        writer = IncrementalSafetensorsWriter(output_path)
+        writer.__enter__()
+        try:
+            for key in tqdm(extraction_keys, desc="Extracting LoRA", unit="layers"):
+                status, layer_sd = _process_layer(key)
+                stats[status] += 1
+                if layer_sd:
+                    writer.write_dict(layer_sd)
 
-            if force_clear_cache:
-                import gc
-                gc.collect()
-                if torch.cuda.is_available():
-                    torch.cuda.empty_cache()
+                if force_clear_cache:
+                    import gc
+                    gc.collect()
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
 
-            pbar.update(1)
+                pbar.update(1)
+        finally:
+            writer.__exit__(None, None, None)
 
         print(f"[LoRA Extract] Done: {stats['extracted']} extracted, {stats['chunked']} chunked, "
               f"{stats['full']} full, {stats['skipped']} skipped")
@@ -613,22 +671,12 @@ def extract_lora_from_files(
         handler_b.__exit__(None, None, None)
         cleanup_after_operation()
 
-    return output_sd
-
-
-def _save_lora(output_sd: dict, output_filename: str) -> str:
-    """Save LoRA to file."""
-    if not output_sd:
-        raise ValueError("No LoRA weights extracted")
 
+def _build_lora_output_path(output_filename: str) -> str:
+    """Build output path for LoRA file."""
     output_dir = folder_paths.get_folder_paths("loras")[0]
     os.makedirs(output_dir, exist_ok=True)
-    output_path = os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
-
-    save_file(output_sd, output_path)
-    print(f"[LoRA Extract] Saved to {output_path}")
-
-    return output_path
+    return os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
 
 
 # =============================================================================
@@ -644,84 +692,59 @@ def _get_model_inputs():
     ]
 
 
-def _reconstruct_dots(key: str) -> str:
-    """Reconstruct dot structure from underscored key using heuristics."""
-    # Block indices: input_blocks_1 -> input_blocks.1
-    blocks = [
-        "input_blocks", "output_blocks", "middle_block",
-        "transformer_blocks", "single_transformer_blocks",
-        "double_blocks", "single_blocks",
-        "down_blocks", "up_blocks", "mid_block",
-        "attentions", "resnets", "upsamplers", "downsamplers"
-    ]
-    for b in blocks:
-        key = re.sub(f"{b}_(\\d+)", f"{b}.\\1", key)
-
-    # Sequential numbers: .1_1 -> .1.1
-    key = re.sub(r"(\.\d+)_(\d+)", r"\1.\2", key)
-
-    # Separators that should be preceded by dot
-    separators = [
-        "transformer_blocks", "single_transformer_blocks",
-        "attn1", "attn2", "attn",
-        "img_attn", "txt_attn",
-        "to_q", "to_k", "to_v", "to_out",
-        "q_proj", "k_proj", "v_proj", "out_proj",
-        "qkv", "proj", "ff", "net",
-        "norm1", "norm2", "norm3", "norm",
-        "time_emb_proj"
-    ]
-    for s in separators:
-        key = key.replace(f"_{s}", f".{s}")
-
-    # Special case: to_out_0 -> to_out.0
-    key = re.sub(r"to_out_(\d+)", r"to_out.\1", key)
-
-    return key
-
-
 def _format_lora_key(key: str) -> str:
     """
     Format the key for LoRA saving.
-    Standardizes to 'diffusion_model.' or 'transformer.' prefix.
+    Always standardizes to 'diffusion_model.' prefix for maximum compatibility.
+
+    ComfyUI maps 'diffusion_model.<layer>' generically for all model types, making
+    it universally compatible. The 'transformer.' prefix only works for a subset of
+    model types via model-specific mapping code and is intentionally not used here.
     """
     if key.endswith(".weight"):
         key = key[:-7]
 
-    # Handle ComfyUI Checkpoint wrapper
+    # Handle ComfyUI Checkpoint wrapper (ldm/sgm format: model.diffusion_model.*)
     if key.startswith("model.diffusion_model."):
-        # Check if it's a Diffusers-format transformer inside a checkpoint
-        inner_key = key[22:]  # len("model.diffusion_model.")
-        if inner_key.startswith("transformer_blocks") or inner_key.startswith("single_transformer_blocks"):
-            return f"transformer.{inner_key}"
-        # Standard checkpoint or Original Flux
+        inner_key = key[22:]  # strip len("model.diffusion_model.")
         return f"diffusion_model.{inner_key}"
 
-    # Handle direct Diffusers keys
+    # Handle net.* wrapper
+    if key.startswith("net."):
+        inner_key = key[4:]  # strip len("net.")
+        return f"diffusion_model.{inner_key}"
+
+    # Handle direct Diffusers keys (transformer_blocks.* / single_transformer_blocks.* without prefix)
     if key.startswith("transformer_blocks") or key.startswith("single_transformer_blocks"):
-        return f"transformer.{key}"
+        return f"diffusion_model.{key}"
 
     # Handle legacy lora_unet_ prefix (for resizing without base)
     if key.startswith("lora_unet_"):
-        core = key[10:]
-        dotted = _reconstruct_dots(core)
-        # Check if reconstructed key implies transformer format
-        if dotted.startswith("transformer_blocks") or dotted.startswith("single_transformer_blocks"):
-            return f"transformer.{dotted}"
-        return f"diffusion_model.{dotted}"
-
-    # Handle already correct prefixes
-    if key.startswith("diffusion_model.") or key.startswith("transformer."):
         return key
 
-    # Handle known Unet blocks (Diffusers format UNet)
+    # Handle already-prefixed keys
+    if key.startswith("diffusion_model."):
+        return key
+
+    # Absorb now-invalid transformer. prefix into diffusion_model. for compatibility
+    if key.startswith("transformer."):
+        return f"diffusion_model.{key[12:]}"
+
+    # Handle known Diffusers UNet blocks
     if any(key.startswith(p) for p in ["down_blocks", "up_blocks", "mid_block", "conv_in", "conv_out", "time_embedding", "class_embedding"]):
         return f"diffusion_model.{key}"
 
-    # Default fallback to diffusion_model
+    # Default fallback
     return f"diffusion_model.{key}"
 
 
+def _format_direct_diff_key(source_key: str, lora_name: str) -> str:
+    """Map a full delta to ComfyUI's canonical direct-patch key."""
+    if source_key.endswith(".bias"):
+        return f"{lora_name[:-5]}.diff_b"
+    return f"{lora_name}.diff"
+
+
 def _get_common_inputs():
     return [
         io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
@@ -734,10 +757,16 @@ def _get_common_inputs():
                       tooltip="Skip layers with max difference below this"),
         io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
         io.String.Input("output_filename", default="extracted_lora"),
-        io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
+        io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16",
+                       tooltip="Output dtype; layers stored as FP32 in either source remain FP32"),
         io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
         io.String.Input("skip_patterns", default="", multiline=True,
-                       tooltip="Regex patterns for layers to skip"),
+                       tooltip="Patterns for layers to skip (regex or glob depending on glob_skip_patterns)"),
+        io.Boolean.Input("glob_skip_patterns", default=False,
+                        tooltip="When True, skip_patterns use glob syntax (* = any sequence, ? = any char, dots are literal). "
+                                "When False (default), patterns are Python regex matched as substrings."),
+        io.Boolean.Input("include_1d_diffs", default=False,
+                        tooltip="Include 1D tensors as FP32 and keys ending in .scale or .lin as full .diff patches"),
     ]
 
 
@@ -768,19 +797,21 @@ class LoRAExtractFixed(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, linear_dim, conv_dim, svd_niter, chunk_large_layers,
                 clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache,
+                include_1d_diffs=False) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
 
-        output_sd = extract_lora_from_files(
+        extract_lora_from_files(
             model_a_path, model_b_path, "fixed", linear_dim, conv_dim,
-            device, save_dtype, linear_dim, conv_dim,
+            device, save_dtype, output_path, linear_dim, conv_dim,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers, svd_niter,
-            lazy_load, force_clear_cache
+            lazy_load, force_clear_cache, glob_skip_patterns, include_1d_diffs
         )
 
-        return io.NodeOutput(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)
 
 
 class LoRAExtractRatio(io.ComfyNode):
@@ -812,19 +843,22 @@ class LoRAExtractRatio(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, linear_ratio, conv_ratio, linear_max_rank, conv_max_rank,
                 chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache,
+                include_1d_diffs=False) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
 
-        output_sd = extract_lora_from_files(
+        extract_lora_from_files(
             model_a_path, model_b_path, "ratio", linear_ratio, conv_ratio,
-            device, save_dtype, linear_max_rank, conv_max_rank,
+            device, save_dtype, output_path, linear_max_rank, conv_max_rank,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
-            lazy_load=lazy_load, force_clear_cache=force_clear_cache
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache,
+            glob_skip_patterns=glob_skip_patterns, include_1d_diffs=include_1d_diffs
         )
 
-        return io.NodeOutput(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)
 
 
 class LoRAExtractQuantile(io.ComfyNode):
@@ -856,19 +890,22 @@ class LoRAExtractQuantile(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, linear_quantile, conv_quantile, linear_max_rank, conv_max_rank,
                 chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache,
+                include_1d_diffs=False) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
 
-        output_sd = extract_lora_from_files(
+        extract_lora_from_files(
             model_a_path, model_b_path, "quantile", linear_quantile, conv_quantile,
-            device, save_dtype, linear_max_rank, conv_max_rank,
+            device, save_dtype, output_path, linear_max_rank, conv_max_rank,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
-            lazy_load=lazy_load, force_clear_cache=force_clear_cache
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache,
+            glob_skip_patterns=glob_skip_patterns, include_1d_diffs=include_1d_diffs
         )
 
-        return io.NodeOutput(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)
 
 
 class LoRAExtractKnee(io.ComfyNode):
@@ -898,19 +935,22 @@ class LoRAExtractKnee(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, knee_method, linear_max_rank, conv_max_rank,
                 chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache,
+                include_1d_diffs=False) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
 
-        output_sd = extract_lora_from_files(
+        extract_lora_from_files(
             model_a_path, model_b_path, knee_method, 0, 0,
-            device, save_dtype, linear_max_rank, conv_max_rank,
+            device, save_dtype, output_path, linear_max_rank, conv_max_rank,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
-            lazy_load=lazy_load, force_clear_cache=force_clear_cache
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache,
+            glob_skip_patterns=glob_skip_patterns, include_1d_diffs=include_1d_diffs
         )
 
-        return io.NodeOutput(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)
 
 
 class LoRAExtractFrobenius(io.ComfyNode):
@@ -942,16 +982,19 @@ class LoRAExtractFrobenius(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, linear_target, conv_target, linear_max_rank, conv_max_rank,
                 chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache,
+                include_1d_diffs=False) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
+        output_path = _build_lora_output_path(output_filename)
 
-        output_sd = extract_lora_from_files(
+        extract_lora_from_files(
             model_a_path, model_b_path, "sv_fro", linear_target, conv_target,
-            device, save_dtype, linear_max_rank, conv_max_rank,
+            device, save_dtype, output_path, linear_max_rank, conv_max_rank,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
-            lazy_load=lazy_load, force_clear_cache=force_clear_cache
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache,
+            glob_skip_patterns=glob_skip_patterns, include_1d_diffs=include_1d_diffs
         )
 
-        return io.NodeOutput(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)

nodes/lora_rename.py

@@ -0,0 +1,119 @@
+import os
+import folder_paths
+import comfy.utils
+from tqdm import tqdm
+from comfy_api.latest import io
+from safetensors import safe_open
+from .utils import convert_pt_to_safetensors
+from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
+
+from unifiedefficientloader import MemoryEfficientSafeOpen, IncrementalSafetensorsWriter
+
+REVERSE_RENAME_DICT = {
+    "transformer_blocks.": "blocks.",
+    "norm1.linear_1": "adaln_modulation_self_attn.1",
+    "norm1.linear_2": "adaln_modulation_self_attn.2",
+    "norm2.linear_1": "adaln_modulation_cross_attn.1",
+    "norm2.linear_2": "adaln_modulation_cross_attn.2",
+    "norm3.linear_1": "adaln_modulation_mlp.1",
+    "norm3.linear_2": "adaln_modulation_mlp.2",
+    "attn1.to_q": "self_attn.q_proj",
+    "attn1.to_k": "self_attn.k_proj",
+    "attn1.to_v": "self_attn.v_proj",
+    "attn1.to_out.0": "self_attn.output_proj",
+    "attn2.to_q": "cross_attn.q_proj",
+    "attn2.to_k": "cross_attn.k_proj",
+    "attn2.to_v": "cross_attn.v_proj",
+    "attn2.to_out.0": "cross_attn.output_proj",
+    "ff.net.0.proj": "mlp.layer1",
+    "ff.net.2": "mlp.layer2",
+    "norm_out.linear_1": "final_layer.adaln_modulation.1",
+    "norm_out.linear_2": "final_layer.adaln_modulation.2",
+    "proj_out": "final_layer.linear",
+    "time_embed.t_embedder": "t_embedder.1",
+    "time_embed.norm": "t_embedding_norm",
+    "patch_embed.proj": "x_embedder.proj.1",
+}
+
+def _convert_diffusers_to_non_diffusers_anima_lora(lora_name: str, output_filename: str) -> str:
+    """
+    Converts a Diffusers Anima LoRA back to its non-diffusers format.
+    """
+    model_type = "loras"
+    model_path = folder_paths.get_full_path_or_raise(model_type, lora_name)
+
+    if model_path.endswith(('.pt', '.pth', '.bin', '.ckpt')):
+        temp_safe_path = model_path + ".safetensors"
+        if not os.path.exists(temp_safe_path):
+            conversion_successful, error_message = convert_pt_to_safetensors(model_path, temp_safe_path)
+            if not conversion_successful:
+                raise Exception(f"Conversion failed: {error_message}")
+        model_path_to_load = temp_safe_path
+    else:
+        model_path_to_load = model_path
+
+    # Prepare memory before operation
+    model_size_gb = estimate_model_size(model_path_to_load)
+    prepare_for_large_operation(model_size_gb * 1.2)
+
+    # Get metadata from safe_open (header only, no tensor loading)
+    with safe_open(model_path_to_load, framework="pt", device="cpu") as f:
+        metadata = f.metadata() or {}
+
+    model_dir = folder_paths.get_folder_paths(model_type)[-1]
+    output_path = os.path.join(model_dir, f"{output_filename.strip()}.safetensors")
+
+    # Stream tensors, rename on the fly, write immediately
+    writer = IncrementalSafetensorsWriter(output_path, metadata=metadata)
+    writer.__enter__()
+    try:
+        with MemoryEfficientSafeOpen(model_path_to_load) as handler:
+            original_keys = handler.keys()
+            pbar = comfy.utils.ProgressBar(len(original_keys))
+            for key in tqdm(original_keys, desc="Renaming Anima LoRA keys", unit="keys"):
+                new_key = key
+
+                if key.startswith("text_conditioner."):
+                    new_key = f"diffusion_model.llm_adapter.{key.removeprefix('text_conditioner.')}"
+                elif key.startswith("transformer."):
+                    new_key = key.removeprefix("transformer.")
+                    for old_part, new_part in REVERSE_RENAME_DICT.items():
+                        new_key = new_key.replace(old_part, new_part)
+                    new_key = f"diffusion_model.{new_key}"
+
+                writer.write(new_key, handler.get_tensor(key).contiguous())
+                pbar.update(1)
+    finally:
+        writer.__exit__(None, None, None)
+
+    # Cleanup after operation
+    cleanup_after_operation()
+
+    return output_path
+
+
+class AnimaLoraRename(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="AnimaLoraRename",
+            display_name="Anima LoRA Rename (Diffusers to Non-Diffusers)",
+            category="ModelUtils/LoRA",
+            description="Loads a Diffusers Anima LoRA, reverses its key names to the non-diffusers format, and saves it as a new safetensors file.",
+            inputs=[
+                io.Combo.Input(
+                    "lora_name",
+                    options=folder_paths.get_filename_list("loras"),
+                ),
+                io.String.Input("output_filename", default="anima_lora_non_diffusers"),
+            ],
+            outputs=[
+                io.String.Output(display_name="output_path"),
+            ],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, lora_name: str, output_filename: str) -> io.NodeOutput:
+        path = _convert_diffusers_to_non_diffusers_anima_lora(lora_name, output_filename)
+        return io.NodeOutput(path)

nodes/merger.py

@@ -1,3 +1,4 @@
+import fnmatch
 import os
 import re
 import torch
@@ -5,13 +6,15 @@ import folder_paths
 import comfy.utils
 from tqdm import tqdm
 from comfy_api.latest import io
-from safetensors.torch import save_file
 from .merger_ops import TWO_MODEL_MODES, THREE_MODEL_MODES, MissingTensorBehavior, MissingTensorError
-from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
 from .device_utils import (
     estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 )
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned, IncrementalSafetensorsWriter
+from .quantization_guard import inspect_low_bit_input, write_preserved_tensor
+from .lora_resize import is_direct_diff_key
+
 
 def load_documentation_from_file(filename):
     """Loads documentation from a markdown file in the ../docs/ directory."""
@@ -24,26 +27,38 @@ def load_documentation_from_file(filename):
         return f"# Documentation File Not Found\n\nPlease ensure `{filename}` exists in the `docs` directory of the custom node."
 
 
-def _compile_patterns(pattern_string):
-    """Compiles whitespace-separated regex patterns into a list of compiled patterns.
+def _compile_patterns(pattern_string, glob_mode=False):
+    """Compiles whitespace-separated patterns into a list.
 
-    Returns empty list if pattern_string is empty or whitespace-only.
-    Raises ValueError if any pattern is invalid regex.
+    In regex mode (default): returns compiled re.Pattern objects.
+      Raises ValueError if any pattern is invalid regex.
+    In glob mode: returns plain strings; fnmatch handles matching.
     """
     if not pattern_string or not pattern_string.strip():
         return []
 
     patterns = []
     for pattern in pattern_string.split():
-        try:
-            patterns.append(re.compile(pattern))
-        except re.error as e:
-            raise ValueError(f"Invalid regex pattern '{pattern}': {e}")
+        if not pattern:
+            continue
+        if glob_mode:
+            patterns.append(pattern)
+        else:
+            try:
+                patterns.append(re.compile(pattern))
+            except re.error as e:
+                raise ValueError(f"Invalid regex pattern '{pattern}': {e}")
     return patterns
 
 
-def _matches_any_pattern(key, patterns):
-    """Returns True if key matches any of the compiled regex patterns (substring search)."""
+def _matches_any_pattern(key, patterns, glob_mode=False):
+    """Returns True if key matches any pattern.
+
+    Glob mode: fnmatch substring match — dots are literal, * matches any sequence.
+    Regex mode: compiled re.Pattern substring search.
+    """
+    if glob_mode:
+        return any(fnmatch.fnmatch(key, f"*{p}*") for p in patterns)
     for pattern in patterns:
         if pattern.search(key):
             return True
@@ -76,6 +91,9 @@ class MergerLogic:
             process_device = recipe_params.get('device', 'cpu')
             print(f"[Merger] Preparing memory for {total_size_gb:.2f}GB merge operation...")
             prepare_for_large_operation(total_size_gb * 1.2, torch.device(process_device))
+            flush_threshold_bytes = int(total_size_gb * 1024**3 / 16)
+        else:
+            flush_threshold_bytes = 256 * 1024 * 1024  # 256MB fallback (min 8GB RAM assumed)
 
         handlers = {}
         for name in model_names.values():
@@ -89,6 +107,18 @@ class MergerLogic:
         all_keys = primary_handler.keys()
         metadata = primary_handler.metadata()
 
+        low_bit_keys_by_name = {}
+        if model_type == "loras":
+            try:
+                for name, handler in handlers.items():
+                    low_bit_keys_by_name[name] = inspect_low_bit_input(
+                        handler, f"LoRA ({name})", "Generic LoRA Merge"
+                    )
+            except Exception:
+                for handler in handlers.values():
+                    handler.__exit__(None, None, None)
+                raise
+
         # Convert mismatch_mode string to enum
         mismatch_mode_str = recipe_params.get('mismatch_mode', 'skip')
         mismatch_mode = MissingTensorBehavior(mismatch_mode_str)
@@ -96,8 +126,9 @@ class MergerLogic:
         alignment_mode = recipe_params.get('alignment_mode', 'pad/crop')
 
         # Compile filter patterns
-        exclude_patterns = _compile_patterns(recipe_params.get('exclude_patterns', ''))
-        discard_patterns = _compile_patterns(recipe_params.get('discard_patterns', ''))
+        glob_mode = recipe_params.get('glob_patterns', False)
+        exclude_patterns = _compile_patterns(recipe_params.get('exclude_patterns', ''), glob_mode=glob_mode)
+        discard_patterns = _compile_patterns(recipe_params.get('discard_patterns', ''), glob_mode=glob_mode)
 
         # Pre-compute key differences for logging
         primary_keys = set(all_keys)
@@ -115,6 +146,26 @@ class MergerLogic:
         pbar = comfy.utils.ProgressBar(len(all_keys))
         save_dtype = recipe_params.pop('save_dtype')
         save_torch_dtype = {"fp32": torch.float32, "fp16": torch.float16, "bf16": torch.bfloat16}.get(save_dtype)
+        override_dtype = recipe_params.pop('override_dtype', False)
+        include_1d_diffs = recipe_params.pop('include_1d_diffs', False)
+
+        def determine_target_dtype(original_dtypes, target_dtype, override):
+            if override:
+                return target_dtype
+            priority_map = {
+                torch.float32: 3,
+                torch.bfloat16: 2,
+                torch.float16: 2,
+            }
+            highest_dtype = target_dtype
+            highest_priority = priority_map.get(highest_dtype, 1)
+            for dtype in original_dtypes:
+                p = priority_map.get(dtype, 1)
+                if p > highest_priority:
+                    highest_priority = p
+                    highest_dtype = dtype
+            return highest_dtype
+
         recipe_params.update({"handlers": handlers})
         process_device = recipe_params.get('device', 'cpu')
         process_dtype = recipe_params.get('dtype', torch.float32)
@@ -124,16 +175,69 @@ class MergerLogic:
         discarded_keys = 0
         error_keys = []
 
+        output_folder = "loras" if calc_mode == "SVD LoRA Extraction" else model_type
+        output_dir = os.path.join(folder_paths.models_dir, output_folder)
+        os.makedirs(output_dir, exist_ok=True)
+        output_filename = recipe_params.get("output_filename")
+        output_path = os.path.join(output_dir, f"{output_filename}.safetensors")
+
+        writer = None
+
+        writer = IncrementalSafetensorsWriter(output_path, metadata=metadata)
+        writer.__enter__()
+
+        batch_buffer = {}
+        batch_bytes = 0
+        batch_key_count = 0
+
+        def _flush_batch():
+            nonlocal batch_buffer, batch_bytes, batch_key_count
+            if not batch_buffer:
+                return
+            writer.write_dict(batch_buffer)
+            pbar.update(batch_key_count)
+            batch_buffer = {}
+            batch_bytes = 0
+            batch_key_count = 0
+            if recipe_params.get('force_clear_cache', True):
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
         with torch.no_grad():
             for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
-                # Check discard patterns first - skip entirely
-                if _matches_any_pattern(key, discard_patterns):
-                    discarded_keys += 1
+                preserve_low_bit = model_type == "loras" and any(
+                    key in low_bit_keys for low_bit_keys in low_bit_keys_by_name.values()
+                )
+                if preserve_low_bit:
+                    _flush_batch()
+                    write_preserved_tensor(writer, key, primary_handler)
                     pbar.update(1)
                     continue
+                # Check discard patterns first - skip entirely
+                if _matches_any_pattern(key, discard_patterns, glob_mode=glob_mode):
+                    discarded_keys += 1
+                    continue
 
                 # Pre-load Model A's tensor with pinned memory for CUDA
                 cpu_tensor = primary_handler.get_tensor(key)
+
+                # Determine original dtypes across all active models for this key
+                original_dtypes = [primary_handler.get_dtype(key)]
+                for name, handler in handlers.items():
+                    if name != primary_model_name and key in handler.keys():
+                        original_dtypes.append(handler.get_dtype(key))
+
+                target_dtype = determine_target_dtype(original_dtypes, save_torch_dtype, override_dtype)
+                preserve_model_a_1d = model_type == "loras" and cpu_tensor.ndim == 1 and not include_1d_diffs
+                if preserve_model_a_1d:
+                    target_dtype = determine_target_dtype(
+                        [primary_handler.get_dtype(key)], save_torch_dtype, override_dtype
+                    )
+                elif model_type == "loras" and is_direct_diff_key(key) and cpu_tensor.ndim == 1:
+                    target_dtype = torch.float32
+
                 if process_device == 'cuda':
                     tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
                 else:
@@ -141,10 +245,17 @@ class MergerLogic:
                 del cpu_tensor
 
                 # Check exclude patterns - use Model A only, no merge
-                if _matches_any_pattern(key, exclude_patterns):
-                    merged_state_dict[key] = tensor_a.detach().to(save_torch_dtype).cpu().clone()
-                    excluded_keys += 1
-                    pbar.update(1)
+                if preserve_model_a_1d or _matches_any_pattern(key, exclude_patterns, glob_mode=glob_mode):
+                    t = tensor_a.detach().to(target_dtype).cpu()
+
+                    batch_buffer[key] = t
+                    batch_bytes += t.numel() * t.element_size()
+                    batch_key_count += 1
+                    if not preserve_model_a_1d:
+                        excluded_keys += 1
+                    del tensor_a
+                    if (batch_bytes >= flush_threshold_bytes or batch_key_count >= 32):
+                        _flush_batch()
                     continue
 
                 # Pass tensor_a metadata to recipes for zeros mode and fallback
@@ -157,6 +268,9 @@ class MergerLogic:
                     result = recipe.merge()
                 except MissingTensorError as e:
                     if mismatch_mode == MissingTensorBehavior.ERROR:
+
+                        _flush_batch()
+                        writer.__exit__(None, None, None)
                         raise ValueError(f"Layer mismatch error (mismatch_mode='error'): {e}")
                     result = None
                     error_keys.append(key)
@@ -168,7 +282,10 @@ class MergerLogic:
 
                 if isinstance(result, dict):
                     for r_key, r_tensor in result.items():
-                        merged_state_dict[r_key] = r_tensor.detach().to(save_torch_dtype).cpu().clone()
+                        t = r_tensor.detach().to(target_dtype).cpu()
+
+                        batch_buffer[r_key] = t
+                        batch_bytes += t.numel() * t.element_size()
                 else:
                     # Ensure compatibility with Model A's architecture.
                     # If alignment_mode is 'pad/crop', we crop results that were padded.
@@ -179,7 +296,13 @@ class MergerLogic:
                             slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(result.shape, target_shape))
                             result = result[slices]
 
-                    merged_state_dict[key] = result.detach().to(save_torch_dtype).cpu().clone()
+                    t = result.detach().to(target_dtype).cpu()
+
+                    batch_buffer[key] = t
+                    batch_bytes += t.numel() * t.element_size()
+
+
+                batch_key_count += 1
 
                 # Clean up references to allow GC immediately.
                 # Local loop variables must be explicitly deleted to prevent PyTorch from keeping tensors in VRAM.
@@ -189,13 +312,12 @@ class MergerLogic:
                 del recipe
                 del result
 
-                if recipe_params.get('force_clear_cache', True):
-                    import gc
-                    gc.collect()
-                    if torch.cuda.is_available():
-                        torch.cuda.empty_cache()
+                if (batch_bytes >= flush_threshold_bytes or batch_key_count >= 32):
+                    _flush_batch()
 
-                pbar.update(1)
+
+        _flush_batch()
+        writer.__exit__(None, None, None)
 
         # Log summary
         if excluded_keys > 0:
@@ -211,15 +333,13 @@ class MergerLogic:
             handler.__exit__(None, None, None)
 
         output_folder = "loras" if calc_mode == "SVD LoRA Extraction" else model_type
-        # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
-        output_dir = folder_paths.get_folder_paths(output_folder)[-1]
+        output_dir = os.path.join(folder_paths.models_dir, output_folder)
         os.makedirs(output_dir, exist_ok=True)
         output_filename = recipe_params.get("output_filename")
         output_path = os.path.join(output_dir, f"{output_filename}.safetensors")
-        save_file(merged_state_dict, output_path, metadata=metadata)
-
+        if os.path.exists(output_path):
+            print(f"[Merger] Output saved to {output_path}")
         # Cleanup after heavy operation
-        del merged_state_dict
         cleanup_after_operation()
 
         return f"{output_filename}.safetensors"
@@ -243,18 +363,24 @@ class CheckpointTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_checkpoint"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force the entire model to be saved as the selected save_dtype. If False (default), higher precision dtypes are preserved."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -265,9 +391,9 @@ class CheckpointTwoMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -275,10 +401,11 @@ class CheckpointTwoMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
@@ -302,18 +429,24 @@ class ModelTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_model"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force the entire model to be saved as the selected save_dtype. If False (default), higher precision dtypes are preserved."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -324,9 +457,9 @@ class ModelTwoMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -334,10 +467,11 @@ class ModelTwoMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
@@ -361,18 +495,24 @@ class TextEncoderTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_textencoder"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force the entire model to be saved as the selected save_dtype. If False (default), higher precision dtypes are preserved."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -383,9 +523,9 @@ class TextEncoderTwoMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -393,10 +533,11 @@ class TextEncoderTwoMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
@@ -420,18 +561,26 @@ class LoRATwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_lora"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force merged non-1D tensors to save_dtype. Enabled 1D direct diffs remain FP32."),
+                io.Boolean.Input("include_1d_diffs", default=False,
+                                 tooltip="Merge 1D tensors. When disabled, preserve Model A's 1D tensors unchanged."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -442,9 +591,10 @@ class LoRATwoMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool,
+                include_1d_diffs: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -452,11 +602,13 @@ class LoRATwoMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
+            "include_1d_diffs": include_1d_diffs,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -479,18 +631,24 @@ class EmbeddingTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_embedding"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force the entire model to be saved as the selected save_dtype. If False (default), higher precision dtypes are preserved."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -501,9 +659,9 @@ class EmbeddingTwoMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -511,10 +669,11 @@ class EmbeddingTwoMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
@@ -541,18 +700,24 @@ class CheckpointThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_checkpoint"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force the entire model to be saved as the selected save_dtype. If False (default), higher precision dtypes are preserved."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -563,9 +728,9 @@ class CheckpointThreeMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -573,10 +738,11 @@ class CheckpointThreeMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
@@ -601,18 +767,24 @@ class ModelThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_model"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force the entire model to be saved as the selected save_dtype. If False (default), higher precision dtypes are preserved."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -623,9 +795,9 @@ class ModelThreeMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -633,10 +805,11 @@ class ModelThreeMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
@@ -661,18 +834,24 @@ class TextEncoderThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_textencoder"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force the entire model to be saved as the selected save_dtype. If False (default), higher precision dtypes are preserved."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -683,9 +862,9 @@ class TextEncoderThreeMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -693,10 +872,11 @@ class TextEncoderThreeMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
@@ -721,18 +901,26 @@ class LoRAThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_lora"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force merged non-1D tensors to save_dtype. Enabled 1D direct diffs remain FP32."),
+                io.Boolean.Input("include_1d_diffs", default=False,
+                                 tooltip="Merge 1D tensors. When disabled, preserve Model A's 1D tensors unchanged."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -743,9 +931,10 @@ class LoRAThreeMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool,
+                include_1d_diffs: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -753,11 +942,13 @@ class LoRAThreeMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
+            "include_1d_diffs": include_1d_diffs,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
         filename = MergerLogic.execute_merge(model_names, calc_mode, THREE_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -781,18 +972,24 @@ class EmbeddingThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("delta", default=2.0, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("epsilon", default=0.01, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("zeta", default=0.0, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_embedding"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.Boolean.Input("glob_patterns", default=False,
+                                 tooltip="When True, exclude/discard patterns use glob syntax (* = any sequence, dots are literal). "
+                                         "When False (default), patterns are Python regex matched as substrings."),
                 io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
                 io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+                io.Boolean.Input("override_dtype", default=False, tooltip="Force the entire model to be saved as the selected save_dtype. If False (default), higher precision dtypes are preserved."),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -803,9 +1000,9 @@ class EmbeddingThreeMerger(io.ComfyNode):
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
-                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
+                gamma: float, delta: float, epsilon: float, zeta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str,
-                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -813,12 +1010,13 @@ class EmbeddingThreeMerger(io.ComfyNode):
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
             "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
+            "output_filename": output_filename, "save_dtype": save_dtype, "override_dtype": override_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "glob_patterns": glob_patterns,
             "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
         filename = MergerLogic.execute_merge(model_names, calc_mode, THREE_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
-        return io.NodeOutput(filename, doc)
+        return io.NodeOutput(filename, doc)

nodes/merger_ops.py

@@ -4,7 +4,7 @@ import torch.nn.functional as F
 from enum import Enum
 
 
-from .merger_utils import transfer_to_gpu_pinned
+from unifiedefficientloader import transfer_to_gpu_pinned
 
 class MissingTensorBehavior(Enum):
     """Controls behavior when a tensor key is missing from a model."""
@@ -319,6 +319,53 @@ class PowerUpOp(Operation):
         return (m * delta) / keep_prob
 
 
+class DARETIESOp(Operation):
+    """DARE + TIES combined operation on a single task vector (delta = b - a).
+
+    Step 1 — DARE: randomly drop `drop_rate` fraction of delta parameters,
+                   rescale survivors by 1/(1-drop_rate).
+    Step 2 — TIES trim: zero out parameters below the `trim_quantile` magnitude threshold.
+    Step 3 — TIES elect: determine dominant sign per parameter position.
+    Step 4 — TIES merge: zero out parameters that disagree with the dominant sign.
+
+    For a two-model merge the TIES elect/merge step degenerates to a no-op (single vector,
+    sign is always self-consistent) but the trim step still reduces interference from
+    near-zero noisy deltas. The dominant sign filter is included for correctness and
+    forward compatibility with n-model extensions.
+    """
+    def __init__(self, key, alpha, beta, gamma, seed, *sources):
+        super().__init__(key, *sources)
+        # alpha = drop_rate (DARE), beta = trim_quantile (TIES), gamma = lambda scale
+        self.alpha, self.beta, self.gamma, self.seed = alpha, beta, gamma, seed
+
+    def oper(self, a, b):
+        if a is None or b is None:
+            return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        delta = (b - a).float()
+
+        # --- DARE: drop and rescale ---
+        keep_prob = 1.0 - self.alpha
+        if keep_prob <= 0:
+            return torch.zeros_like(a)
+        rng = torch.Generator(device=delta.device).manual_seed(self.seed)
+        dare_mask = (torch.empty_like(delta).uniform_(0, 1, generator=rng) < keep_prob).float()
+        delta = (delta * dare_mask) / keep_prob
+
+        # --- TIES trim: zero parameters below magnitude threshold ---
+        if self.beta > 0:
+            flat_abs = delta.abs().flatten()
+            k = max(1, int(flat_abs.numel() * self.beta))
+            threshold = torch.kthvalue(flat_abs, k).values
+            delta = torch.where(delta.abs() < threshold, torch.zeros_like(delta), delta)
+
+        # --- TIES elect: dominant sign (trivially self-consistent for single vector) ---
+        dominant_sign = torch.sign(delta)
+        delta = torch.where(torch.sign(delta) == dominant_sign, delta, torch.zeros_like(delta))
+
+        return (a + self.gamma * delta).to(a.dtype)
+
+
 class InterpolateDifference(Operation):
     def __init__(self, key, alpha, beta, gamma, seed, *sources):
         super().__init__(key, *sources)
@@ -710,5 +757,179 @@ class WeightSumCutoffMode(CalcMode):
         return res
 
 
-TWO_MODEL_MODES = [WeightSum(), InterpDifference(), PowerUp(), SVDMode(), ManEnhInterpDifference(), AutoEnhInterpDifference(), WeightSumCutoffMode()]
+class DARETIESMode(CalcMode):
+    name = 'Power-Up (DARE+TIES)'
+    description = 'Adds capabilities of B to A using DARE (Drop and Rescale) followed by TIES (Trim, Elect Sign). Best practice for LoRA-tuned model merging.'
+    models_used = ['A', 'B']
+    param_docs = {
+        'alpha': 'DARE drop rate — fraction of delta parameters randomly zeroed. Higher = more aggressive sparsification.',
+        'beta':  'TIES trim quantile — fraction of smallest-magnitude delta parameters zeroed after DARE. Removes residual noise.',
+        'gamma': 'Lambda — scale multiplier applied to the final merged delta before adding to Model A.',
+        'seed':  'Random seed for the DARE dropout mask.',
+    }
+
+    def create_recipe(self, key, **kwargs):
+        mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
+        if '_tensor_a' in kwargs:
+            a = PreloadedTensor(key, kwargs['_tensor_a'])
+        else:
+            loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
+            a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
+
+        loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
+        b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
+
+        op = DARETIESOp(
+            key,
+            kwargs['alpha'],   # drop_rate
+            kwargs['beta'],    # trim_quantile
+            kwargs['gamma'],   # lambda scale
+            kwargs['seed'],
+            a, b,
+        )
+        op.alignment_mode = alignment_mode
+        return op
+
+
+class EnhancedPowerUpOp(Operation):
+    def __init__(self, key, alpha, beta, gamma, delta, seed, *sources):
+        super().__init__(key, *sources)
+        self.mask_power = alpha       # Mask power
+        self.min_keep_prob = beta     # Min keep prob
+        self.mask_smooth = gamma      # Smoothness (0 or 1)
+        self.delta = delta            # Final multiplier
+        self.seed = seed
+
+    def oper(self, a, b):
+        if a is None or b is None: return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        delta_val = (b - a).float()
+        abs_delta = torch.abs(delta_val)
+        max_delta = torch.max(abs_delta)
+        if max_delta == 0: return torch.zeros_like(delta_val)
+
+        prob = torch.clamp((abs_delta / max_delta) ** max(self.mask_power, 0.001), min=self.min_keep_prob, max=1.0)
+
+        rng = torch.Generator(device=a.device).manual_seed(self.seed)
+        random_mask = torch.bernoulli(prob, generator=rng)
+        interpolated_mask = torch.lerp(random_mask, prob, self.mask_smooth)
+
+        rescaled_delta = delta_val * interpolated_mask
+        return rescaled_delta
+
+
+class EnhancedDARETIESOp(Operation):
+    def __init__(self, key, alpha, beta, gamma, delta, epsilon, zeta, seed, *sources):
+        super().__init__(key, *sources)
+        self.alpha = alpha       # Unused directly for drop scale
+        self.beta = beta         # TIES trim quantile
+        self.gamma = gamma       # TIES lambda scale
+        self.mask_power = delta
+        self.min_keep_prob = epsilon
+        self.mask_smooth = zeta
+        self.seed = seed
+
+    def oper(self, a, b):
+        if a is None or b is None: return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        delta_val = (b - a).float()
+        abs_delta = torch.abs(delta_val)
+        max_delta = torch.max(abs_delta)
+
+        if max_delta > 0:
+            prob = torch.clamp((abs_delta / max_delta) ** max(self.mask_power, 0.001), min=self.min_keep_prob, max=1.0)
+
+            rng = torch.Generator(device=a.device).manual_seed(self.seed)
+            random_mask = torch.bernoulli(prob, generator=rng)
+            interpolated_mask = torch.lerp(random_mask, prob, self.mask_smooth)
+
+            delta_val = delta_val * interpolated_mask
+
+        if self.beta > 0:
+            flat_abs = delta_val.abs().flatten()
+            k = max(1, int(flat_abs.numel() * self.beta))
+            threshold = torch.kthvalue(flat_abs, k).values
+            delta_val = torch.where(delta_val.abs() < threshold, torch.zeros_like(delta_val), delta_val)
+
+        dominant_sign = torch.sign(delta_val)
+        delta_val = torch.where(torch.sign(delta_val) == dominant_sign, delta_val, torch.zeros_like(delta_val))
+
+        return (a + self.gamma * delta_val).to(a.dtype)
+
+
+class EnhancedPowerUp(CalcMode):
+    name = 'Power-Up Enhanced (DARE)'
+    description = 'Adds capabilities of B to A using Enhanced Drop and Rescale based on relative tensor differences.'
+    models_used = ['A', 'B']
+    param_docs = {
+        'alpha': 'Mask power (Curve. 2.0 = quadratic).',
+        'beta': 'Minimum keep probability (Floor to prevent explosion).',
+        'gamma': 'Mask smoothness factor (0.0 = pure dropout, 1.0 = soft scale).',
+        'delta': 'Multiplier for the final difference.',
+    }
+    def create_recipe(self, key, **kwargs):
+        mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
+        if '_tensor_a' in kwargs:
+            a = PreloadedTensor(key, kwargs['_tensor_a'])
+        else:
+            loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
+            a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
+
+        loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
+        b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
+
+        deltahat = EnhancedPowerUpOp(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'], kwargs['delta'], kwargs['seed'], a, b)
+        deltahat.alignment_mode = alignment_mode
+        multiplied = Multiply(key, kwargs['delta'], deltahat)
+        res = Add(key, a, multiplied)
+        res.alignment_mode = alignment_mode
+        return res
+
+
+class EnhancedDARETIESMode(CalcMode):
+    name = 'Power-Up Enhanced (DARE+TIES)'
+    description = 'Adds capabilities of B to A using Enhanced DARE mask followed by TIES.'
+    models_used = ['A', 'B']
+    param_docs = {
+        'alpha': 'Unused (Drop scale handled via mask interpolation).',
+        'beta': 'TIES trim quantile (fraction of smallest-magnitude delta zeroed).',
+        'gamma': 'TIES Lambda scale.',
+        'delta': 'Mask power (Curve. 2.0 = quadratic).',
+        'epsilon': 'Minimum keep probability (Floor to prevent explosion).',
+        'zeta': 'Mask smoothness factor (0.0 = pure dropout, 1.0 = soft scale).',
+    }
+
+    def create_recipe(self, key, **kwargs):
+        mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
+        if '_tensor_a' in kwargs:
+            a = PreloadedTensor(key, kwargs['_tensor_a'])
+        else:
+            loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
+            a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
+
+        loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
+        b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
+
+        op = EnhancedDARETIESOp(
+            key,
+            kwargs.get('alpha', 0.5),
+            kwargs.get('beta', 0.2),
+            kwargs.get('gamma', 1.0),
+            kwargs.get('delta', 0.5),
+            kwargs.get('epsilon', 0.5),
+            kwargs.get('zeta', 0.5),
+            kwargs['seed'],
+            a, b,
+        )
+        op.alignment_mode = alignment_mode
+        return op
+
+
+TWO_MODEL_MODES = [WeightSum(), InterpDifference(), PowerUp(), DARETIESMode(), SVDMode(), ManEnhInterpDifference(), AutoEnhInterpDifference(), WeightSumCutoffMode(), EnhancedPowerUp(), EnhancedDARETIESMode()]
 THREE_MODEL_MODES = [AddDifference(), TrainDifference(), Extract(), AddDisimilarity()]

nodes/merger_utils.py

@@ -1,292 +1,7 @@
 """
-Memory-efficient safetensors utilities with pinned memory support.
-Forces direct file reads to minimize OS page cache usage and avoids VRAM accumulation.
+Utilities for model merging operations.
 """
-import os
-import gc
-import json
-import struct
-import torch
-from safetensors import safe_open
-from typing import Dict, List, Optional, Any
-
-# Module-level configuration for pinned transfer
-_verbose = False
-_pinned_transfer_stats = {"pinned": 0, "fallback": 0}
-
-def set_verbose(enabled: bool):
-    """Enable/disable verbose output for pinned transfers."""
-    global _verbose
-    _verbose = enabled
-
-def get_pinned_transfer_stats():
-    """Return pinned transfer statistics."""
-    return _pinned_transfer_stats.copy()
-
-def reset_pinned_transfer_stats():
-    """Reset transfer statistics."""
-    global _pinned_transfer_stats
-    _pinned_transfer_stats = {"pinned": 0, "fallback": 0}
-
-def transfer_to_gpu_pinned(
-    tensor: torch.Tensor,
-    device: str = 'cuda',
-    dtype: Optional[torch.dtype] = None
-) -> torch.Tensor:
-    """Transfer tensor to GPU using pinned memory for faster transfer."""
-    global _pinned_transfer_stats
-
-    if tensor.device.type != 'cpu' or not torch.cuda.is_available():
-        if dtype is not None:
-            return tensor.to(device=device, dtype=dtype)
-        return tensor.to(device=device)
-
-    if not str(device).startswith('cuda'):
-        if dtype is not None:
-            return tensor.to(device=device, dtype=dtype)
-        return tensor.to(device=device)
-
-    try:
-        pinned = tensor.pin_memory()
-        if dtype is not None:
-            result = pinned.to(device=device, dtype=dtype, non_blocking=True)
-        else:
-            result = pinned.to(device=device, non_blocking=True)
-
-        torch.cuda.current_stream().synchronize()
-        _pinned_transfer_stats["pinned"] += 1
-        return result
-    except Exception as e:
-        _pinned_transfer_stats["fallback"] += 1
-        if dtype is not None:
-            return tensor.to(device=device, dtype=dtype)
-        return tensor.to(device=device)
-
-
-class MemoryEfficientSafeOpen:
-    """Memory-efficient safetensors file reader with streaming and pinned memory support.
-
-    Operates strictly in streaming mode (low memory). Tensors are read byte-by-byte from
-    their exact offsets without using safe_open or memory mapping, ensuring tensors do not
-    remain persistently in OS page caches or RAM.
-    """
-
-    def __init__(self, filename: str, device: str = 'cpu', **kwargs):
-        # Ignore mmap_mode, low_memory, or any other legacy kwargs. We always stream.
-        self.filename = filename
-        self.device = device
-        self.header, self.header_size = self._read_header()
-        self.file = open(filename, "rb")
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        self.file.close()
-
-    def keys(self) -> List[str]:
-        """Return all tensor keys (excluding metadata)."""
-        return [k for k in self.header.keys() if k != "__metadata__"]
-
-    def metadata(self) -> Dict[str, str]:
-        """Return file metadata."""
-        return self.header.get("__metadata__", {})
-
-    def get_shape(self, key: str) -> tuple:
-        """Get tensor shape without loading tensor data."""
-        if key not in self.header:
-            raise KeyError(f"Tensor '{key}' not found in file")
-        return tuple(self.header[key]["shape"])
-
-    def get_ndim(self, key: str) -> int:
-        """Get tensor ndim without loading tensor data."""
-        return len(self.get_shape(key))
-
-    def keys_sorted_by_offset(self) -> List[str]:
-        keys_with_offsets = []
-        for key in self.keys():
-            offset = self.header[key]["data_offsets"][0]
-            keys_with_offsets.append((key, offset))
-        keys_with_offsets.sort(key=lambda x: x[1])
-        return [k for k, _ in keys_with_offsets]
-
-    def get_tensor(self, key: str) -> torch.Tensor:
-        """Load a single tensor by key directly from file byte offset."""
-        if key not in self.header:
-            raise KeyError(f"Tensor '{key}' not found in the file")
-
-        metadata = self.header[key]
-        offset_start, offset_end = metadata["data_offsets"]
-
-        if offset_start != offset_end:
-            self.file.seek(self.header_size + 8 + offset_start)
-            tensor_bytes = bytearray(offset_end - offset_start)
-            self.file.readinto(tensor_bytes)
-        else:
-            tensor_bytes = None
-
-        return self._deserialize_tensor(tensor_bytes, metadata)
-
-    def get_tensor_to_gpu(
-        self,
-        key: str,
-        device: str = 'cuda',
-        dtype: Optional[torch.dtype] = None
-    ) -> torch.Tensor:
-        cpu_tensor = self.get_tensor(key)
-        return transfer_to_gpu_pinned(cpu_tensor, device, dtype)
-
-    def get_tensors_parallel(
-        self,
-        keys: Optional[List[str]] = None,
-        workers: Optional[int] = None,
-        sort_by_offset: bool = True
-    ) -> Dict[str, torch.Tensor]:
-        if keys is None:
-            keys = self.keys()
-
-        if workers is None:
-            try:
-                from .device_utils import get_optimal_workers, estimate_model_size
-                model_size = estimate_model_size(self.filename)
-                workers = get_optimal_workers(model_size)
-            except ImportError:
-                workers = min(4, os.cpu_count() or 4)
-
-        if sort_by_offset:
-            keys_with_offsets = [(k, self.header[k]["data_offsets"][0]) for k in keys]
-            keys_with_offsets.sort(key=lambda x: x[1])
-            keys = [k for k, _ in keys_with_offsets]
-
-        result = {}
-        with ThreadPoolExecutor(max_workers=workers) as executor:
-            futures = {
-                executor.submit(
-                    _read_tensor_from_file,
-                    self.filename,
-                    key,
-                    self.header[key],
-                    self.header_size
-                ): key
-                for key in keys
-            }
-
-            for future in as_completed(futures):
-                key = futures[future]
-                try:
-                    tensor = future.result()
-                    if tensor is not None:
-                        result[key] = tensor
-                except Exception as e:
-                    print(f"[MemoryEfficientSafeOpen] Failed to load {key}: {e}")
-
-        return result
-
-    def get_tensor_as_dict(self, key: str) -> Dict[str, Any]:
-        if key not in self.header:
-            raise KeyError(f"Tensor '{key}' not found")
-
-        metadata = self.header[key]
-        offset_start, offset_end = metadata["data_offsets"]
-
-        if offset_start == offset_end:
-            return {}
-
-        self.file.seek(self.header_size + 8 + offset_start)
-        tensor_bytes = self.file.read(offset_end - offset_start)
-        return json.loads(tensor_bytes.decode("utf-8"))
-
-    def _read_header(self):
-        with open(self.filename, "rb") as f:
-            header_size = struct.unpack("<Q", f.read(8))[0]
-            header_json = f.read(header_size).decode("utf-8")
-        return json.loads(header_json), header_size
-
-    def _deserialize_tensor(self, tensor_bytes, metadata):
-        dtype_str = metadata["dtype"]
-        shape = metadata["shape"]
-        dtype = self._get_torch_dtype(dtype_str)
-
-        if tensor_bytes is None:
-            byte_tensor = torch.empty(0, dtype=torch.uint8)
-        else:
-            byte_tensor = torch.frombuffer(tensor_bytes, dtype=torch.uint8)
-
-        if dtype_str in ["F8_E5M2", "F8_E4M3"]:
-            return self._convert_float8(byte_tensor, dtype_str, shape)
-
-        return byte_tensor.view(dtype).reshape(shape)
-
-    @staticmethod
-    def _get_torch_dtype(dtype_str):
-        dtype_map = {
-            "F64": torch.float64, "F32": torch.float32, "F16": torch.float16, "BF16": torch.bfloat16,
-            "I64": torch.int64, "I32": torch.int32, "I16": torch.int16, "I8": torch.int8,
-            "U8": torch.uint8, "BOOL": torch.bool,
-        }
-        if hasattr(torch, "float8_e5m2"):
-            dtype_map["F8_E5M2"] = torch.float8_e5m2
-        if hasattr(torch, "float8_e4m3fn"):
-            dtype_map["F8_E4M3"] = torch.float8_e4m3fn
-
-        dtype = dtype_map.get(dtype_str)
-        if dtype is None:
-            raise ValueError(f"Unsupported dtype: {dtype_str}")
-        return dtype
-
-    @staticmethod
-    def _convert_float8(byte_tensor, dtype_str, shape):
-        if dtype_str == "F8_E5M2" and hasattr(torch, "float8_e5m2"):
-            return byte_tensor.view(torch.float8_e5m2).reshape(shape)
-        elif dtype_str == "F8_E4M3" and hasattr(torch, "float8_e4m3fn"):
-            return byte_tensor.view(torch.float8_e4m3fn).reshape(shape)
-        else:
-            raise ValueError(f"Unsupported float8 type: {dtype_str}")
-
-
-def _read_tensor_from_file(
-    filepath: str,
-    key: str,
-    metadata: Dict,
-    header_size: int
-) -> Optional[torch.Tensor]:
-    """Helper function to read a single tensor from file (for parallel execution).
-
-    Each worker opens its own file handle for thread safety.
-    """
-    offset_start, offset_end = metadata["data_offsets"]
-    dtype_str = metadata["dtype"]
-    shape = metadata["shape"]
-
-    if offset_start == offset_end:
-        return None
-
-    dtype_map = {
-        "F64": torch.float64, "F32": torch.float32, "F16": torch.float16, "BF16": torch.bfloat16,
-        "I64": torch.int64, "I32": torch.int32, "I16": torch.int16, "I8": torch.int8,
-        "U8": torch.uint8, "BOOL": torch.bool,
-    }
-    if hasattr(torch, "float8_e5m2"):
-        dtype_map["F8_E5M2"] = torch.float8_e5m2
-    if hasattr(torch, "float8_e4m3fn"):
-        dtype_map["F8_E4M3"] = torch.float8_e4m3fn
-
-    dtype = dtype_map.get(dtype_str, torch.float32)
-
-    with open(filepath, "rb") as f:
-        f.seek(header_size + 8 + offset_start)
-        tensor_bytes = f.read(offset_end - offset_start)
-
-    byte_tensor = torch.frombuffer(bytearray(tensor_bytes), dtype=torch.uint8)
-
-    if dtype_str in ["F8_E5M2", "F8_E4M3"]:
-        if dtype_str == "F8_E5M2" and hasattr(torch, "float8_e5m2"):
-            return byte_tensor.view(torch.float8_e5m2).reshape(shape)
-        elif dtype_str == "F8_E4M3" and hasattr(torch, "float8_e4m3fn"):
-            return byte_tensor.view(torch.float8_e4m3fn).reshape(shape)
-
-    return byte_tensor.view(dtype).reshape(shape)
-
+from typing import Dict, List
 
 def get_union_keys(*handlers) -> List[str]:
     """Returns the union of all tensor keys across multiple model handlers."""
@@ -305,4 +20,4 @@ def get_intersection_keys(*handlers) -> List[str]:
     all_keys = set(handlers[0].keys())
     for handler in handlers[1:]:
         all_keys &= set(handler.keys())
-    return sorted(all_keys)
+    return sorted(all_keys)

nodes/prunekeys.py

@@ -4,12 +4,12 @@ import folder_paths
 import comfy.utils
 from tqdm import tqdm
 from comfy_api.latest import io
-from safetensors.torch import save_file
 from safetensors import safe_open
 from .utils import convert_pt_to_safetensors
-from .merger_utils import MemoryEfficientSafeOpen
 from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, IncrementalSafetensorsWriter
+
 
 def _prune_keys(model_name: str, model_type: str, keys_to_prune_str: str,
                 use_regex: bool, output_filename: str) -> str:
@@ -39,32 +39,33 @@ def _prune_keys(model_name: str, model_type: str, keys_to_prune_str: str,
     if not patterns:
         raise ValueError("No keys/patterns provided to prune.")
 
-    # Stream tensors and filter on the fly
-    pruned_tensors = {}
-    with MemoryEfficientSafeOpen(model_path_to_load) as handler:
-        all_keys = handler.keys()
-        pbar = comfy.utils.ProgressBar(len(all_keys))
-        for key in tqdm(all_keys, desc="Pruning keys", unit="keys"):
-            is_match = False
-            if use_regex:
-                if any(re.search(pattern, key) for pattern in patterns):
-                    is_match = True
-            else:
-                if any(pattern in key for pattern in patterns):
-                    is_match = True
-
-            if not is_match:
-                pruned_tensors[key] = handler.get_tensor(key)
-            pbar.update(1)
-
     # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
     model_dir = folder_paths.get_folder_paths(model_type)[-1]
     output_path = os.path.join(model_dir, f"{output_filename.strip()}.safetensors")
 
-    save_file(pruned_tensors, output_path, metadata)
+    # Stream tensors, filter on the fly, write immediately
+    writer = IncrementalSafetensorsWriter(output_path, metadata=metadata)
+    writer.__enter__()
+    try:
+        with MemoryEfficientSafeOpen(model_path_to_load) as handler:
+            all_keys = handler.keys()
+            pbar = comfy.utils.ProgressBar(len(all_keys))
+            for key in tqdm(all_keys, desc="Pruning keys", unit="keys"):
+                is_match = False
+                if use_regex:
+                    if any(re.search(pattern, key) for pattern in patterns):
+                        is_match = True
+                else:
+                    if any(pattern in key for pattern in patterns):
+                        is_match = True
+
+                if not is_match:
+                    writer.write(key, handler.get_tensor(key).contiguous())
+                pbar.update(1)
+    finally:
+        writer.__exit__(None, None, None)
 
     # Cleanup after operation
-    del pruned_tensors
     cleanup_after_operation()
 
     return output_path

nodes/quantization_guard.py

@@ -0,0 +1,145 @@
+"""Fail-safe detection for quantized or isolated low-bit safetensors inputs."""
+
+import logging
+import json
+import struct
+
+import torch
+
+
+MAX_ISOLATED_LOW_BIT_TENSORS = 2
+
+
+def _low_bit_dtypes():
+    names = (
+        "int8",
+        "uint8",
+        "float4_e2m1fn_x2",
+        "float8_e4m3fn",
+        "float8_e4m3fnuz",
+        "float8_e5m2",
+        "float8_e5m2fnuz",
+        "float8_e8m0fnu",
+    )
+    return {getattr(torch, name) for name in names if hasattr(torch, name)}
+
+
+LOW_BIT_DTYPES = _low_bit_dtypes()
+LOW_BIT_STORAGE_CODES = {"I8", "U8", "F4"}
+UPSTREAM_WRITER_STORAGE_CODES = {
+    "F64", "F32", "F16", "BF16", "I64", "I32", "I16", "I8", "U8",
+    "U64", "U32", "U16", "BOOL", "C64", "F8_E5M2", "F8_E4M3",
+}
+
+
+def _header(handler):
+    if getattr(handler, "_header", None) is not None:
+        return handler._header
+    with open(handler.filename, "rb") as source:
+        header_size = struct.unpack("<Q", source.read(8))[0]
+        return json.loads(source.read(header_size).decode("utf-8"))
+
+
+def storage_dtype_code(handler, key: str) -> str | None:
+    entry = _header(handler).get(key)
+    return entry.get("dtype") if entry else None
+
+
+def _is_low_bit(handler, key: str) -> bool:
+    storage_code = storage_dtype_code(handler, key)
+    if storage_code in LOW_BIT_STORAGE_CODES or (
+        storage_code is not None and storage_code.startswith("F8_")
+    ):
+        return True
+    return handler.get_dtype(key) in LOW_BIT_DTYPES
+
+
+def inspect_low_bit_input(handler, label: str, operation: str) -> set[str]:
+    """Reject quantized models; return isolated low-bit keys that must not be transformed."""
+    keys = list(handler.keys())
+    metadata = handler.metadata() or {}
+    quant_keys = [key for key in keys if key.endswith(".comfy_quant")]
+    has_quant_metadata = "_quantization_metadata" in metadata
+    has_legacy_scaled_fp8 = any(key.endswith("scaled_fp8") for key in keys)
+    if quant_keys or has_quant_metadata or has_legacy_scaled_fp8:
+        raise ValueError(
+            f"[{operation}] {label} uses ComfyUI quantization metadata; quantized inputs are unsupported"
+        )
+
+    low_bit_keys = {key for key in keys if _is_low_bit(handler, key)}
+    if len(low_bit_keys) > MAX_ISOLATED_LOW_BIT_TENSORS:
+        raise ValueError(
+            f"[{operation}] {label} contains {len(low_bit_keys)} low-bit tensors; "
+            "casted or quantized inputs are unsupported"
+        )
+    if low_bit_keys:
+        logging.warning(
+            "[%s] %s contains %d isolated low-bit tensor(s); "
+            "preserving/excluding unchanged: %s",
+            operation,
+            label,
+            len(low_bit_keys),
+            ", ".join(sorted(low_bit_keys)),
+        )
+    return low_bit_keys
+
+
+def layer_has_low_bit(block_keys: dict, low_bit_keys: set[str]) -> bool:
+    return any(
+        block_keys.get(name) in low_bit_keys
+        for name in (
+            "down", "up", "alpha", "mid", "reshape", "dora_scale",
+            "diff", "diff_b", "w_norm", "b_norm", "set_weight",
+        )
+    )
+
+
+def write_preserved_tensor(writer, key: str, handler, output_key: str | None = None) -> None:
+    """Copy one tensor's original safetensors bytes and dtype into an active writer."""
+    destination = output_key or key
+    header = _header(handler)
+    entry = header[key]
+    if entry["dtype"] in UPSTREAM_WRITER_STORAGE_CODES:
+        writer.write(destination, handler.get_tensor(key).cpu().contiguous())
+        return
+
+    source_start, source_end = entry["data_offsets"]
+    byte_size = source_end - source_start
+
+    with writer._lock:
+        if destination in writer._manifest:
+            raise ValueError(f"Tensor '{destination}' has already been written.")
+        target_start = writer._current_data_offset
+        target_end = target_start + byte_size
+        writer._manifest[destination] = {
+            "dtype": entry["dtype"],
+            "shape": entry["shape"],
+            "data_offsets": [target_start, target_end],
+        }
+        writer._current_data_offset = target_end
+
+    source_path = handler.filename
+    source_absolute = 8 + getattr(handler, "_header_size", 0) + source_start
+    if not getattr(handler, "_header_size", None):
+        with open(source_path, "rb") as source:
+            source_absolute = 8 + struct.unpack("<Q", source.read(8))[0] + source_start
+    target_absolute = 8 + writer.max_header_bytes + target_start
+
+    def copy_raw():
+        try:
+            with open(source_path, "rb") as source:
+                source.seek(source_absolute)
+                remaining = byte_size
+                with writer._lock:
+                    writer._file.seek(target_absolute)
+                    while remaining:
+                        chunk = source.read(min(8 * 1024 * 1024, remaining))
+                        if not chunk:
+                            raise EOFError(f"Unexpected EOF while preserving tensor '{key}'")
+                        writer._file.write(chunk)
+                        remaining -= len(chunk)
+        finally:
+            writer._semaphore.release()
+
+    writer._semaphore.acquire()
+    writer._futures.append(writer._executor.submit(copy_raw))

nodes/renamekeys.py

@@ -3,12 +3,13 @@ import folder_paths
 import comfy.utils
 from tqdm import tqdm
 from comfy_api.latest import io
-from safetensors.torch import save_file
 from safetensors import safe_open
 from .utils import convert_pt_to_safetensors
-from .merger_utils import MemoryEfficientSafeOpen
 from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, IncrementalSafetensorsWriter
+
+
 
 def _rename_keys(model_name: str, model_type: str, old_keys_str: str,
                  new_keys_str: str, output_filename: str) -> str:
@@ -41,24 +42,25 @@ def _rename_keys(model_name: str, model_type: str, old_keys_str: str,
 
     key_map = dict(zip(old_keys, new_keys))
 
-    # Stream tensors and rename on the fly
-    renamed_tensors = {}
-    with MemoryEfficientSafeOpen(model_path_to_load) as handler:
-        original_keys = handler.keys()
-        pbar = comfy.utils.ProgressBar(len(original_keys))
-        for key in tqdm(original_keys, desc="Renaming keys", unit="keys"):
-            new_key = key_map.get(key, key)
-            renamed_tensors[new_key] = handler.get_tensor(key)
-            pbar.update(1)
-
     # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
     model_dir = folder_paths.get_folder_paths(model_type)[-1]
     output_path = os.path.join(model_dir, f"{output_filename.strip()}.safetensors")
 
-    save_file(renamed_tensors, output_path, metadata)
+    # Stream tensors, rename on the fly, write immediately
+    writer = IncrementalSafetensorsWriter(output_path, metadata=metadata)
+    writer.__enter__()
+    try:
+        with MemoryEfficientSafeOpen(model_path_to_load) as handler:
+            original_keys = handler.keys()
+            pbar = comfy.utils.ProgressBar(len(original_keys))
+            for key in tqdm(original_keys, desc="Renaming keys", unit="keys"):
+                new_key = key_map.get(key, key)
+                writer.write(new_key, handler.get_tensor(key).contiguous())
+                pbar.update(1)
+    finally:
+        writer.__exit__(None, None, None)
 
     # Cleanup after operation
-    del renamed_tensors
     cleanup_after_operation()
 
     return output_path

nodes/text_encoder_extract.py

@@ -0,0 +1,721 @@
+"""
+Text Encoder LoRA and DoRA Extraction.
+
+Specifically targets extracting text encoder deltas as LoRA/DoRA with custom mapping rules
+for layer name prefixes and PEFT weight formatting.
+"""
+import fnmatch
+import os
+import re
+import torch
+import torch.linalg as linalg
+import folder_paths
+import comfy.utils
+from tqdm import tqdm
+from comfy_api.latest import io
+from .device_utils import (
+    estimate_model_size, prepare_for_large_operation,
+    cleanup_after_operation
+)
+
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned, IncrementalSafetensorsWriter
+
+# Import SVD rank utilities from standard lora extract file to avoid duplication
+from .lora_extract_svd import (
+    _compute_rank,
+    _svd_extract_linear,
+    _svd_extract_conv,
+    _extract_chunked_layer,
+    _detect_fused_layer,
+    _compile_patterns,
+    _matches_any_pattern
+)
+
+
+def _format_te_lora_key(key: str) -> str:
+    """
+    Format key specifically for Text Encoder LoRA/DoRA saving.
+    Corrects layer name prefixing based on specific target mappings.
+    """
+    if key.endswith(".weight"):
+        key = key[:-7]
+
+    rules = [
+        ("model.language_model.layers", "text_encoders.transformer.model.layers"),
+        ("model.visual.blocks", "text_encoders.transformer.visual.blocks"),
+        ("model.visual.deepstack_merger_list", "text_encoders.transformer.visual.deepstack_merger_list"),
+        ("model.visual.merger", "text_encoders.transformer.visual.merger"),
+        ("model.visual.pos_embed", "text_encoders.transformer.visual.pos_embed"),
+        ("model.layers", "text_encoders.transformer.model.layers"),
+        ("vision_model.embeddings.position_embedding", "text_encoders.transformer.vision_model.embeddings.position_embedding"),
+        ("vision_model.encoder.layers", "text_encoders.transformer.vision_model.encoder.layers"),
+    ]
+
+    for prefix, target in rules:
+        if key.startswith(prefix):
+            return target + key[len(prefix):]
+
+    # Prepend text_encoders. as default generic fallback
+    if key.startswith("text_encoders."):
+        return key
+    return f"text_encoders.{key}"
+
+
+def extract_te_from_files(
+    model_a_path: str,
+    model_b_path: str,
+    mode: str,
+    linear_param: float,
+    conv_param: float,
+    device: str,
+    save_dtype: str,
+    output_path: str,
+    is_dora: bool = False,
+    linear_max_rank: int = None,
+    conv_max_rank: int = None,
+    clamp_quantile: float = 0.99,
+    min_diff: float = 0.0,
+    skip_patterns_str: str = "",
+    mismatch_mode: str = "skip",
+    chunk_large_layers: bool = True,
+    svd_niter: int = 2,
+    lazy_load: bool = True,
+    force_clear_cache: bool = True,
+    glob_skip_patterns: bool = False,
+) -> None:
+    """
+    Extract LoRA/DoRA from difference between two Text Encoder models, writing incrementally to disk.
+    """
+    save_torch_dtype = {
+        "fp32": torch.float32,
+        "fp16": torch.float16,
+        "bf16": torch.bfloat16,
+    }.get(save_dtype, torch.float16)
+
+    skip_patterns = _compile_patterns(skip_patterns_str, glob_mode=glob_skip_patterns)
+
+    # Prepare memory before heavy operation
+    total_size_gb = estimate_model_size(model_a_path) + estimate_model_size(model_b_path)
+    print(f"[TE Extract] Preparing memory for {total_size_gb:.2f}GB operation...")
+    prepare_for_large_operation(total_size_gb * 1.5, torch.device(device))
+
+    handler_a = MemoryEfficientSafeOpen(model_a_path, low_memory=lazy_load)
+    handler_b = MemoryEfficientSafeOpen(model_b_path, low_memory=lazy_load)
+
+    try:
+        keys_a = set(handler_a.keys())
+        keys_b = set(handler_b.keys())
+        weight_keys = [k for k in keys_a if k.endswith(".weight")]
+        pbar = comfy.utils.ProgressBar(len(weight_keys))
+        stats = {"extracted": 0, "full": 0, "skipped": 0, "chunked": 0}
+
+        def _process_layer(key):
+            lora_name = _format_te_lora_key(key)
+
+            if _matches_any_pattern(key, skip_patterns, glob_mode=glob_skip_patterns):
+                return "skipped", None
+
+            # Load tensors with pinned memory for CUDA
+            use_pinned = device == 'cuda'
+
+            if key not in keys_b:
+                if mismatch_mode == "skip":
+                    return "skipped", None
+                if mismatch_mode == "error":
+                    raise ValueError(f"Key {key} not found in model B")
+
+                cpu_a = handler_a.get_tensor(key)
+                if use_pinned:
+                    weight_diff = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
+                else:
+                    weight_diff = cpu_a.to(device=device, dtype=torch.float32)
+                del cpu_a
+
+                dora_scale = None
+            else:
+                cpu_a = handler_a.get_tensor(key)
+                cpu_b = handler_b.get_tensor(key)
+
+                if use_pinned:
+                    tensor_a = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
+                    tensor_b = transfer_to_gpu_pinned(cpu_b, device, torch.float32)
+                else:
+                    tensor_a = cpu_a.to(device=device, dtype=torch.float32)
+                    tensor_b = cpu_b.to(device=device, dtype=torch.float32)
+                del cpu_a, cpu_b
+
+                if tensor_a.shape != tensor_b.shape:
+                    if mismatch_mode == "skip":
+                        del tensor_a, tensor_b
+                        return "skipped", None
+                    if mismatch_mode == "error":
+                        raise ValueError(f"Shape mismatch for {key}: {tensor_a.shape} vs {tensor_b.shape}")
+
+                    weight_diff = tensor_a
+                    dora_scale = None
+                    del tensor_b
+                else:
+                    # Skip 1D weights
+                    if tensor_a.ndim < 2:
+                        del tensor_a, tensor_b
+                        return "skipped", None
+
+                    if is_dora:
+                        # Weight Decomposition calculation
+                        is_conv_dim = tensor_a.ndim == 4
+
+                        if is_conv_dim:
+                            out_ch = tensor_a.shape[0]
+                            flat_a = tensor_a.reshape(out_ch, -1)
+                            flat_b = tensor_b.reshape(out_ch, -1)
+                            norm_a = torch.linalg.norm(flat_a, dim=1, keepdim=True)
+                            norm_b = torch.linalg.norm(flat_b, dim=1, keepdim=True)
+                        else:
+                            norm_a = torch.linalg.norm(tensor_a, dim=1, keepdim=True)
+                            norm_b = torch.linalg.norm(tensor_b, dim=1, keepdim=True)
+
+                        # Compute directional diff
+                        eps = 1e-8
+                        norm_ratio = norm_b / (norm_a + eps)
+                        if is_conv_dim:
+                            norm_ratio = norm_ratio.view(-1, 1, 1, 1)
+                            dora_scale = norm_a.view(-1, 1, 1, 1)
+                        else:
+                            dora_scale = norm_a
+
+                        w_dir = tensor_a * norm_ratio
+                        weight_diff = w_dir - tensor_b
+                        del tensor_a, tensor_b, w_dir
+                    else:
+                        weight_diff = tensor_a - tensor_b
+                        dora_scale = None
+                        del tensor_a, tensor_b
+
+            # Skip small differences
+            if min_diff > 0 and weight_diff.abs().max() < min_diff:
+                del weight_diff
+                return "skipped", None
+
+            # Skip 1D tensors
+            if weight_diff.ndim < 2:
+                del weight_diff
+                return "skipped", None
+
+            is_conv = weight_diff.ndim == 4
+            layer_results = {}
+
+            try:
+                if is_conv:
+                    result, mode_str = _svd_extract_conv(
+                        weight_diff, mode, conv_param, device, conv_max_rank, clamp_quantile
+                    )
+                else:
+                    result, mode_str = _svd_extract_linear(
+                        weight_diff, mode, linear_param, device, linear_max_rank, clamp_quantile, svd_niter
+                    )
+            except Exception as e:
+                # Try chunked extraction for large tensors
+                if chunk_large_layers and not is_conv:
+                    num_chunks = _detect_fused_layer(key, weight_diff.shape)
+                    if num_chunks > 1:
+                        print(f"[TE Extract] Chunked: {key} ({num_chunks} chunks)")
+                        lora_up, lora_down, rank = _extract_chunked_layer(
+                            weight_diff, num_chunks, mode, linear_param, device, linear_max_rank
+                        )
+                        if lora_up is not None:
+                            # PEFT suffixes
+                            layer_results[f"{lora_name}.lora_B.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.lora_A.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                            if is_dora and dora_scale is not None:
+                                layer_results[f"{lora_name}.dora_scale"] = dora_scale.to(save_torch_dtype).cpu().contiguous()
+                            del weight_diff
+                            return "chunked", layer_results
+
+                print(f"[TE Extract] Failed: {key}: {e}")
+                del weight_diff
+                return "skipped", None
+
+            # Store result
+            if mode_str == "full":
+                layer_results[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
+                status = "full"
+            else:
+                lora_down, lora_up, _ = result
+                # Write to PEFT weight suffixes
+                layer_results[f"{lora_name}.lora_A.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.lora_B.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                if is_dora and dora_scale is not None:
+                    layer_results[f"{lora_name}.dora_scale"] = dora_scale.to(save_torch_dtype).cpu().contiguous()
+                status = "extracted"
+
+            del weight_diff
+            return status, layer_results
+
+        writer = IncrementalSafetensorsWriter(output_path)
+        writer.__enter__()
+        try:
+            for key in tqdm(weight_keys, desc="Extracting TE Layers", unit="layers"):
+                status, layer_sd = _process_layer(key)
+                stats[status] += 1
+                if layer_sd:
+                    writer.write_dict(layer_sd)
+
+                if force_clear_cache:
+                    import gc
+                    gc.collect()
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+
+                pbar.update(1)
+        finally:
+            writer.__exit__(None, None, None)
+
+        print(f"[TE Extract] Done: {stats['extracted']} extracted, {stats['chunked']} chunked, "
+              f"{stats['full']} full, {stats['skipped']} skipped")
+
+    finally:
+        handler_a.__exit__(None, None, None)
+        handler_b.__exit__(None, None, None)
+        cleanup_after_operation()
+
+
+def _build_lora_output_path(output_filename: str) -> str:
+    """Build output path for LoRA file."""
+    output_dir = folder_paths.get_folder_paths("loras")[0]
+    os.makedirs(output_dir, exist_ok=True)
+    return os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
+
+
+def _get_te_model_inputs():
+    return [
+        io.Combo.Input("model_a", options=folder_paths.get_filename_list("text_encoders"),
+                      tooltip="Finetuned Text Encoder model (A - B = LoRA)"),
+        io.Combo.Input("model_b", options=folder_paths.get_filename_list("text_encoders"),
+                      tooltip="Base Text Encoder model (A - B = LoRA)"),
+    ]
+
+
+def _get_common_inputs():
+    return [
+        io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+        io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
+        io.Boolean.Input("chunk_large_layers", default=False,
+                        tooltip="Split large fused layers (QKV, MLP) into chunks"),
+        io.Float.Input("clamp_quantile", default=0.99, min=0.5, max=1.0, step=0.01,
+                      tooltip="Clamp outlier singular values"),
+        io.Float.Input("min_diff", default=0.0, min=0.0, max=1.0, step=0.001,
+                      tooltip="Skip layers with max difference below this"),
+        io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+        io.String.Input("output_filename", default="extracted_te_lora"),
+        io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
+        io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+        io.String.Input("skip_patterns", default="", multiline=True,
+                       tooltip="Patterns for layers to skip (regex or glob depending on glob_skip_patterns)"),
+        io.Boolean.Input("glob_skip_patterns", default=False,
+                        tooltip="When True, skip_patterns use glob syntax (* = any sequence, ? = any char, dots are literal). "
+                                "When False (default), patterns are Python regex matched as substrings."),
+    ]
+
+
+# =============================================================================
+# LoRA Text Encoder Extract Nodes
+# =============================================================================
+
+class TextEncoderLoRAExtractFixed(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderLoRAExtractFixed",
+            display_name="TE LoRA Extract (Fixed Rank)",
+            category="ModelUtils/LoRA Extract (TE)",
+            description="Extract Text Encoder LoRA with specified fixed rank for each layer type.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Int.Input("linear_dim", default=64, min=1, max=16384, tooltip="Rank for linear/attention layers"),
+                io.Int.Input("conv_dim", default=32, min=1, max=16384, tooltip="Rank for conv layers"),
+                io.Int.Input("svd_niter", default=2, min=0, max=10, tooltip="SVD power iterations"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_dim, conv_dim, svd_niter, chunk_large_layers,
+                clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, "fixed", linear_dim, conv_dim,
+            device, save_dtype, output_path, is_dora=False,
+            linear_max_rank=linear_dim, conv_max_rank=conv_dim,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers, svd_niter=svd_niter,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+class TextEncoderLoRAExtractRatio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderLoRAExtractRatio",
+            display_name="TE LoRA Extract (Ratio)",
+            category="ModelUtils/LoRA Extract (TE)",
+            description="Keep singular values > max(S) / ratio.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Float.Input("linear_ratio", default=2.0, min=1.0, max=100.0, step=0.1),
+                io.Float.Input("conv_ratio", default=2.0, min=1.0, max=100.0, step=0.1),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_ratio, conv_ratio, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, "ratio", linear_ratio, conv_ratio,
+            device, save_dtype, output_path, is_dora=False,
+            linear_max_rank=linear_max_rank, conv_max_rank=conv_max_rank,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+class TextEncoderLoRAExtractQuantile(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderLoRAExtractQuantile",
+            display_name="TE LoRA Extract (Quantile)",
+            category="ModelUtils/LoRA Extract (TE)",
+            description="Keep enough singular values to reach target percentage.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Float.Input("linear_quantile", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Float.Input("conv_quantile", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_quantile, conv_quantile, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, "quantile", linear_quantile, conv_quantile,
+            device, save_dtype, output_path, is_dora=False,
+            linear_max_rank=linear_max_rank, conv_max_rank=conv_max_rank,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+class TextEncoderLoRAExtractKnee(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderLoRAExtractKnee",
+            display_name="TE LoRA Extract (Knee)",
+            category="ModelUtils/LoRA Extract (TE)",
+            description="Automatically find optimal rank using knee detection on singular value curve.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Combo.Input("knee_method", options=["sv_knee", "sv_cumulative_knee"], default="sv_knee"),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, knee_method, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, knee_method, 0, 0,
+            device, save_dtype, output_path, is_dora=False,
+            linear_max_rank=linear_max_rank, conv_max_rank=conv_max_rank,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+class TextEncoderLoRAExtractFrobenius(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderLoRAExtractFrobenius",
+            display_name="TE LoRA Extract (Frobenius)",
+            category="ModelUtils/LoRA Extract (TE)",
+            description="Preserve target fraction of Frobenius norm.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Float.Input("linear_target", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Float.Input("conv_target", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_target, conv_target, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, "sv_fro", linear_target, conv_target,
+            device, save_dtype, output_path, is_dora=False,
+            linear_max_rank=linear_max_rank, conv_max_rank=conv_max_rank,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+# =============================================================================
+# DoRA Text Encoder Extract Nodes
+# =============================================================================
+
+class TextEncoderDoRAExtractFixed(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderDoRAExtractFixed",
+            display_name="TE DoRA Extract (Fixed Rank)",
+            category="ModelUtils/DoRA Extract (TE)",
+            description="Extract Text Encoder DoRA with specified fixed rank for each layer type.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Int.Input("linear_dim", default=64, min=1, max=16384, tooltip="Rank for linear/attention layers"),
+                io.Int.Input("conv_dim", default=32, min=1, max=16384, tooltip="Rank for conv layers"),
+                io.Int.Input("svd_niter", default=2, min=0, max=10, tooltip="SVD power iterations"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_dim, conv_dim, svd_niter, chunk_large_layers,
+                clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, "fixed", linear_dim, conv_dim,
+            device, save_dtype, output_path, is_dora=True,
+            linear_max_rank=linear_dim, conv_max_rank=conv_dim,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers, svd_niter=svd_niter,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+class TextEncoderDoRAExtractRatio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderDoRAExtractRatio",
+            display_name="TE DoRA Extract (Ratio)",
+            category="ModelUtils/DoRA Extract (TE)",
+            description="Keep singular values > max(S) / ratio.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Float.Input("linear_ratio", default=2.0, min=1.0, max=100.0, step=0.1),
+                io.Float.Input("conv_ratio", default=2.0, min=1.0, max=100.0, step=0.1),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_ratio, conv_ratio, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, "ratio", linear_ratio, conv_ratio,
+            device, save_dtype, output_path, is_dora=True,
+            linear_max_rank=linear_max_rank, conv_max_rank=conv_max_rank,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+class TextEncoderDoRAExtractQuantile(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderDoRAExtractQuantile",
+            display_name="TE DoRA Extract (Quantile)",
+            category="ModelUtils/DoRA Extract (TE)",
+            description="Keep enough singular values to reach target percentage.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Float.Input("linear_quantile", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Float.Input("conv_quantile", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_quantile, conv_quantile, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, "quantile", linear_quantile, conv_quantile,
+            device, save_dtype, output_path, is_dora=True,
+            linear_max_rank=linear_max_rank, conv_max_rank=conv_max_rank,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+class TextEncoderDoRAExtractKnee(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderDoRAExtractKnee",
+            display_name="TE DoRA Extract (Knee)",
+            category="ModelUtils/DoRA Extract (TE)",
+            description="Automatically find optimal rank using knee detection on singular value curve.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Combo.Input("knee_method", options=["sv_knee", "sv_cumulative_knee"], default="sv_knee"),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, knee_method, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, knee_method, 0, 0,
+            device, save_dtype, output_path, is_dora=True,
+            linear_max_rank=linear_max_rank, conv_max_rank=conv_max_rank,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)
+
+
+class TextEncoderDoRAExtractFrobenius(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncoderDoRAExtractFrobenius",
+            display_name="TE DoRA Extract (Frobenius)",
+            category="ModelUtils/DoRA Extract (TE)",
+            description="Preserve target fraction of Frobenius norm.",
+            inputs=[
+                *_get_te_model_inputs(),
+                io.Float.Input("linear_target", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Float.Input("conv_target", default=0.9, min=0.0, max=1.0, step=0.01),
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_target, conv_target, linear_max_rank, conv_max_rank,
+                chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, glob_skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+        output_path = _build_lora_output_path(output_filename)
+
+        extract_te_from_files(
+            model_a_path, model_b_path, "sv_fro", linear_target, conv_target,
+            device, save_dtype, output_path, is_dora=True,
+            linear_max_rank=linear_max_rank, conv_max_rank=conv_max_rank,
+            clamp_quantile=clamp_quantile, min_diff=min_diff, skip_patterns_str=skip_patterns,
+            mismatch_mode=mismatch_mode, chunk_large_layers=chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache, glob_skip_patterns=glob_skip_patterns
+        )
+        return io.NodeOutput(output_path)

pyproject.toml

@@ -1,9 +1,20 @@
 [project]
 name = "ComfyUI-ModelUtils"
 description = "[WIP]Custom nodes for handling, inspecting, modifying and creating various model files."
-version = "0.2.9"
+version = "0.5.2"
 license = { file = "LICENSE" }
 
+classifiers = [
+    "Development Status :: 5 - Production/Stable",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
 [project.urls]
 Repository = "https://github.com/silveroxides/ComfyUI-ModelUtils"
 #  Used by Comfy Registry https://comfyregistry.org

requirements.txt

@@ -1,5 +1,5 @@
 requests
 Pillow
 mutagen
-unifiedefficientloader
+unifiedefficientloader>=0.5.2
 av

tests/test_consensus_merger.py

@@ -0,0 +1,845 @@
+import importlib
+import sys
+import types
+from pathlib import Path
+
+import pytest
+import torch
+from safetensors.torch import load_file, save_file
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+PACKAGE_NAME = "modelutils_cwb_tests"
+
+
+def _load_module(name):
+    if PACKAGE_NAME not in sys.modules:
+        package = types.ModuleType(PACKAGE_NAME)
+        package.__path__ = [str(REPO_ROOT)]
+        sys.modules[PACKAGE_NAME] = package
+    return importlib.import_module(f"{PACKAGE_NAME}.{name}")
+
+
+@pytest.fixture(scope="module")
+def cwb():
+    return _load_module("nodes.consensus_merger")
+
+
+def _patch_io(monkeypatch, module, tmp_path, paths):
+    monkeypatch.setattr(module.folder_paths, "get_full_path", lambda _, name: paths.get(name))
+    monkeypatch.setattr(module.folder_paths, "get_folder_paths", lambda _: [str(tmp_path)])
+    monkeypatch.setattr(module.folder_paths, "models_dir", str(tmp_path))
+    monkeypatch.setattr(module, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(module, "cleanup_after_operation", lambda: None)
+
+
+def _result_path(tmp_path, category, result):
+    return tmp_path / category / result
+
+
+def _params(output_filename, **overrides):
+    params = {
+        "cwb_preset": "custom",
+        "consensus_type": "mean",
+        "alignment_method": "index",
+        "alignment_threshold": 0.4,
+        "similarity_threshold": 0.0,
+        "power_alpha": 2.0,
+        "diversity_beta": 0.0,
+        "rescale_norm": False,
+        "global_scale": 1.0,
+        "dynamic_similarity_contrast": False,
+        "soft_comfort_bandpass": False,
+        "position_weight": 0.0,
+        "preserve_common_prefix": False,
+        "mismatch_mode": "skip",
+        "output_filename": output_filename,
+        "save_dtype": "fp16",
+        "process_device": "cpu",
+        "exclude_patterns": "",
+        "discard_patterns": "",
+        "glob_patterns": False,
+        "lazy_load": True,
+        "force_clear_cache": False,
+        "override_dtype": False,
+        "include_1d_diffs": False,
+    }
+    params.update(overrides)
+    return params
+
+
+def test_preset_resolution_matches_extended_contract(cwb):
+    assert all(
+        preset["global_scale"] == pytest.approx(1.0)
+        for preset in cwb.CWB_PRESETS.values()
+    )
+
+    power = cwb.resolve_cwb_settings(cwb_preset="power_blend")
+    assert power.consensus_type == "median"
+    assert power.alignment_threshold == pytest.approx(0.9)
+    assert power.power_alpha == pytest.approx(8.0)
+    assert power.dynamic_similarity_contrast is True
+    assert power.soft_comfort_bandpass is False
+
+    baseline = cwb.resolve_cwb_settings(
+        cwb_preset="baseline",
+        alignment_method="index",
+        dynamic_similarity_contrast=True,
+        soft_comfort_bandpass=True,
+    )
+    assert baseline.alignment_method == "similarity"
+    assert baseline.dynamic_similarity_contrast is False
+    assert baseline.soft_comfort_bandpass is False
+
+    varied = cwb.resolve_cwb_settings(
+        cwb_preset="varied_merge",
+        global_scale=1.25,
+        position_weight=0.2,
+        preserve_common_prefix=True,
+    )
+    assert varied.global_scale == pytest.approx(1.25)
+    assert varied.position_weight == pytest.approx(0.2)
+    assert varied.preserve_common_prefix is True
+
+
+def test_cwb_math_index_prefix_and_scale(cwb):
+    settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        alignment_method="index",
+        global_scale=1.0,
+    )
+    first = torch.tensor([[1.0, 0.0], [0.0, 1.0]])
+    second = torch.tensor([[3.0, 0.0], [0.0, 3.0]])
+    torch.testing.assert_close(
+        cwb.merge_cwb_tensors([first, second], settings),
+        torch.tensor([[2.0, 0.0], [0.0, 2.0]]),
+    )
+
+    prefix_settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        alignment_method="index",
+        global_scale=2.0,
+        preserve_common_prefix=True,
+    )
+    prefixed = cwb.merge_cwb_tensors(
+        [torch.tensor([[9.0, 9.0], [1.0, 0.0]]),
+         torch.tensor([[9.0, 9.0], [3.0, 0.0]])],
+        prefix_settings,
+    )
+    torch.testing.assert_close(prefixed[0], torch.tensor([9.0, 9.0]))
+    torch.testing.assert_close(prefixed[1], torch.tensor([4.0, 0.0]))
+
+
+def test_similarity_alignment_reorders_matching_rows(cwb):
+    settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        alignment_method="similarity",
+        alignment_threshold=0.5,
+    )
+    reference = torch.tensor([[1.0, 0.0], [0.0, 1.0]])
+    reversed_source = torch.tensor([[0.0, 2.0], [2.0, 0.0]])
+    torch.testing.assert_close(
+        cwb.merge_cwb_tensors([reference, reversed_source], settings),
+        torch.tensor([[1.5, 0.0], [0.0, 1.5]]),
+    )
+
+    scores = torch.ones((2, 2))
+    biased = cwb._position_biased_scores(scores, 1.0)
+    assert biased[0, 0] > biased[0, 1]
+
+
+def test_fixed_coordinate_merge_never_reorders_model_rows(cwb):
+    settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        alignment_method="similarity",
+        alignment_threshold=0.5,
+    )
+    reference = torch.tensor([[1.0, 0.0], [0.0, 1.0]])
+    reversed_source = torch.tensor([[0.0, 2.0], [2.0, 0.0]])
+    expected = torch.stack([
+        cwb.merge_consensus_group(
+            torch.stack([reference[row], reversed_source[row]]), settings
+        )
+        for row in range(reference.shape[0])
+    ])
+    torch.testing.assert_close(
+        cwb.merge_cwb_tensors(
+            [reference, reversed_source],
+            settings,
+            allow_similarity_alignment=False,
+        ),
+        expected,
+    )
+
+
+def test_lora_similarity_alignment_pairs_a_rows_with_b_columns(cwb):
+    settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        alignment_method="similarity",
+        alignment_threshold=0.5,
+    )
+    reference_down = torch.tensor([[1.0, 0.0], [0.0, 1.0]])
+    reference_up = torch.tensor([[2.0, 0.0], [0.0, 3.0]])
+    source_down = torch.tensor([[0.0, 1.0], [-1.0, 0.0]])
+    source_up = torch.tensor([[0.0, -2.0], [3.0, 0.0]])
+
+    merged_down, merged_up = cwb.merge_cwb_lora_pairs(
+        [reference_down, source_down],
+        [reference_up, source_up],
+        settings,
+        reference_index=0,
+    )
+    torch.testing.assert_close(merged_down, reference_down)
+    torch.testing.assert_close(merged_up, reference_up)
+
+
+def test_lora_krea_shape_scores_only_rank_components(monkeypatch, cwb):
+    rank = 256
+    output_features = 36864
+    down = torch.ones((rank, 2))
+    up = torch.ones((output_features, rank))
+    matrix_shapes = []
+
+    def bounded_mm(left, right):
+        matrix_shapes.append((left.shape, right.shape))
+        return torch.eye(left.shape[0], right.shape[1], dtype=left.dtype)
+
+    monkeypatch.setattr(cwb.torch, "mm", bounded_mm)
+    monkeypatch.setattr(
+        cwb,
+        "merge_consensus_group",
+        lambda stacked, settings, **kwargs: stacked[0].clone(),
+    )
+    settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        alignment_method="similarity",
+        alignment_threshold=0.5,
+    )
+    merged_down, merged_up = cwb.merge_cwb_lora_pairs(
+        [down, down], [up, up], settings, reference_index=0
+    )
+
+    assert merged_down.shape == down.shape
+    assert merged_up.shape == up.shape
+    assert matrix_shapes
+    assert all(left[0] == rank and right[1] == rank for left, right in matrix_shapes)
+
+
+def test_lora_alpha_and_global_scale_apply_once_to_pair(cwb):
+    settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        alignment_method="index",
+        global_scale=3.0,
+    )
+    down = torch.ones((1, 1))
+    up = torch.ones((1, 1))
+    merged_down, merged_up, rank = cwb._merge_lora_pair_to_cpu(
+        [(0, down, up, 2.0), (1, down, up, 2.0)],
+        settings,
+        "cpu",
+        torch.float32,
+    )
+    assert rank == 1
+    torch.testing.assert_close(merged_down, torch.ones((1, 1)))
+    torch.testing.assert_close(merged_up, torch.full((1, 1), 6.0))
+
+
+def test_lora_pair_alignment_supports_convolution_factors(cwb):
+    settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        alignment_method="index",
+    )
+    down = torch.arange(6, dtype=torch.float32).reshape(2, 3, 1, 1)
+    up = torch.arange(8, dtype=torch.float32).reshape(4, 2, 1, 1)
+    merged_down, merged_up = cwb.merge_cwb_lora_pairs(
+        [down, down], [up, up], settings, reference_index=0
+    )
+    torch.testing.assert_close(merged_down, down)
+    torch.testing.assert_close(merged_up, up)
+
+
+def test_lora_cuda_oom_retries_current_pair_on_cpu(monkeypatch, caplog, cwb):
+    settings = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        alignment_method="index",
+    )
+    original_to_compute = cwb._to_compute
+    attempted_devices = []
+
+    def fail_cuda(tensor, device):
+        attempted_devices.append(device)
+        if str(device).startswith("cuda"):
+            raise torch.OutOfMemoryError("injected allocation failure")
+        return original_to_compute(tensor, device)
+
+    monkeypatch.setattr(cwb, "_to_compute", fail_cuda)
+    monkeypatch.setattr(cwb, "_release_failed_cuda_operation", lambda: None)
+    down = torch.ones((1, 2))
+    up = torch.ones((2, 1))
+    with caplog.at_level("WARNING"):
+        merged_down, merged_up, rank = cwb._merge_lora_pair_to_cpu(
+            [(0, down, up, 1.0), (1, down, up, 1.0)],
+            settings,
+            "cuda",
+            torch.float32,
+            operation_label="diffusion_model.foo",
+        )
+
+    assert rank == 1
+    torch.testing.assert_close(merged_down, down)
+    torch.testing.assert_close(merged_up, up)
+    assert attempted_devices[0] == "cuda"
+    assert "cpu" in attempted_devices
+    assert "retrying this layer on CPU" in caplog.text
+    assert "diffusion_model.foo" in caplog.text
+
+
+def test_norm_rescale_and_dsc_bandpass_are_finite(cwb):
+    rescale = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="mean",
+        rescale_norm=True,
+    )
+    merged = cwb.merge_consensus_group(
+        torch.tensor([[2.0, 0.0], [0.0, 2.0]]), rescale
+    )
+    assert torch.linalg.vector_norm(merged).item() == pytest.approx(2.0)
+
+    dsc = cwb.resolve_cwb_settings(
+        cwb_preset="custom",
+        consensus_type="median",
+        diversity_beta=1.5,
+        dynamic_similarity_contrast=True,
+        soft_comfort_bandpass=True,
+    )
+    result = cwb.merge_consensus_group(
+        torch.tensor([[1.0, 0.0], [0.8, 0.2], [0.0, 1.0]]), dsc
+    )
+    assert torch.isfinite(result).all()
+
+
+def test_all_ten_schemas_and_lora_switch_contract(cwb):
+    assert len(cwb.CWB_MERGER_NODES) == 10
+    assert len({node.NODE_ID for node in cwb.CWB_MERGER_NODES}) == 10
+    for node in cwb.CWB_MERGER_NODES:
+        schema = node.define_schema()
+        ids = [value.id for value in schema.inputs]
+        assert ids[:3] == ["execution_mode", "model_a", "model_b"]
+        if node.INPUT_COUNT == 3:
+            assert ids[3] == "model_c"
+        if node.LORA_MODE:
+            assert schema.inputs[-1].id == "include_1d_diffs"
+            assert schema.inputs[-1].default is False
+        else:
+            assert "include_1d_diffs" not in ids
+
+
+def test_streaming_dtype_nonfloat_and_secondary_preservation(
+    monkeypatch, tmp_path, cwb
+):
+    a = tmp_path / "a.safetensors"
+    b = tmp_path / "b.safetensors"
+    save_file({
+        "shared": torch.tensor([[1.0, 0.0], [0.0, 1.0]], dtype=torch.float32),
+        "only_a": torch.tensor([5.0, 6.0], dtype=torch.float16),
+        "metadata_tensor": torch.tensor([1, 2], dtype=torch.int64),
+    }, str(a), metadata={"source": "A"})
+    save_file({
+        "shared": torch.tensor([[3.0, 0.0], [0.0, 3.0]], dtype=torch.float16),
+        "extra": torch.ones((2, 2)),
+        "metadata_tensor": torch.tensor([9, 9], dtype=torch.int64),
+    }, str(b))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"], "diffusion_models", _params("anchored")
+    )
+    tensors = load_file(str(_result_path(tmp_path, "diffusion_models", result)))
+    assert set(tensors) == {"shared", "only_a", "extra", "metadata_tensor"}
+    torch.testing.assert_close(tensors["shared"], torch.tensor([[2.0, 0.0], [0.0, 2.0]]))
+    assert tensors["shared"].dtype == torch.float32
+    torch.testing.assert_close(tensors["only_a"], torch.tensor([5.0, 6.0], dtype=torch.float16))
+    torch.testing.assert_close(tensors["metadata_tensor"], torch.tensor([1, 2]))
+    torch.testing.assert_close(tensors["extra"], torch.ones((2, 2)))
+
+    overridden = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"],
+        "diffusion_models",
+        _params("anchored_override", override_dtype=True),
+    )
+    overridden_tensors = load_file(
+        str(_result_path(tmp_path, "diffusion_models", overridden))
+    )
+    assert overridden_tensors["shared"].dtype == torch.float16
+    assert overridden_tensors["metadata_tensor"].dtype == torch.int64
+
+
+def test_failed_merge_preserves_existing_output_and_removes_temporary_file(
+    monkeypatch, tmp_path, cwb
+):
+    a = tmp_path / "atomic_a.safetensors"
+    b = tmp_path / "atomic_b.safetensors"
+    output = tmp_path / "diffusion_models" / "atomic_output.safetensors"
+    output.parent.mkdir(parents=True)
+    save_file({"layer": torch.ones((2, 2))}, str(a))
+    save_file({"layer": torch.full((2, 2), 2.0)}, str(b))
+    save_file({"existing": torch.tensor([7.0])}, str(output))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+
+    def fail_merge(*args, **kwargs):
+        raise RuntimeError("injected CWB failure")
+
+    monkeypatch.setattr(cwb, "merge_cwb_tensors", fail_merge)
+    with pytest.raises(RuntimeError, match="injected CWB failure"):
+        cwb.ConsensusMergerLogic.execute(
+            ["a", "b"], "diffusion_models", _params("atomic_output")
+        )
+
+    existing = load_file(str(output))
+    assert set(existing) == {"existing"}
+    torch.testing.assert_close(existing["existing"], torch.tensor([7.0]))
+    assert not list(output.parent.glob(".atomic_output.safetensors.*.tmp"))
+
+
+def test_embedding_union_uses_longest_first_dimension(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "embed_a.safetensors"
+    b = tmp_path / "embed_b.safetensors"
+    save_file({"emb": torch.tensor([[1.0, 0.0], [0.0, 1.0]])}, str(a))
+    save_file({
+        "emb": torch.tensor([[3.0, 0.0], [0.0, 3.0], [4.0, 4.0]]),
+        "secondary_only": torch.tensor([[7.0, 0.0]]),
+    }, str(b))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"],
+        "embeddings",
+        _params("embedding_union"),
+        embedding_union=True,
+    )
+    tensors = load_file(str(_result_path(tmp_path, "embeddings", result)))
+    assert tensors["emb"].shape == (3, 2)
+    torch.testing.assert_close(tensors["emb"][:2], torch.tensor([[2.0, 0.0], [0.0, 2.0]]))
+    torch.testing.assert_close(tensors["emb"][2], torch.tensor([4.0, 4.0]))
+    torch.testing.assert_close(tensors["secondary_only"], torch.tensor([[7.0, 0.0]]))
+
+
+def test_three_input_streaming_merge(monkeypatch, tmp_path, cwb):
+    paths = {}
+    for name, value in (("a", 1.0), ("b", 3.0), ("c", 5.0)):
+        path = tmp_path / f"three_{name}.safetensors"
+        save_file({"layer": torch.tensor([[value, 0.0]])}, str(path))
+        paths[name] = str(path)
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b", "c"], "diffusion_models", _params("three_inputs")
+    )
+    tensor = load_file(str(_result_path(tmp_path, "diffusion_models", result)))["layer"]
+    torch.testing.assert_close(tensor, torch.tensor([[3.0, 0.0]]))
+
+
+def test_two_input_preserves_secondary_only_tensor_exactly(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "secondary_a.safetensors"
+    b = tmp_path / "secondary_b.safetensors"
+    save_file({"shared": torch.tensor([1.0])}, str(a))
+    save_file({
+        "shared": torch.tensor([1.0]),
+        "secondary": torch.tensor([2.0, 3.0], dtype=torch.float32),
+    }, str(b))
+    _patch_io(monkeypatch, cwb, tmp_path, {"a": str(a), "b": str(b)})
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"],
+        "diffusion_models",
+        _params(
+            "secondary_exact",
+            global_scale=2.0,
+            override_dtype=True,
+            save_dtype="fp16",
+        ),
+    )
+    tensor = load_file(str(_result_path(tmp_path, "diffusion_models", result)))["secondary"]
+    torch.testing.assert_close(tensor, torch.tensor([2.0, 3.0]))
+    assert tensor.dtype == torch.float32
+
+
+def test_three_input_merges_available_secondary_sources(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "union_a.safetensors"
+    b = tmp_path / "union_b.safetensors"
+    c = tmp_path / "union_c.safetensors"
+    save_file({"shared": torch.tensor([1.0])}, str(a))
+    save_file({
+        "shared": torch.tensor([1.0]),
+        "secondary_shared": torch.tensor([2.0]),
+        "b_only": torch.tensor([7.0]),
+    }, str(b))
+    save_file({
+        "shared": torch.tensor([1.0]),
+        "secondary_shared": torch.tensor([4.0]),
+        "c_only": torch.tensor([9.0]),
+    }, str(c))
+    _patch_io(monkeypatch, cwb, tmp_path, {"a": str(a), "b": str(b), "c": str(c)})
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b", "c"], "diffusion_models", _params("secondary_union")
+    )
+    tensors = load_file(str(_result_path(tmp_path, "diffusion_models", result)))
+    torch.testing.assert_close(tensors["secondary_shared"], torch.tensor([3.0]))
+    torch.testing.assert_close(tensors["b_only"], torch.tensor([7.0]))
+    torch.testing.assert_close(tensors["c_only"], torch.tensor([9.0]))
+
+
+def test_secondary_shape_conflict_skip_preserves_earliest_source(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "shape_a.safetensors"
+    b = tmp_path / "shape_b.safetensors"
+    c = tmp_path / "shape_c.safetensors"
+    save_file({"shared": torch.tensor([1.0])}, str(a))
+    save_file({"secondary": torch.tensor([2.0, 3.0])}, str(b))
+    save_file({"secondary": torch.tensor([[8.0, 9.0]])}, str(c))
+    _patch_io(monkeypatch, cwb, tmp_path, {"a": str(a), "b": str(b), "c": str(c)})
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b", "c"],
+        "diffusion_models",
+        _params("secondary_shape_skip", mismatch_mode="skip"),
+    )
+    tensor = load_file(str(_result_path(tmp_path, "diffusion_models", result)))["secondary"]
+    torch.testing.assert_close(tensor, torch.tensor([2.0, 3.0]))
+
+
+def test_secondary_only_filters_and_low_bit_preservation(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "filter_a.safetensors"
+    b = tmp_path / "filter_b.safetensors"
+    save_file({"shared": torch.tensor([1.0])}, str(a))
+    save_file({
+        "secondary.keep": torch.tensor([4, 5], dtype=torch.uint8),
+        "secondary.drop": torch.tensor([6.0]),
+    }, str(b))
+    _patch_io(monkeypatch, cwb, tmp_path, {"a": str(a), "b": str(b)})
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"],
+        "diffusion_models",
+        _params("secondary_filters", discard_patterns=r"secondary\.drop"),
+    )
+    tensors = load_file(str(_result_path(tmp_path, "diffusion_models", result)))
+    assert "secondary.drop" not in tensors
+    torch.testing.assert_close(
+        tensors["secondary.keep"], torch.tensor([4, 5], dtype=torch.uint8)
+    )
+
+
+def test_lora_cross_format_companion_group_preserves_model_a_canonically(
+    monkeypatch, tmp_path, cwb
+):
+    a = tmp_path / "lora_a.safetensors"
+    b = tmp_path / "lora_b.safetensors"
+    save_file({
+        "diffusion_model.foo.lora_A.weight": torch.tensor([[1.0, 2.0]], dtype=torch.float32),
+        "diffusion_model.foo.lora_B.weight": torch.tensor([[1.0], [2.0]], dtype=torch.float32),
+        "diffusion_model.foo.dora_scale": torch.tensor([1.0, 1.0]),
+    }, str(a))
+    save_file({
+        "lora_unet_foo.lora_down.weight": torch.tensor([[3.0, 4.0], [5.0, 6.0]]),
+        "lora_unet_foo.lora_up.weight": torch.tensor([[3.0, 5.0], [4.0, 6.0]]),
+        "lora_unet_secondary.lora_down.weight": torch.ones((1, 2)),
+        "lora_unet_secondary.lora_up.weight": torch.ones((2, 1)),
+    }, str(b))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"], "loras", _params("lora_cross_format"), lora_mode=True
+    )
+    tensors = load_file(str(_result_path(tmp_path, "loras", result)))
+    assert set(tensors) == {
+        "diffusion_model.foo.lora_A.weight",
+        "diffusion_model.foo.lora_B.weight",
+        "diffusion_model.foo.dora_scale",
+        "lora_unet_secondary.lora_A.weight",
+        "lora_unet_secondary.lora_B.weight",
+    }
+    assert tensors["diffusion_model.foo.lora_A.weight"].shape == (1, 2)
+    assert tensors["diffusion_model.foo.lora_B.weight"].shape == (2, 1)
+    assert tensors["diffusion_model.foo.lora_A.weight"].dtype == torch.float32
+    torch.testing.assert_close(
+        tensors["diffusion_model.foo.dora_scale"], torch.tensor([1.0, 1.0])
+    )
+    torch.testing.assert_close(
+        tensors["lora_unet_secondary.lora_A.weight"], torch.ones((1, 2))
+    )
+    torch.testing.assert_close(
+        tensors["lora_unet_secondary.lora_B.weight"], torch.ones((2, 1))
+    )
+
+
+def test_lora_secondary_only_pair_is_preserved_without_scaling(
+    monkeypatch, tmp_path, cwb
+):
+    a = tmp_path / "lora_secondary_a.safetensors"
+    b = tmp_path / "lora_secondary_b.safetensors"
+    save_file({
+        "diffusion_model.base.diff": torch.tensor([[1.0]]),
+    }, str(a))
+    save_file({
+        "transformer.extra.lora_down.weight": torch.tensor([[2.0, 3.0]]),
+        "transformer.extra.lora_up.weight": torch.tensor([[4.0], [5.0]]),
+        "transformer.extra.alpha": torch.tensor(1.0),
+    }, str(b))
+    _patch_io(monkeypatch, cwb, tmp_path, {"a": str(a), "b": str(b)})
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"],
+        "loras",
+        _params(
+            "lora_secondary_exact",
+            global_scale=2.0,
+            override_dtype=True,
+            save_dtype="fp16",
+        ),
+        lora_mode=True,
+    )
+    tensors = load_file(str(_result_path(tmp_path, "loras", result)))
+    torch.testing.assert_close(
+        tensors["diffusion_model.extra.lora_A.weight"],
+        torch.tensor([[2.0, 3.0]]),
+    )
+    torch.testing.assert_close(
+        tensors["diffusion_model.extra.lora_B.weight"],
+        torch.tensor([[4.0], [5.0]]),
+    )
+    assert tensors["diffusion_model.extra.lora_A.weight"].dtype == torch.float32
+    assert tensors["diffusion_model.extra.alpha"].item() == pytest.approx(1.0)
+
+
+def test_lora_three_input_merges_b_and_c_group_without_a(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "lora_union_a.safetensors"
+    b = tmp_path / "lora_union_b.safetensors"
+    c = tmp_path / "lora_union_c.safetensors"
+    save_file({"diffusion_model.base.diff": torch.tensor([[1.0]])}, str(a))
+    save_file({
+        "diffusion_model.extra.lora_A.weight": torch.full((1, 2), 1.0),
+        "diffusion_model.extra.lora_B.weight": torch.full((2, 1), 1.0),
+    }, str(b))
+    save_file({
+        "transformer.extra.lora_down.weight": torch.full((1, 2), 3.0),
+        "transformer.extra.lora_up.weight": torch.full((2, 1), 3.0),
+    }, str(c))
+    _patch_io(monkeypatch, cwb, tmp_path, {"a": str(a), "b": str(b), "c": str(c)})
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b", "c"], "loras", _params("lora_secondary_union"), lora_mode=True
+    )
+    tensors = load_file(str(_result_path(tmp_path, "loras", result)))
+    torch.testing.assert_close(
+        tensors["diffusion_model.extra.lora_A.weight"], torch.full((1, 2), 2.0)
+    )
+    torch.testing.assert_close(
+        tensors["diffusion_model.extra.lora_B.weight"], torch.full((2, 1), 2.0)
+    )
+
+
+def test_lora_secondary_direct_and_companion_groups_are_atomic(
+    monkeypatch, tmp_path, cwb
+):
+    a = tmp_path / "lora_atomic_a.safetensors"
+    b = tmp_path / "lora_atomic_b.safetensors"
+    save_file({"diffusion_model.base.diff": torch.tensor([[1.0]])}, str(a))
+    save_file({
+        "diffusion_model.norm.diff": torch.tensor([2.0, 3.0], dtype=torch.bfloat16),
+        "diffusion_model.extra.lora_A.weight": torch.tensor([[4.0, 5.0]]),
+        "diffusion_model.extra.lora_B.weight": torch.tensor([[6.0], [7.0]]),
+        "diffusion_model.extra.dora_scale": torch.tensor([8.0, 9.0]),
+    }, str(b))
+    _patch_io(monkeypatch, cwb, tmp_path, {"a": str(a), "b": str(b)})
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"],
+        "loras",
+        _params("lora_secondary_atomic", global_scale=3.0),
+        lora_mode=True,
+    )
+    tensors = load_file(str(_result_path(tmp_path, "loras", result)))
+    torch.testing.assert_close(
+        tensors["diffusion_model.norm.diff"],
+        torch.tensor([2.0, 3.0], dtype=torch.bfloat16),
+    )
+    assert tensors["diffusion_model.norm.diff"].dtype == torch.bfloat16
+    torch.testing.assert_close(
+        tensors["diffusion_model.extra.lora_A.weight"], torch.tensor([[4.0, 5.0]])
+    )
+    torch.testing.assert_close(
+        tensors["diffusion_model.extra.lora_B.weight"], torch.tensor([[6.0], [7.0]])
+    )
+    torch.testing.assert_close(
+        tensors["diffusion_model.extra.dora_scale"], torch.tensor([8.0, 9.0])
+    )
+
+
+def test_lora_rejects_duplicate_normalized_cores_in_secondary_input(cwb):
+    with pytest.raises(ValueError, match="same logical layer 'foo'.*input 2"):
+        cwb._secondary_lora_map(
+            {
+                "diffusion_model.foo": {},
+                "transformer.foo": {},
+            },
+            input_index=1,
+        )
+
+
+def test_lora_peft_prefix_matching_and_existing_alpha(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "peft_a.safetensors"
+    b = tmp_path / "peft_b.safetensors"
+    save_file({
+        "base_model.model.diffusion_model.foo.lora_A.weight": torch.ones((1, 2)),
+        "base_model.model.diffusion_model.foo.lora_B.weight": torch.ones((2, 1)),
+        "base_model.model.diffusion_model.foo.alpha": torch.tensor(1.0),
+    }, str(a))
+    save_file({
+        "diffusion_model.foo.lora_down.weight": torch.ones((2, 2)),
+        "diffusion_model.foo.lora_up.weight": torch.ones((2, 2)),
+        "diffusion_model.foo.alpha": torch.tensor(2.0),
+    }, str(b))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"], "loras", _params("peft_prefix"), lora_mode=True
+    )
+    tensors = load_file(str(_result_path(tmp_path, "loras", result)))
+    assert set(tensors) == {
+        "diffusion_model.foo.lora_A.weight",
+        "diffusion_model.foo.lora_B.weight",
+        "diffusion_model.foo.alpha",
+    }
+    assert tensors["diffusion_model.foo.alpha"].item() == pytest.approx(2.0)
+
+
+def test_lora_mochi_inputs_emit_preferred_canonical_output(
+    monkeypatch, tmp_path, cwb
+):
+    a = tmp_path / "mochi_a.safetensors"
+    b = tmp_path / "mochi_b.safetensors"
+    for path, value in ((a, 1.0), (b, 3.0)):
+        save_file({
+            "diffusion_model.foo.lora_A": torch.full((1, 2), value),
+            "diffusion_model.foo.lora_B": torch.full((2, 1), value),
+            "diffusion_model.foo.alpha": torch.tensor(1.0),
+        }, str(path))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"], "loras", _params("mochi_canonical"), lora_mode=True
+    )
+    tensors = load_file(str(_result_path(tmp_path, "loras", result)))
+
+    assert set(tensors) == {
+        "diffusion_model.foo.lora_A.weight",
+        "diffusion_model.foo.lora_B.weight",
+        "diffusion_model.foo.alpha",
+    }
+
+
+def test_lora_1d_direct_switch_is_default_fallback_and_fp32(
+    monkeypatch, tmp_path, cwb
+):
+    a = tmp_path / "direct_a.safetensors"
+    b = tmp_path / "direct_b.safetensors"
+    save_file({"diffusion_model.norm.diff": torch.tensor([1.0, 2.0], dtype=torch.bfloat16)}, str(a))
+    save_file({"diffusion_model.norm.diff": torch.tensor([3.0, 4.0], dtype=torch.bfloat16)}, str(b))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+
+    default = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"], "loras", _params("direct_default"), lora_mode=True
+    )
+    default_tensor = load_file(str(_result_path(tmp_path, "loras", default)))[
+        "diffusion_model.norm.diff"
+    ]
+    torch.testing.assert_close(default_tensor, torch.tensor([1.0, 2.0], dtype=torch.bfloat16))
+    assert default_tensor.dtype == torch.bfloat16
+
+    enabled = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"],
+        "loras",
+        _params("direct_enabled", include_1d_diffs=True, override_dtype=True),
+        lora_mode=True,
+    )
+    enabled_tensor = load_file(str(_result_path(tmp_path, "loras", enabled)))[
+        "diffusion_model.norm.diff"
+    ]
+    torch.testing.assert_close(enabled_tensor, torch.tensor([2.0, 3.0]))
+    assert enabled_tensor.dtype == torch.float32
+
+
+def test_quantized_marker_errors_before_output(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "quant_a.safetensors"
+    b = tmp_path / "quant_b.safetensors"
+    save_file({
+        "layer.weight": torch.ones((2, 2)),
+        "layer.comfy_quant": torch.tensor([1], dtype=torch.uint8),
+    }, str(a))
+    save_file({"layer.weight": torch.ones((2, 2))}, str(b))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+    with pytest.raises(ValueError, match="ComfyUI quantization metadata"):
+        cwb.ConsensusMergerLogic.execute(
+            ["a", "b"], "diffusion_models", _params("must_not_exist")
+        )
+    assert not (tmp_path / "diffusion_models" / "must_not_exist.safetensors").exists()
+
+
+def test_isolated_low_bit_preserves_anchored_tensor(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "guard_a.safetensors"
+    b = tmp_path / "guard_b.safetensors"
+    save_file({"layer": torch.tensor([1.0, 2.0])}, str(a))
+    save_file({"layer": torch.tensor([8, 9], dtype=torch.uint8)}, str(b))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"], "diffusion_models", _params("guarded_generic")
+    )
+    tensor = load_file(str(_result_path(tmp_path, "diffusion_models", result)))["layer"]
+    torch.testing.assert_close(tensor, torch.tensor([1.0, 2.0]))
+    assert tensor.dtype == torch.float32
+
+
+def test_isolated_low_bit_preserves_complete_lora_layer(monkeypatch, tmp_path, cwb):
+    a = tmp_path / "guard_lora_a.safetensors"
+    b = tmp_path / "guard_lora_b.safetensors"
+    save_file({
+        "diffusion_model.foo.lora_A.weight": torch.tensor([[1.0, 2.0]]),
+        "diffusion_model.foo.lora_B.weight": torch.tensor([[3.0], [4.0]]),
+        "diffusion_model.foo.alpha": torch.tensor(1.0),
+        "diffusion_model.foo.dora_scale": torch.tensor([5.0, 6.0]),
+    }, str(a))
+    save_file({
+        "diffusion_model.foo.lora_A.weight": torch.tensor([[9.0, 9.0]]),
+        "diffusion_model.foo.lora_B.weight": torch.tensor([[9.0], [9.0]]),
+        "diffusion_model.foo.alpha": torch.tensor(1, dtype=torch.uint8),
+    }, str(b))
+    paths = {"a": str(a), "b": str(b)}
+    _patch_io(monkeypatch, cwb, tmp_path, paths)
+    result = cwb.ConsensusMergerLogic.execute(
+        ["a", "b"], "loras", _params("guarded_lora"), lora_mode=True
+    )
+    tensors = load_file(str(_result_path(tmp_path, "loras", result)))
+    torch.testing.assert_close(
+        tensors["diffusion_model.foo.lora_A.weight"], torch.tensor([[1.0, 2.0]])
+    )
+    torch.testing.assert_close(
+        tensors["diffusion_model.foo.lora_B.weight"], torch.tensor([[3.0], [4.0]])
+    )
+    assert tensors["diffusion_model.foo.alpha"].item() == pytest.approx(1.0)
+    torch.testing.assert_close(
+        tensors["diffusion_model.foo.dora_scale"], torch.tensor([5.0, 6.0])
+    )

tests/test_lora_diff_dtype_contract.py

@@ -0,0 +1,899 @@
+import importlib
+import logging
+import sys
+import types
+from pathlib import Path
+
+import pytest
+import torch
+from safetensors.torch import load_file, save_file
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+PACKAGE_NAME = "modelutils_contract_tests"
+
+
+def _load_module(name):
+    if PACKAGE_NAME not in sys.modules:
+        package = types.ModuleType(PACKAGE_NAME)
+        package.__path__ = [str(REPO_ROOT)]
+        sys.modules[PACKAGE_NAME] = package
+    return importlib.import_module(f"{PACKAGE_NAME}.{name}")
+
+
+@pytest.fixture(scope="module")
+def modules():
+    return _load_module("nodes.lora_resize"), _load_module("nodes.lora_merger")
+
+
+@pytest.fixture(scope="module")
+def generic_modules():
+    return _load_module("nodes.merger"), _load_module("nodes.merger_ops")
+
+
+def _patch_output(monkeypatch, module, output_dir):
+    monkeypatch.setattr(module.folder_paths, "get_folder_paths", lambda _: [str(output_dir)])
+    monkeypatch.setattr(module.folder_paths, "models_dir", str(output_dir))
+    monkeypatch.setattr(module, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(module, "cleanup_after_operation", lambda: None)
+
+
+COMFY_LORA_PAIR_SUFFIXES = [
+    (".lora_down.weight", ".lora_up.weight"),
+    ("_lora.down.weight", "_lora.up.weight"),
+    (".lora_A.weight", ".lora_B.weight"),
+    (".lora.down.weight", ".lora.up.weight"),
+    (".lora_A", ".lora_B"),
+    (".lora_linear_layer.down.weight", ".lora_linear_layer.up.weight"),
+    (".lora_A.default.weight", ".lora_B.default.weight"),
+]
+
+
+@pytest.mark.parametrize("down_suffix,up_suffix", COMFY_LORA_PAIR_SUFFIXES)
+def test_all_comfy_lora_pair_spellings_parse_to_canonical_keys(
+    modules, down_suffix, up_suffix
+):
+    resize, _ = modules
+    block = "diffusion_model.layer"
+    layers, passthrough = resize.parse_lora_layers([
+        f"{block}{down_suffix}",
+        f"{block}{up_suffix}",
+        f"{block}.alpha",
+    ])
+
+    assert not passthrough
+    assert layers[block]["down"] == f"{block}{down_suffix}"
+    assert layers[block]["up"] == f"{block}{up_suffix}"
+    assert resize.canonical_lora_key(block, "down") == f"{block}.lora_A.weight"
+    assert resize.canonical_lora_key(block, "up") == f"{block}.lora_B.weight"
+
+
+@pytest.mark.parametrize("down_suffix,up_suffix", COMFY_LORA_PAIR_SUFFIXES)
+def test_merge_to_model_applies_every_comfy_lora_pair_spelling(
+    monkeypatch, tmp_path, modules, down_suffix, up_suffix
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    variant = COMFY_LORA_PAIR_SUFFIXES.index((down_suffix, up_suffix))
+    base = tmp_path / f"base_{variant}.safetensors"
+    adapter = tmp_path / f"adapter_{variant}.safetensors"
+    block = "diffusion_model.layer"
+    save_file({
+        "model.diffusion_model.layer.weight": torch.zeros((2, 2), dtype=torch.bfloat16),
+    }, str(base))
+    save_file({
+        f"{block}{down_suffix}": torch.tensor([[1.0, 2.0]]),
+        f"{block}{up_suffix}": torch.tensor([[3.0], [4.0]]),
+        f"{block}.alpha": torch.tensor(0.5),
+    }, str(adapter))
+
+    output = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float16,
+        f"merged_variant_{variant}", verbose=False,
+    )
+    result = load_file(output)["model.diffusion_model.layer.weight"]
+    torch.testing.assert_close(
+        result,
+        torch.tensor([[1.5, 3.0], [2.0, 4.0]]),
+    )
+    assert result.dtype == torch.float32
+
+
+def test_merge_to_model_applies_comfy_direct_norm_and_set_forms(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "direct_base.safetensors"
+    adapter = tmp_path / "direct_adapter.safetensors"
+    save_file({
+        "model.diffusion_model.norm.weight": torch.zeros(2),
+        "model.diffusion_model.norm.bias": torch.zeros(2),
+        "model.diffusion_model.replace.weight": torch.zeros((2, 2)),
+    }, str(base))
+    save_file({
+        "diffusion_model.norm.w_norm": torch.tensor([1.0, 2.0]),
+        "diffusion_model.norm.b_norm": torch.tensor([3.0, 4.0]),
+        "diffusion_model.replace.set_weight": torch.full((2, 2), 7.0),
+    }, str(adapter))
+
+    output = resize.merge_loras_to_model(
+        [str(adapter)], [0.5], str(base), "cpu", torch.float16,
+        "direct_forms", verbose=False, include_1d_diffs=True,
+    )
+    result = load_file(output)
+    torch.testing.assert_close(
+        result["model.diffusion_model.norm.weight"], torch.tensor([0.5, 1.0])
+    )
+    torch.testing.assert_close(
+        result["model.diffusion_model.norm.bias"], torch.tensor([1.5, 2.0])
+    )
+    torch.testing.assert_close(
+        result["model.diffusion_model.replace.weight"], torch.full((2, 2), 7.0)
+    )
+
+
+def test_merge_to_model_uses_comfy_reshape_companion(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "reshape_base.safetensors"
+    adapter = tmp_path / "reshape_adapter.safetensors"
+    save_file({
+        "model.diffusion_model.layer.weight": torch.zeros((1, 2)),
+    }, str(base))
+    save_file({
+        "diffusion_model.layer.lora_A.weight": torch.tensor([[1.0, 2.0]]),
+        "diffusion_model.layer.lora_B.weight": torch.tensor([[3.0], [4.0]]),
+        "diffusion_model.layer.reshape_weight": torch.tensor([2, 2]),
+    }, str(adapter))
+
+    output = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float16,
+        "reshape_companion", verbose=False,
+    )
+    result = load_file(output)["model.diffusion_model.layer.weight"]
+    assert result.shape == (2, 2)
+    torch.testing.assert_close(result, torch.tensor([[3.0, 6.0], [4.0, 8.0]]))
+
+
+def test_merge_to_model_reports_success_and_separates_base_only_biases(
+    monkeypatch, tmp_path, modules, capsys
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "summary_base.safetensors"
+    adapter = tmp_path / "summary_adapter.safetensors"
+    save_file({
+        "model.diffusion_model.layer.weight": torch.zeros((2, 2)),
+        "model.diffusion_model.layer.bias": torch.zeros(2),
+    }, str(base))
+    save_file({
+        "diffusion_model.layer.lora_A.weight": torch.ones((1, 2)),
+        "diffusion_model.layer.lora_B.weight": torch.ones((2, 1)),
+    }, str(adapter))
+
+    path, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "summary", verbose=True, return_report=True,
+    )
+    output = capsys.readouterr().out
+
+    assert Path(path).is_file()
+    assert report.startswith("SUCCESS: 1/1 adapter groups fully applied")
+    assert "[PATCHED]" not in report
+    assert "[APPLIED]" not in report
+    assert "L1: summary_adapter.safetensors, strength=1.0" in report
+    assert "[BASE-ONLY BIASES RETAINED] (1)" in report
+    assert "model.diffusion_model.layer.bias" in report
+    assert "SUCCESS: 1/1 groups applied" in output
+    assert "1 tensors patched, 1 base tensors not patched, 0 adapter issues" in output
+    assert "model.diffusion_model.layer.bias" not in output
+
+
+def test_success_report_stays_compact_and_console_is_summary_only(
+    monkeypatch, tmp_path, modules, capsys
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "many_success_base.safetensors"
+    adapter = tmp_path / "many_success_adapter.safetensors"
+    save_file({
+        f"model.diffusion_model.layer_{index}.weight": torch.zeros((2, 2))
+        for index in range(40)
+    }, str(base))
+    save_file({
+        f"diffusion_model.layer_{index}.diff": torch.ones((2, 2))
+        for index in range(40)
+    }, str(adapter))
+
+    _, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "many_success", verbose=True, return_report=True,
+    )
+    console = capsys.readouterr().out
+
+    assert report.startswith("SUCCESS: 40/40")
+    assert "layer_0.weight" not in report
+    assert "ADAPTER GROUP ISSUES\n- None" in report
+    assert len(report) < 1000
+    assert max(map(len, report.splitlines())) < 160
+    assert "layer_0.weight" not in console
+    assert len(console.splitlines()) <= 9
+
+
+def test_merge_to_model_report_identifies_shape_mismatch_and_unresolved_group(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "diagnostic_base.safetensors"
+    adapter = tmp_path / "diagnostic_adapter.safetensors"
+    save_file({
+        "model.diffusion_model.layer.weight": torch.zeros((2, 2)),
+    }, str(base))
+    save_file({
+        "diffusion_model.layer.diff": torch.ones(3),
+        "diffusion_model.orphan.lora_A.weight": torch.ones((1, 2)),
+        "diffusion_model.orphan.lora_B.weight": torch.ones((2, 1)),
+    }, str(adapter))
+
+    _, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "diagnostic", verbose=False, include_1d_diffs=True, return_report=True,
+    )
+
+    assert report.startswith("COMPLETED WITH ERRORS: 0/2")
+    assert "[MAPPED BUT UNCHANGED] (1)" in report
+    assert "[SHAPE MISMATCH] (1)" in report
+    assert "source_shape=(3,), target_shape=(2, 2)" in report
+    assert "[UNRESOLVED] (1)" in report
+    assert "diffusion_model.orphan" in report
+    assert "no matching base tensor" in report
+
+
+def test_merge_to_model_detects_comfy_adapter_logged_failure_atomically(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "logged_failure_base.safetensors"
+    adapter = tmp_path / "logged_failure_adapter.safetensors"
+    base_weight = torch.arange(4, dtype=torch.float32).reshape(2, 2)
+    save_file({"model.diffusion_model.layer.weight": base_weight}, str(base))
+    save_file({
+        "diffusion_model.layer.lora_A.weight": torch.ones((1, 2)),
+        "diffusion_model.layer.lora_B.weight": torch.ones((2, 1)),
+    }, str(adapter))
+
+    class LoggedFailureAdapter:
+        def calculate_weight(self, weight, key, *args, **kwargs):
+            weight.add_(99.0)
+            logging.error("ERROR lora %s synthetic internal failure", key)
+            return weight
+
+    class FakeLoRAAdapter:
+        @staticmethod
+        def load(*args, **kwargs):
+            return LoggedFailureAdapter()
+
+    monkeypatch.setattr(resize, "LoRAAdapter", FakeLoRAAdapter)
+    path, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "logged_failure", verbose=False, return_report=True,
+    )
+
+    torch.testing.assert_close(
+        load_file(path)["model.diffusion_model.layer.weight"], base_weight
+    )
+    assert report.startswith("COMPLETED WITH ERRORS: 0/1")
+    assert "[CALCULATION FAILURE] (1)" in report
+    assert "synthetic internal failure" in report
+    assert "[MAPPED BUT UNCHANGED] (1)" in report
+
+
+def test_merge_to_model_report_classifies_disabled_1d_as_exclusion(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "disabled_1d_base.safetensors"
+    adapter = tmp_path / "disabled_1d_adapter.safetensors"
+    save_file({"model.diffusion_model.norm.weight": torch.zeros(2)}, str(base))
+    save_file({"diffusion_model.norm.diff": torch.ones(2)}, str(adapter))
+
+    path, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "disabled_1d", verbose=False, return_report=True,
+    )
+
+    torch.testing.assert_close(
+        load_file(path)["model.diffusion_model.norm.weight"], torch.zeros(2)
+    )
+    assert report.startswith("COMPLETED WITH EXCLUSIONS: 0/1")
+    assert "[1D DISABLED] (1)" in report
+    assert "source=diffusion_model.norm.diff" in report
+    assert "include_1d_diffs is disabled" in report
+
+
+def test_merge_to_model_report_marks_partially_applied_group(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "partial_base.safetensors"
+    adapter = tmp_path / "partial_adapter.safetensors"
+    save_file({
+        "model.diffusion_model.layer.weight": torch.zeros((2, 2)),
+        "model.diffusion_model.layer.bias": torch.zeros(2),
+    }, str(base))
+    save_file({
+        "diffusion_model.layer.lora_A.weight": torch.ones((1, 2)),
+        "diffusion_model.layer.lora_B.weight": torch.ones((2, 1)),
+        "diffusion_model.layer.diff_b": torch.ones(3),
+    }, str(adapter))
+
+    path, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "partial", verbose=False, include_1d_diffs=True, return_report=True,
+    )
+    result = load_file(path)
+
+    torch.testing.assert_close(
+        result["model.diffusion_model.layer.weight"], torch.ones((2, 2))
+    )
+    torch.testing.assert_close(
+        result["model.diffusion_model.layer.bias"], torch.zeros(2)
+    )
+    assert report.startswith("COMPLETED WITH ERRORS: 0/1")
+    assert "[PARTIALLY APPLIED] (1)" in report
+    assert "source_shape=(3,), target_shape=(2,)" in report
+
+
+def test_merge_to_model_detects_raised_adapter_failure_atomically(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "raised_failure_base.safetensors"
+    adapter = tmp_path / "raised_failure_adapter.safetensors"
+    base_weight = torch.eye(2)
+    save_file({"model.diffusion_model.layer.weight": base_weight}, str(base))
+    save_file({
+        "diffusion_model.layer.lora_A.weight": torch.ones((1, 2)),
+        "diffusion_model.layer.lora_B.weight": torch.ones((2, 1)),
+    }, str(adapter))
+
+    class RaisedFailureAdapter:
+        def calculate_weight(self, weight, key, *args, **kwargs):
+            weight.zero_()
+            raise RuntimeError("synthetic raised failure")
+
+    class FakeLoRAAdapter:
+        @staticmethod
+        def load(*args, **kwargs):
+            return RaisedFailureAdapter()
+
+    monkeypatch.setattr(resize, "LoRAAdapter", FakeLoRAAdapter)
+    path, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "raised_failure", verbose=False, return_report=True,
+    )
+
+    torch.testing.assert_close(
+        load_file(path)["model.diffusion_model.layer.weight"], base_weight
+    )
+    assert "[CALCULATION FAILURE] (1)" in report
+    assert "RuntimeError: synthetic raised failure" in report
+
+
+def test_merge_to_model_report_lists_skipped_tensor_and_pattern(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "skip_base.safetensors"
+    adapter = tmp_path / "skip_adapter.safetensors"
+    save_file({
+        "model.diffusion_model.layer.weight": torch.zeros((2, 2)),
+        "model.diffusion_model.layer.bias": torch.zeros(2),
+    }, str(base))
+    save_file({
+        "diffusion_model.layer.lora_A.weight": torch.ones((1, 2)),
+        "diffusion_model.layer.lora_B.weight": torch.ones((2, 1)),
+    }, str(adapter))
+
+    path, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "skip", skip_patterns_str=r"layer\.weight$", verbose=False,
+        return_report=True,
+    )
+
+    assert "model.diffusion_model.layer.weight" not in load_file(path)
+    assert report.startswith("COMPLETED WITH EXCLUSIONS: 0/1")
+    assert "[OMITTED BY SKIP PATTERN] (1)" in report
+    assert r"pattern=layer\.weight$" in report
+    assert "[SKIPPED BY PATTERN] (1)" in report
+
+
+def test_merge_to_model_node_returns_path_blank_line_report(
+    monkeypatch, modules
+):
+    resize, _ = modules
+    monkeypatch.setattr(
+        resize.folder_paths, "get_full_path_or_raise",
+        lambda category, name: f"{category}/{name}",
+    )
+    monkeypatch.setattr(
+        resize, "merge_loras_to_model",
+        lambda **kwargs: ("saved/model.safetensors", "SUCCESS\nDETAILS"),
+    )
+
+    result = resize.LoRAMergeToModel.execute(
+        "base.safetensors", "1",
+        "one.safetensors", 1.0,
+        "None", 1.0, "None", 1.0, "None", 1.0,
+        "None", 1.0, "None", 1.0, "None", 1.0, "None", 1.0,
+        "", "merged", "fp32", "cpu", True, False, False,
+    )
+
+    assert result.result == ("saved/model.safetensors\n\nSUCCESS\nDETAILS",)
+
+
+def test_mixed_parser_scans_every_key_and_retains_unpaired(modules):
+    resize, _ = modules
+    filler = [f"metadata.tensor_{index}" for index in range(60)]
+    keys = filler + [
+        "diffusion_model.linear.lora_down.weight",
+        "diffusion_model.linear.lora_up.weight",
+        "diffusion_model.linear.alpha",
+        "diffusion_model.norm.diff",
+        "diffusion_model.block.diff_b",
+        "diffusion_model.orphan.lora_down.weight",
+        "diffusion_model.linear.dora_scale",
+    ]
+
+    info = resize.detect_lora_format(keys)
+    layers, passthrough = resize.parse_lora_layers(keys)
+
+    assert info["format"] == "mixed"
+    assert layers["diffusion_model.linear"]["down"].endswith("lora_down.weight")
+    assert layers["diffusion_model.norm"]["diff"].endswith(".diff")
+    assert layers["diffusion_model.block"]["diff_b"].endswith(".diff_b")
+    assert "diffusion_model.orphan.lora_down.weight" in passthrough
+    assert layers["diffusion_model.linear"]["dora_scale"].endswith(".dora_scale")
+
+
+def test_dtype_contract(modules):
+    resize, _ = modules
+    assert resize.select_output_dtype([torch.bfloat16], torch.float16, is_1d_diff=True) == torch.float32
+    assert resize.select_output_dtype([torch.float32, torch.bfloat16], torch.float16) == torch.float32
+    assert resize.select_output_dtype([torch.float32], torch.bfloat16, force=True) == torch.bfloat16
+
+
+def test_include_1d_switches_are_appended_and_default_off(modules, generic_modules):
+    resize, multi = modules
+    generic, _ = generic_modules
+    node_classes = [
+        generic.LoRATwoMerger,
+        generic.LoRAThreeMerger,
+        multi.LoRAMultiMerge,
+        multi.LoRAMultiMergeDARE,
+        multi.LoRAMultiMergeDAREEnhanced,
+        resize.LoRAMergeToModel,
+    ]
+
+    for node_class in node_classes:
+        final_input = node_class.define_schema().inputs[-1]
+        assert final_input.id == "include_1d_diffs"
+        assert final_input.default is False
+
+    assert len(resize.LoRAMergeToModel.define_schema().outputs) == 1
+
+
+def test_resize_preserves_mixed_direct_auxiliary_and_fp32(monkeypatch, tmp_path, modules):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    source = tmp_path / "mixed.safetensors"
+    save_file({
+        "diffusion_model.linear.lora_down.weight": torch.arange(8, dtype=torch.float32).reshape(2, 4),
+        "diffusion_model.linear.lora_up.weight": torch.arange(8, dtype=torch.float32).reshape(4, 2),
+        "diffusion_model.linear.alpha": torch.tensor(2.0, dtype=torch.float32),
+        "diffusion_model.norm.diff": torch.tensor([0.125, -0.25], dtype=torch.bfloat16),
+        "diffusion_model.scale.diff": torch.tensor(0.5, dtype=torch.bfloat16),
+        "diffusion_model.lin.diff": torch.full((2, 2), 1e-8, dtype=torch.float32),
+        "diffusion_model.linear.dora_scale": torch.ones(4, dtype=torch.float32),
+        "diffusion_model.orphan.lora_down.weight": torch.ones((1, 2), dtype=torch.bfloat16),
+    }, str(source))
+
+    output = resize.resize_lora_file(
+        str(source), 1, None, None, "cpu", torch.float16, "resized", verbose=False
+    )
+    tensors = load_file(output)
+
+    assert tensors["diffusion_model.linear.lora_A.weight"].shape[0] == 2
+    assert tensors["diffusion_model.linear.lora_A.weight"].dtype == torch.float32
+    assert tensors["diffusion_model.linear.lora_B.weight"].dtype == torch.float32
+    assert tensors["diffusion_model.linear.alpha"].dtype == torch.float32
+    assert tensors["diffusion_model.norm.diff"].dtype == torch.float32
+    assert tensors["diffusion_model.scale.diff"].dtype == torch.float16
+    assert tensors["diffusion_model.lin.diff"].dtype == torch.float32
+    assert tensors["diffusion_model.linear.dora_scale"].dtype == torch.float32
+    assert tensors["diffusion_model.orphan.lora_down.weight"].dtype == torch.float16
+
+    pure_source = tmp_path / "pure_diff.safetensors"
+    save_file({
+        "diffusion_model.pure.diff": torch.ones((2, 2), dtype=torch.bfloat16),
+        "diffusion_model.pure_norm.diff": torch.ones(2, dtype=torch.bfloat16),
+    }, str(pure_source))
+    pure_output = resize.resize_lora_file(
+        str(pure_source), 1, None, None, "cpu", torch.float16, "pure_resized", verbose=False
+    )
+    pure_tensors = load_file(pure_output)
+    assert set(pure_tensors) == {"diffusion_model.pure.diff", "diffusion_model.pure_norm.diff"}
+    assert pure_tensors["diffusion_model.pure.diff"].dtype == torch.float16
+    assert pure_tensors["diffusion_model.pure_norm.diff"].dtype == torch.float32
+
+
+def test_resize_normalizes_mochi_pair_to_preferred_output(
+    monkeypatch, tmp_path, modules
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    source = tmp_path / "mochi.safetensors"
+    save_file({
+        "diffusion_model.layer.lora_A": torch.ones((2, 2)),
+        "diffusion_model.layer.lora_B": torch.ones((2, 2)),
+        "diffusion_model.layer.alpha": torch.tensor(2.0),
+    }, str(source))
+
+    output = resize.resize_lora_file(
+        str(source), 1, None, None, "cpu", torch.float32,
+        "mochi_normalized", verbose=False,
+    )
+    tensors = load_file(output)
+    assert set(tensors) == {
+        "diffusion_model.layer.lora_A.weight",
+        "diffusion_model.layer.lora_B.weight",
+        "diffusion_model.layer.alpha",
+    }
+
+
+def test_standard_multi_merge_normalizes_mochi_pair_to_preferred_output(
+    monkeypatch, tmp_path, modules
+):
+    _, merger = modules
+    _patch_output(monkeypatch, merger, tmp_path)
+    first = tmp_path / "multi_mochi_a.safetensors"
+    second = tmp_path / "multi_mochi_b.safetensors"
+    for path, value in ((first, 1.0), (second, 3.0)):
+        save_file({
+            "diffusion_model.layer.lora_A": torch.full((1, 2), value),
+            "diffusion_model.layer.lora_B": torch.full((2, 1), value),
+            "diffusion_model.layer.alpha": torch.tensor(1.0),
+        }, str(path))
+
+    output = merger.merge_multi_loras(
+        [str(first), str(second)], [1.0, 1.0], "weighted_sum", "cpu",
+        torch.float32, "multi_mochi_normalized", verbose=False,
+    )
+    tensors = load_file(output)
+
+    assert set(tensors) == {
+        "diffusion_model.layer.lora_A.weight",
+        "diffusion_model.layer.lora_B.weight",
+        "diffusion_model.layer.alpha",
+    }
+
+
+def test_standard_multi_merge_emits_weighted_direct_union(monkeypatch, tmp_path, modules):
+    _, merger = modules
+    _patch_output(monkeypatch, merger, tmp_path)
+    first = tmp_path / "first.safetensors"
+    second = tmp_path / "second.safetensors"
+    base = tmp_path / "multi_base.safetensors"
+    save_file({
+        "model.diffusion_model.norm.weight": torch.zeros(2),
+        "model.diffusion_model.matrix.weight": torch.zeros((2, 2)),
+        "model.diffusion_model.only_first.weight": torch.zeros(()),
+        "model.diffusion_model.scale": torch.zeros(()),
+        "model.diffusion_model.lowrank.weight": torch.zeros((2, 2)),
+    }, str(base))
+    save_file({
+        "diffusion_model.norm.diff": torch.tensor([1.0, 2.0], dtype=torch.bfloat16),
+        "diffusion_model.matrix.diff": torch.ones((2, 2), dtype=torch.float32),
+        "diffusion_model.only_first.diff": torch.tensor(4.0, dtype=torch.bfloat16),
+        "diffusion_model.scale.diff": torch.tensor(1.0, dtype=torch.bfloat16),
+        "diffusion_model.lowrank.lora_down.weight": torch.ones((1, 2), dtype=torch.float32),
+        "diffusion_model.lowrank.lora_up.weight": torch.ones((2, 1), dtype=torch.float32),
+        "diffusion_model.lowrank.alpha": torch.tensor(1.0, dtype=torch.float32),
+    }, str(first))
+    save_file({
+        "diffusion_model.norm.diff": torch.tensor([3.0, 4.0], dtype=torch.bfloat16),
+        "diffusion_model.matrix.diff": torch.full((2, 2), 2.0, dtype=torch.bfloat16),
+        "diffusion_model.lowrank.lora_down.weight": torch.full((1, 2), 2.0, dtype=torch.bfloat16),
+        "diffusion_model.lowrank.lora_up.weight": torch.full((2, 1), 2.0, dtype=torch.bfloat16),
+        "diffusion_model.lowrank.alpha": torch.tensor(1.0, dtype=torch.bfloat16),
+    }, str(second))
+
+    output = merger.merge_multi_loras(
+        [str(first), str(second)], [2.0, -1.0], "concatenate", "cpu",
+        torch.float16, "merged", base_model_path=str(base), verbose=False,
+        include_1d_diffs=True,
+    )
+    tensors = load_file(output)
+
+    torch.testing.assert_close(tensors["diffusion_model.norm.diff"], torch.tensor([-1.0, 0.0]))
+    torch.testing.assert_close(tensors["diffusion_model.matrix.diff"], torch.zeros((2, 2)))
+    assert tensors["diffusion_model.norm.diff"].dtype == torch.float32
+    assert tensors["diffusion_model.matrix.diff"].dtype == torch.float32
+    assert tensors["diffusion_model.only_first.diff"].item() == pytest.approx(8.0)
+    assert tensors["diffusion_model.only_first.diff"].dtype == torch.float16
+    assert tensors["diffusion_model.scale.diff"].item() == pytest.approx(2.0)
+    assert tensors["diffusion_model.lowrank.lora_A.weight"].shape[0] == 2
+    assert tensors["diffusion_model.lowrank.lora_A.weight"].dtype == torch.float32
+    assert tensors["diffusion_model.lowrank.lora_B.weight"].dtype == torch.float32
+    assert tensors["diffusion_model.lowrank.alpha"].dtype == torch.float32
+
+    default_output = merger.merge_multi_loras(
+        [str(first), str(second)], [2.0, -1.0], "concatenate", "cpu",
+        torch.float16, "merged_default", base_model_path=str(base), verbose=False,
+    )
+    default_tensors = load_file(default_output)
+    assert "diffusion_model.norm.diff" not in default_tensors
+    assert "diffusion_model.matrix.diff" in default_tensors
+    assert "diffusion_model.scale.diff" in default_tensors
+    assert "diffusion_model.lowrank.lora_A.weight" in default_tensors
+
+
+def test_multi_merge_rejects_direct_shape_mismatch(monkeypatch, tmp_path, modules):
+    _, merger = modules
+    _patch_output(monkeypatch, merger, tmp_path)
+    first = tmp_path / "shape_first.safetensors"
+    second = tmp_path / "shape_second.safetensors"
+    save_file({"diffusion_model.bad.diff": torch.ones(2)}, str(first))
+    save_file({"diffusion_model.bad.diff": torch.ones(3)}, str(second))
+
+    with pytest.raises(ValueError, match="Direct LoRA shape mismatch"):
+        merger.merge_multi_loras(
+            [str(first), str(second)], [1.0, 1.0], "weighted_sum", "cpu",
+            torch.float16, "bad_shapes", verbose=False, include_1d_diffs=True,
+        )
+
+
+@pytest.mark.parametrize("enhanced", [False, True])
+def test_dare_variants_emit_direct_layers(monkeypatch, tmp_path, modules, enhanced):
+    _, merger = modules
+    _patch_output(monkeypatch, merger, tmp_path)
+    first = tmp_path / f"dare_first_{enhanced}.safetensors"
+    second = tmp_path / f"dare_second_{enhanced}.safetensors"
+    save_file({
+        "diffusion_model.norm.diff": torch.tensor([1.0, 2.0], dtype=torch.bfloat16),
+        "diffusion_model.lowrank.lora_down.weight": torch.ones((1, 2), dtype=torch.float32),
+        "diffusion_model.lowrank.lora_up.weight": torch.ones((2, 1), dtype=torch.float32),
+    }, str(first))
+    save_file({
+        "diffusion_model.norm.diff": torch.tensor([3.0, 4.0], dtype=torch.float32),
+        "diffusion_model.lowrank.lora_down.weight": torch.ones((1, 2), dtype=torch.bfloat16),
+        "diffusion_model.lowrank.lora_up.weight": torch.ones((2, 1), dtype=torch.bfloat16),
+    }, str(second))
+
+    if enhanced:
+        output = merger.merge_multi_loras_dare_enhanced(
+            [str(first), str(second)], [1.0, 1.0], 1.0, 1.0, 0.0, 0.0, 7,
+            "cpu", torch.float16, "enhanced", verbose=False, include_1d_diffs=True,
+        )
+    else:
+        output = merger.merge_multi_loras_dare(
+            [str(first), str(second)], [1.0, 1.0], 0.0, 0.0, 7,
+            "cpu", torch.float16, "dare", verbose=False, include_1d_diffs=True,
+        )
+    tensors = load_file(output)
+
+    torch.testing.assert_close(tensors["diffusion_model.norm.diff"], torch.tensor([2.0, 3.0]))
+    assert tensors["diffusion_model.norm.diff"].dtype == torch.float32
+    assert tensors["diffusion_model.lowrank.lora_A.weight"].dtype == torch.float32
+    assert tensors["diffusion_model.lowrank.lora_B.weight"].dtype == torch.float32
+
+    if enhanced:
+        default_output = merger.merge_multi_loras_dare_enhanced(
+            [str(first), str(second)], [1.0, 1.0], 1.0, 1.0, 0.0, 0.0, 7,
+            "cpu", torch.float16, "enhanced_default", verbose=False,
+        )
+    else:
+        default_output = merger.merge_multi_loras_dare(
+            [str(first), str(second)], [1.0, 1.0], 0.0, 0.0, 7,
+            "cpu", torch.float16, "dare_default", verbose=False,
+        )
+    default_tensors = load_file(default_output)
+    assert "diffusion_model.norm.diff" not in default_tensors
+    assert "diffusion_model.lowrank.lora_A.weight" in default_tensors
+
+
+def test_merge_to_model_applies_generic_direct_keys_and_preserves_dtype(
+    monkeypatch, tmp_path, modules, caplog
+):
+    resize, _ = modules
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "base.safetensors"
+    adapter = tmp_path / "adapter.safetensors"
+    save_file({
+        "model.diffusion_model.block.weight": torch.zeros((2, 2), dtype=torch.bfloat16),
+        "model.diffusion_model.norm.weight": torch.zeros(2, dtype=torch.bfloat16),
+        "model.diffusion_model.block.bias": torch.zeros(2, dtype=torch.bfloat16),
+        "model.diffusion_model.scale": torch.tensor(1.0, dtype=torch.bfloat16),
+        "model.diffusion_model.lin": torch.zeros((2, 2), dtype=torch.bfloat16),
+        "model.diffusion_model.bad.weight": torch.ones((2, 2), dtype=torch.bfloat16),
+        "model.diffusion_model.untouched": torch.tensor(1e-8, dtype=torch.float32),
+    }, str(base))
+    save_file({
+        "diffusion_model.block.diff": torch.full((2, 2), 1e-8, dtype=torch.float32),
+        "diffusion_model.norm.diff": torch.tensor([1.0, 2.0], dtype=torch.bfloat16),
+        "diffusion_model.block.diff_b": torch.tensor([3.0, 4.0], dtype=torch.bfloat16),
+        "diffusion_model.scale.diff": torch.tensor(0.5, dtype=torch.bfloat16),
+        "diffusion_model.lin.diff": torch.ones((2, 2), dtype=torch.float32),
+        "diffusion_model.bad.diff": torch.ones(3, dtype=torch.float32),
+    }, str(adapter))
+
+    output = resize.merge_loras_to_model(
+        [str(adapter)], [0.5], str(base), "cpu", torch.float16, "merged_model",
+        verbose=False, include_1d_diffs=True,
+    )
+    tensors = load_file(output)
+
+    torch.testing.assert_close(tensors["model.diffusion_model.norm.weight"], torch.tensor([0.5, 1.0]))
+    torch.testing.assert_close(tensors["model.diffusion_model.block.bias"], torch.tensor([1.5, 2.0]))
+    assert tensors["model.diffusion_model.scale"].item() == pytest.approx(1.25)
+    torch.testing.assert_close(tensors["model.diffusion_model.lin"], torch.full((2, 2), 0.5))
+    torch.testing.assert_close(tensors["model.diffusion_model.bad.weight"].float(), torch.ones((2, 2)))
+    assert "Shape mismatch" in caplog.text
+    assert tensors["model.diffusion_model.block.weight"].dtype == torch.float32
+    assert tensors["model.diffusion_model.norm.weight"].dtype == torch.float32
+    assert tensors["model.diffusion_model.block.bias"].dtype == torch.float32
+    assert tensors["model.diffusion_model.scale"].dtype == torch.float16
+    assert tensors["model.diffusion_model.lin"].dtype == torch.float32
+    assert tensors["model.diffusion_model.untouched"].dtype == torch.float32
+
+    default_output = resize.merge_loras_to_model(
+        [str(adapter)], [0.5], str(base), "cpu", torch.float16, "merged_model_default",
+        verbose=False,
+    )
+    default_tensors = load_file(default_output)
+    torch.testing.assert_close(
+        default_tensors["model.diffusion_model.norm.weight"].float(), torch.zeros(2)
+    )
+    torch.testing.assert_close(
+        default_tensors["model.diffusion_model.block.bias"].float(), torch.zeros(2)
+    )
+    assert default_tensors["model.diffusion_model.scale"].item() == pytest.approx(1.25)
+    torch.testing.assert_close(
+        default_tensors["model.diffusion_model.lin"], torch.full((2, 2), 0.5)
+    )
+
+
+def _generic_params(output_filename, override_dtype, include_1d_diffs=False):
+    return {
+        "mismatch_mode": "skip",
+        "alignment_mode": "pad/crop",
+        "alpha": 0.5,
+        "beta": 0.0,
+        "gamma": 0.5,
+        "delta": 2.0,
+        "epsilon": 0.01,
+        "zeta": 0.0,
+        "seed": 0,
+        "output_filename": output_filename,
+        "save_dtype": "bf16",
+        "override_dtype": override_dtype,
+        "device": "cpu",
+        "dtype": torch.float32,
+        "exclude_patterns": "",
+        "discard_patterns": "",
+        "glob_patterns": False,
+        "lazy_load": True,
+        "force_clear_cache": False,
+        "include_1d_diffs": include_1d_diffs,
+    }
+
+
+def test_generic_two_and_three_mergers_enforce_direct_dtype_contract(
+    monkeypatch, tmp_path, generic_modules
+):
+    generic, operations = generic_modules
+    paths = {}
+    for name, tensors in {
+        "a": {
+            "diffusion_model.norm.diff": torch.tensor([1.0, 2.0], dtype=torch.bfloat16),
+            "diffusion_model.matrix.diff": torch.ones((2, 2), dtype=torch.float32),
+        },
+        "b": {
+            "diffusion_model.norm.diff": torch.tensor([3.0, 4.0], dtype=torch.bfloat16),
+            "diffusion_model.matrix.diff": torch.full((2, 2), 3.0, dtype=torch.bfloat16),
+            "diffusion_model.secondary_only.diff": torch.ones(1, dtype=torch.float32),
+        },
+        "c": {
+            "diffusion_model.norm.diff": torch.tensor([0.5, 1.0], dtype=torch.float32),
+            "diffusion_model.matrix.diff": torch.full((2, 2), 0.5, dtype=torch.bfloat16),
+        },
+    }.items():
+        path = tmp_path / f"{name}.safetensors"
+        save_file(tensors, str(path))
+        paths[name] = str(path)
+
+    monkeypatch.setattr(generic.folder_paths, "get_full_path", lambda _, name: paths.get(name))
+    monkeypatch.setattr(generic.folder_paths, "get_folder_paths", lambda _: [str(tmp_path)])
+    monkeypatch.setattr(generic.folder_paths, "models_dir", str(tmp_path))
+    monkeypatch.setattr(generic, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(generic, "cleanup_after_operation", lambda: None)
+
+    two_params = _generic_params("generic_two", override_dtype=True)
+    two_params.update({"model_a": "a", "model_b": "b"})
+    generic.MergerLogic.execute_merge(
+        {"model_a": "a", "model_b": "b"},
+        "Weight-Sum",
+        operations.TWO_MODEL_MODES,
+        two_params,
+        "loras",
+    )
+    two = load_file(str(tmp_path / "loras" / "generic_two.safetensors"))
+    torch.testing.assert_close(
+        two["diffusion_model.norm.diff"].float(), torch.tensor([1.0, 2.0])
+    )
+    assert two["diffusion_model.norm.diff"].dtype == torch.bfloat16
+    assert two["diffusion_model.matrix.diff"].dtype == torch.bfloat16
+    assert "diffusion_model.secondary_only.diff" not in two
+
+    two_enabled_params = _generic_params(
+        "generic_two_enabled", override_dtype=True, include_1d_diffs=True
+    )
+    two_enabled_params.update({"model_a": "a", "model_b": "b"})
+    generic.MergerLogic.execute_merge(
+        {"model_a": "a", "model_b": "b"},
+        "Weight-Sum",
+        operations.TWO_MODEL_MODES,
+        two_enabled_params,
+        "loras",
+    )
+    two_enabled = load_file(
+        str(tmp_path / "loras" / "generic_two_enabled.safetensors")
+    )
+    torch.testing.assert_close(
+        two_enabled["diffusion_model.norm.diff"], torch.tensor([2.0, 3.0])
+    )
+    assert two_enabled["diffusion_model.norm.diff"].dtype == torch.float32
+
+    three_params = _generic_params("generic_three", override_dtype=False)
+    three_params.update({"model_a": "a", "model_b": "b", "model_c": "c"})
+    generic.MergerLogic.execute_merge(
+        {"model_a": "a", "model_b": "b", "model_c": "c"},
+        "Add-Difference",
+        operations.THREE_MODEL_MODES,
+        three_params,
+        "loras",
+    )
+    three = load_file(str(tmp_path / "loras" / "generic_three.safetensors"))
+    torch.testing.assert_close(
+        three["diffusion_model.norm.diff"].float(), torch.tensor([1.0, 2.0])
+    )
+    assert three["diffusion_model.norm.diff"].dtype == torch.bfloat16
+    assert three["diffusion_model.matrix.diff"].dtype == torch.float32
+
+    three_enabled_params = _generic_params(
+        "generic_three_enabled", override_dtype=False, include_1d_diffs=True
+    )
+    three_enabled_params.update({"model_a": "a", "model_b": "b", "model_c": "c"})
+    generic.MergerLogic.execute_merge(
+        {"model_a": "a", "model_b": "b", "model_c": "c"},
+        "Add-Difference",
+        operations.THREE_MODEL_MODES,
+        three_enabled_params,
+        "loras",
+    )
+    three_enabled = load_file(
+        str(tmp_path / "loras" / "generic_three_enabled.safetensors")
+    )
+    torch.testing.assert_close(
+        three_enabled["diffusion_model.norm.diff"], torch.tensor([2.25, 3.5])
+    )
+    assert three_enabled["diffusion_model.norm.diff"].dtype == torch.float32

tests/test_lora_quantization_guard.py

@@ -0,0 +1,437 @@
+import importlib
+import json
+import logging
+import struct
+import sys
+import types
+from pathlib import Path
+
+import pytest
+import torch
+from safetensors.torch import load_file, save_file
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+PACKAGE_NAME = "modelutils_quant_guard_tests"
+
+
+def _load_module(name):
+    if PACKAGE_NAME not in sys.modules:
+        package = types.ModuleType(PACKAGE_NAME)
+        package.__path__ = [str(REPO_ROOT)]
+        sys.modules[PACKAGE_NAME] = package
+    return importlib.import_module(f"{PACKAGE_NAME}.{name}")
+
+
+@pytest.fixture(scope="module")
+def modules():
+    return {
+        "guard": _load_module("nodes.quantization_guard"),
+        "extract": _load_module("nodes.lora_extract_svd"),
+        "resize": _load_module("nodes.lora_resize"),
+        "multi": _load_module("nodes.lora_merger"),
+        "generic": _load_module("nodes.merger"),
+        "operations": _load_module("nodes.merger_ops"),
+    }
+
+
+def _patch_output(monkeypatch, module, output_dir):
+    monkeypatch.setattr(module.folder_paths, "get_folder_paths", lambda _: [str(output_dir)])
+    monkeypatch.setattr(module.folder_paths, "models_dir", str(output_dir))
+    monkeypatch.setattr(module, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(module, "cleanup_after_operation", lambda: None)
+
+
+def _inspect_file(path, modules, label="input"):
+    handler = modules["resize"].MemoryEfficientSafeOpen(str(path), low_memory=True)
+    try:
+        return modules["guard"].inspect_low_bit_input(handler, label, "Guard Test")
+    finally:
+        handler.__exit__(None, None, None)
+
+
+def _round_trip_preserved(path, output, modules):
+    handler = modules["resize"].MemoryEfficientSafeOpen(str(path), low_memory=True)
+    try:
+        with modules["resize"].IncrementalSafetensorsWriter(str(output)) as writer:
+            modules["guard"].write_preserved_tensor(writer, "low", handler)
+    finally:
+        handler.__exit__(None, None, None)
+    return load_file(str(output))["low"]
+
+
+def _tensor_payload(path, key="low"):
+    with open(path, "rb") as source:
+        header_size = struct.unpack("<Q", source.read(8))[0]
+        header = json.loads(source.read(header_size).decode("utf-8"))
+        start, end = header[key]["data_offsets"]
+        source.seek(8 + header_size + start)
+        return source.read(end - start)
+
+
+@pytest.mark.parametrize("dtype", [torch.int8, torch.uint8])
+def test_int8_uint8_are_detected_and_warn_once(tmp_path, modules, caplog, dtype):
+    path = tmp_path / f"{dtype}.safetensors"
+    save_file({"low": torch.arange(4, dtype=dtype), "normal": torch.ones(1)}, str(path))
+
+    with caplog.at_level(logging.WARNING):
+        detected = _inspect_file(path, modules, "threshold input")
+
+    assert detected == {"low"}
+    records = [record for record in caplog.records if "threshold input" in record.message]
+    assert len(records) == 1
+    assert "1 isolated low-bit tensor" in records[0].message
+    assert "low" in records[0].message
+
+
+def test_exact_two_tensor_threshold(tmp_path, modules):
+    for count in range(4):
+        path = tmp_path / f"threshold_{count}.safetensors"
+        tensors = {f"low_{i}": torch.ones(2, dtype=torch.uint8) for i in range(count)}
+        tensors["normal"] = torch.ones(1)
+        save_file(tensors, str(path))
+        if count <= 2:
+            assert len(_inspect_file(path, modules)) == count
+        else:
+            with pytest.raises(ValueError, match="3 low-bit tensors.*casted or quantized"):
+                _inspect_file(path, modules)
+
+
+@pytest.mark.parametrize(
+    ("tensors", "metadata"),
+    [
+        ({"layer.comfy_quant": torch.tensor([123], dtype=torch.uint8)}, None),
+        ({"normal": torch.ones(1)}, {"_quantization_metadata": "{}"}),
+        ({"scaled_fp8": torch.ones(2)}, None),
+    ],
+)
+def test_recognized_comfy_quantization_hard_errors(tmp_path, modules, tensors, metadata):
+    path = tmp_path / f"quant_{len(list(tmp_path.iterdir()))}.safetensors"
+    save_file(tensors, str(path), metadata=metadata)
+    with pytest.raises(ValueError, match="ComfyUI quantization metadata.*unsupported"):
+        _inspect_file(path, modules, "recognized input")
+
+
+def test_supported_fp8_storage_round_trip(tmp_path, modules):
+    candidates = [
+        dtype
+        for name in (
+            "float8_e4m3fn", "float8_e4m3fnuz", "float8_e5m2", "float8_e5m2fnuz",
+            "float8_e8m0fnu",
+        )
+        if (dtype := getattr(torch, name, None)) is not None
+    ]
+    supported = []
+    for index, dtype in enumerate(candidates):
+        path = tmp_path / f"low_float_{index}.safetensors"
+        try:
+            save_file({"low": torch.zeros(4, dtype=dtype)}, str(path))
+            loaded = load_file(str(path))["low"]
+        except (KeyError, RuntimeError, TypeError, ValueError):
+            continue
+        supported.append(dtype)
+        assert loaded.dtype == dtype
+        assert _inspect_file(path, modules) == {"low"}
+        preserved = _round_trip_preserved(path, tmp_path / f"preserved_{index}.safetensors", modules)
+        assert preserved.dtype == dtype
+        torch.testing.assert_close(preserved.float(), loaded.float())
+
+    if candidates and not supported:
+        pytest.skip("Installed safetensors cannot write any installed torch FP8 dtype")
+
+
+def test_fp4_storage_when_supported(tmp_path, modules):
+    dtype = getattr(torch, "float4_e2m1fn_x2", None)
+    if dtype is None:
+        pytest.skip("Installed PyTorch has no torch.float4_e2m1fn_x2")
+    path = tmp_path / "fp4.safetensors"
+    try:
+        save_file({"low": torch.zeros(4, dtype=dtype)}, str(path))
+        loaded = load_file(str(path))["low"]
+    except (KeyError, RuntimeError, TypeError, ValueError):
+        pytest.skip("Installed safetensors cannot write torch.float4_e2m1fn_x2")
+    assert loaded.dtype == dtype
+    assert _inspect_file(path, modules) == {"low"}
+    preserved = _round_trip_preserved(path, tmp_path / "preserved_fp4.safetensors", modules)
+    assert preserved.dtype == dtype
+    assert _tensor_payload(path) == _tensor_payload(tmp_path / "preserved_fp4.safetensors")
+
+
+def test_realistic_packed_uint8_comfy_sidecar_is_rejected(tmp_path, modules):
+    path = tmp_path / "packed_comfy.safetensors"
+    save_file({
+        "layer.weight": torch.arange(8, dtype=torch.uint8).reshape(2, 4),
+        "layer.comfy_quant": torch.tensor(list(b'{"format":"int8"}'), dtype=torch.uint8),
+        "layer.weight_scale": torch.ones(2),
+    }, str(path))
+    with pytest.raises(ValueError, match="ComfyUI quantization metadata"):
+        _inspect_file(path, modules)
+
+
+def test_resize_preserves_complete_low_bit_layer_and_resizes_normal(
+    monkeypatch, tmp_path, modules
+):
+    resize = modules["resize"]
+    _patch_output(monkeypatch, resize, tmp_path)
+    source = tmp_path / "resize_quant.safetensors"
+    guarded_down = torch.tensor([[1, 2], [3, 4]], dtype=torch.uint8)
+    save_file({
+        "diffusion_model.guarded.lora_down.weight": guarded_down,
+        "diffusion_model.guarded.lora_up.weight": torch.ones((2, 2)),
+        "diffusion_model.guarded.alpha": torch.tensor(2.0),
+        "diffusion_model.guarded.dora_scale": torch.tensor([0.125, 0.25]),
+        "diffusion_model.normal.lora_down.weight": torch.ones((2, 2)),
+        "diffusion_model.normal.lora_up.weight": torch.ones((2, 2)),
+        "isolated_aux": torch.tensor([7], dtype=torch.int8),
+    }, str(source))
+
+    output = resize.resize_lora_file(
+        str(source), 1, None, None, "cpu", torch.float16, "resize_guarded", verbose=False
+    )
+    result = load_file(output)
+    torch.testing.assert_close(result["diffusion_model.guarded.lora_A.weight"], guarded_down)
+    assert result["diffusion_model.guarded.lora_A.weight"].dtype == torch.uint8
+    assert result["diffusion_model.guarded.lora_B.weight"].dtype == torch.float32
+    assert result["diffusion_model.guarded.dora_scale"].dtype == torch.float32
+    assert result["diffusion_model.normal.lora_A.weight"].shape[0] == 1
+    assert result["isolated_aux"].dtype == torch.int8
+
+
+def test_extract_excludes_low_bit_candidate_and_extracts_float_layer(
+    monkeypatch, tmp_path, modules
+):
+    extract = modules["extract"]
+    monkeypatch.setattr(extract, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(extract, "cleanup_after_operation", lambda: None)
+    model_a = tmp_path / "extract_a.safetensors"
+    model_b = tmp_path / "extract_b.safetensors"
+    output = tmp_path / "extracted.safetensors"
+    save_file({
+        "model.diffusion_model.normal.weight": torch.eye(2),
+        "model.diffusion_model.quant.weight": torch.ones((2, 2), dtype=torch.uint8),
+    }, str(model_a))
+    save_file({
+        "model.diffusion_model.normal.weight": torch.zeros((2, 2)),
+        "model.diffusion_model.quant.weight": torch.zeros((2, 2), dtype=torch.uint8),
+    }, str(model_b))
+
+    extract.extract_lora_from_files(
+        str(model_a), str(model_b), "fixed", 1, 1, "cpu", "fp32", str(output),
+        force_clear_cache=False,
+    )
+    result = load_file(str(output))
+    assert any("normal" in key for key in result)
+    assert not any("quant" in key for key in result)
+
+
+def test_extract_bias_diff_round_trips_through_merge_to_model(
+    monkeypatch, tmp_path, modules
+):
+    extract = modules["extract"]
+    resize = modules["resize"]
+    monkeypatch.setattr(extract, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(extract, "cleanup_after_operation", lambda: None)
+    _patch_output(monkeypatch, resize, tmp_path)
+
+    finetuned = tmp_path / "bias_finetuned.safetensors"
+    base = tmp_path / "bias_base.safetensors"
+    adapter = tmp_path / "bias_adapter.safetensors"
+    save_file({
+        "model.diffusion_model.layer.bias": torch.tensor([0.25, -0.5]),
+    }, str(finetuned))
+    save_file({
+        "model.diffusion_model.layer.bias": torch.tensor([0.0, 0.5]),
+    }, str(base))
+
+    extract.extract_lora_from_files(
+        str(finetuned), str(base), "fixed", 1, 1, "cpu", "fp16", str(adapter),
+        force_clear_cache=False, include_1d_diffs=True,
+    )
+    extracted = load_file(str(adapter))
+    assert set(extracted) == {"diffusion_model.layer.diff_b"}
+    assert extracted["diffusion_model.layer.diff_b"].dtype == torch.float32
+
+    merged_path = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float32,
+        "bias_round_trip", verbose=False, include_1d_diffs=True,
+    )
+    merged = load_file(merged_path)
+    torch.testing.assert_close(
+        merged["model.diffusion_model.layer.bias"],
+        load_file(str(finetuned))["model.diffusion_model.layer.bias"],
+    )
+
+
+def _run_multi_variant(module, variant, paths, output_name):
+    common = (paths, [1.0, 1.0])
+    if variant == "standard":
+        return module.merge_multi_loras(
+            *common, "weighted_sum", "cpu", torch.float16, output_name,
+            verbose=False, include_1d_diffs=True,
+        )
+    if variant == "dare":
+        return module.merge_multi_loras_dare(
+            *common, 0.0, 0.0, 11, "cpu", torch.float16, output_name,
+            verbose=False, include_1d_diffs=True,
+        )
+    return module.merge_multi_loras_dare_enhanced(
+        *common, 1.0, 1.0, 0.0, 0.0, 11, "cpu", torch.float16, output_name,
+        verbose=False, include_1d_diffs=True,
+    )
+
+
+@pytest.mark.parametrize("variant", ["standard", "dare", "enhanced"])
+def test_multi_variants_preserve_earliest_complete_affected_layer(
+    monkeypatch, tmp_path, modules, variant
+):
+    multi = modules["multi"]
+    _patch_output(monkeypatch, multi, tmp_path)
+    first = tmp_path / f"{variant}_first.safetensors"
+    second = tmp_path / f"{variant}_second.safetensors"
+    first_down = torch.tensor([[1.0, 2.0]])
+    save_file({
+        "diffusion_model.guarded.lora_down.weight": first_down,
+        "diffusion_model.guarded.lora_up.weight": torch.tensor([[3.0], [4.0]]),
+        "diffusion_model.guarded.dora_scale": torch.tensor([0.125, 0.25]),
+        "diffusion_model.normal.diff": torch.ones((2, 2)),
+    }, str(first))
+    save_file({
+        "diffusion_model.guarded.lora_down.weight": torch.tensor([[1, 2]], dtype=torch.uint8),
+        "diffusion_model.guarded.lora_up.weight": torch.tensor([[9.0], [9.0]]),
+        "diffusion_model.guarded.dora_scale": torch.tensor([9.0, 9.0]),
+        "diffusion_model.normal.diff": torch.full((2, 2), 3.0),
+    }, str(second))
+
+    output = _run_multi_variant(multi, variant, [str(first), str(second)], f"guard_{variant}")
+    result = load_file(output)
+    torch.testing.assert_close(result["diffusion_model.guarded.lora_A.weight"], first_down)
+    torch.testing.assert_close(
+        result["diffusion_model.guarded.lora_B.weight"], torch.tensor([[3.0], [4.0]])
+    )
+    torch.testing.assert_close(
+        result["diffusion_model.guarded.dora_scale"], torch.tensor([0.125, 0.25])
+    )
+    assert "diffusion_model.normal.diff" in result
+
+
+def test_hard_error_happens_before_resize_output_creation(monkeypatch, tmp_path, modules):
+    resize = modules["resize"]
+    _patch_output(monkeypatch, resize, tmp_path)
+    source = tmp_path / "three_low.safetensors"
+    save_file({f"low_{i}": torch.ones(1, dtype=torch.uint8) for i in range(3)}, str(source))
+    output = tmp_path / "must_not_exist.safetensors"
+
+    with pytest.raises(ValueError, match="3 low-bit tensors"):
+        resize.resize_lora_file(
+            str(source), 1, None, None, "cpu", torch.float16, output.stem, verbose=False
+        )
+    assert not output.exists()
+
+
+def test_recognized_quantization_creates_no_resize_output(monkeypatch, tmp_path, modules):
+    resize = modules["resize"]
+    _patch_output(monkeypatch, resize, tmp_path)
+    source = tmp_path / "recognized_quant.safetensors"
+    save_file({
+        "layer.weight": torch.ones((2, 2)),
+        "layer.comfy_quant": torch.tensor(list(b'{"format":"int8"}'), dtype=torch.uint8),
+    }, str(source))
+    output = tmp_path / "recognized_must_not_exist.safetensors"
+
+    with pytest.raises(ValueError, match="ComfyUI quantization metadata"):
+        resize.resize_lora_file(
+            str(source), 1, None, None, "cpu", torch.float16, output.stem, verbose=False
+        )
+    assert not output.exists()
+
+
+def _generic_params(output_filename):
+    return {
+        "mismatch_mode": "skip", "alignment_mode": "pad/crop", "alpha": 0.5,
+        "beta": 0.0, "gamma": 0.5, "delta": 2.0, "epsilon": 0.01, "zeta": 0.0,
+        "seed": 0, "output_filename": output_filename, "save_dtype": "bf16",
+        "override_dtype": True, "device": "cpu", "dtype": torch.float32,
+        "exclude_patterns": "", "discard_patterns": "", "glob_patterns": False,
+        "lazy_load": True, "force_clear_cache": False, "include_1d_diffs": True,
+    }
+
+
+def test_generic_merge_preserves_model_a_when_either_source_key_is_low_bit(
+    monkeypatch, tmp_path, modules
+):
+    generic = modules["generic"]
+    operations = modules["operations"]
+    model_a = tmp_path / "generic_guard_a.safetensors"
+    model_b = tmp_path / "generic_guard_b.safetensors"
+    a_tensors = {
+        "a_low": torch.tensor([[1, 2]], dtype=torch.uint8),
+        "b_low": torch.tensor([[0.125, 0.25]], dtype=torch.float32),
+        "normal": torch.ones((1, 2)),
+    }
+    save_file(a_tensors, str(model_a))
+    save_file({
+        "a_low": torch.tensor([[9.0, 9.0]]),
+        "b_low": torch.tensor([[3, 4]], dtype=torch.int8),
+        "normal": torch.full((1, 2), 3.0),
+    }, str(model_b))
+    paths = {"a": str(model_a), "b": str(model_b)}
+    monkeypatch.setattr(generic.folder_paths, "get_full_path", lambda _, name: paths[name])
+    monkeypatch.setattr(generic.folder_paths, "get_folder_paths", lambda _: [str(tmp_path)])
+    monkeypatch.setattr(generic.folder_paths, "models_dir", str(tmp_path))
+    monkeypatch.setattr(generic, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(generic, "cleanup_after_operation", lambda: None)
+    params = _generic_params("generic_guarded")
+    params.update({"model_a": "a", "model_b": "b"})
+
+    generic.MergerLogic.execute_merge(
+        {"model_a": "a", "model_b": "b"}, "Weight-Sum",
+        operations.TWO_MODEL_MODES, params, "loras",
+    )
+    result = load_file(
+        str(tmp_path / "loras" / "generic_guarded.safetensors")
+    )
+    torch.testing.assert_close(result["a_low"], a_tensors["a_low"])
+    torch.testing.assert_close(result["b_low"], a_tensors["b_low"])
+    assert result["a_low"].dtype == torch.uint8
+    assert result["b_low"].dtype == torch.float32
+    torch.testing.assert_close(result["normal"].float(), torch.full((1, 2), 2.0))
+
+
+def test_merge_to_model_preserves_low_bit_base_and_skips_low_bit_adapter_layer(
+    monkeypatch, tmp_path, modules
+):
+    resize = modules["resize"]
+    _patch_output(monkeypatch, resize, tmp_path)
+    base = tmp_path / "merge_guard_base.safetensors"
+    adapter = tmp_path / "merge_guard_adapter.safetensors"
+    base_quant = torch.tensor([[1, 2], [3, 4]], dtype=torch.uint8)
+    save_file({
+        "model.diffusion_model.base_quant.weight": base_quant,
+        "model.diffusion_model.guarded.weight": torch.zeros((2, 2)),
+        "model.diffusion_model.normal.weight": torch.zeros((2, 2)),
+    }, str(base))
+    save_file({
+        "diffusion_model.guarded.lora_down.weight": torch.tensor([[1, 1]], dtype=torch.int8),
+        "diffusion_model.guarded.lora_up.weight": torch.ones((2, 1)),
+        "diffusion_model.guarded.diff": torch.full((2, 2), 5.0),
+        "diffusion_model.normal.diff": torch.ones((2, 2)),
+    }, str(adapter))
+
+    output, report = resize.merge_loras_to_model(
+        [str(adapter)], [1.0], str(base), "cpu", torch.float16, "merge_guarded",
+        verbose=False, include_1d_diffs=True, return_report=True,
+    )
+    result = load_file(output)
+    torch.testing.assert_close(result["model.diffusion_model.base_quant.weight"], base_quant)
+    assert result["model.diffusion_model.base_quant.weight"].dtype == torch.uint8
+    torch.testing.assert_close(
+        result["model.diffusion_model.guarded.weight"].float(), torch.zeros((2, 2))
+    )
+    torch.testing.assert_close(
+        result["model.diffusion_model.normal.weight"].float(), torch.ones((2, 2))
+    )
+    assert report.startswith("COMPLETED WITH EXCLUSIONS: 1/2")
+    assert "[GUARDED] (1)" in report
+    assert "diffusion_model.guarded.lora_down.weight" in report
+    assert "torch.int8" in report
+    assert "model.diffusion_model.base_quant.weight" in report

tests/test_model_analysis.py

@@ -0,0 +1,282 @@
+import importlib
+import math
+import sys
+import types
+from pathlib import Path
+
+import pytest
+import torch
+from safetensors.torch import save_file
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+PACKAGE_NAME = "modelutils_analysis_tests"
+
+
+def _load_module(name):
+    if PACKAGE_NAME not in sys.modules:
+        package = types.ModuleType(PACKAGE_NAME)
+        package.__path__ = [str(REPO_ROOT)]
+        sys.modules[PACKAGE_NAME] = package
+    return importlib.import_module(f"{PACKAGE_NAME}.{name}")
+
+
+@pytest.fixture(scope="module")
+def analysis():
+    return _load_module("nodes.model_analysis")
+
+
+def _params(**overrides):
+    params = {
+        "cwb_similarity_alignment": False,
+        "top_weight_differences": 3,
+        "process_device": "cpu",
+        "force_clear_cache": False,
+    }
+    params.update(overrides)
+    return params
+
+
+def test_five_node_schema_contract(analysis):
+    assert len(analysis.MODEL_ANALYSIS_NODES) == 5
+    assert len({node.NODE_ID for node in analysis.MODEL_ANALYSIS_NODES}) == 5
+    for node in analysis.MODEL_ANALYSIS_NODES:
+        schema = node.define_schema()
+        assert schema.category == "ModelUtils/Analysis"
+        assert [item.id for item in schema.inputs[:3]] == [
+            "execution_mode", "model_a", "model_b",
+        ]
+        assert schema.inputs[0].options == ["ANALYZE", "DOCUMENTATION ONLY"]
+        assert "lazy_load" not in [item.id for item in schema.inputs]
+        input_ids = [item.id for item in schema.inputs]
+        if node in (analysis.LoRAModelAnalysis, analysis.EmbeddingModelAnalysis):
+            assert input_ids[3] == "cwb_similarity_alignment"
+        else:
+            assert "cwb_similarity_alignment" not in input_ids
+        assert not {
+            "cwb_preset", "consensus_type", "alignment_threshold",
+            "similarity_threshold", "power_alpha", "diversity_beta",
+            "dynamic_similarity_contrast", "soft_comfort_bandpass",
+            "position_weight", "preserve_common_prefix",
+        } & set(input_ids)
+        assert [item.display_name for item in schema.outputs] == [
+            "comparison_report", "cwb_report", "documentation",
+        ]
+
+
+def test_documentation_mode_has_three_separate_outputs(analysis):
+    result = analysis.DiffusionModelAnalysis.execute(execution_mode="DOCUMENTATION ONLY")
+    assert len(result.result) == 3
+    assert "No comparison performed" in result[0]
+    assert "No CWB analysis performed" in result[1]
+    assert result[2].startswith("# Two-Model Similarity Analysis")
+
+
+def test_raw_metrics_have_independent_expected_values(analysis):
+    a = torch.tensor([[1.0, 2.0], [3.0, 4.0]])
+    b = torch.tensor([[1.0, 0.0], [5.0, 4.0]])
+    raw, top = analysis._compute_raw(a, b, 2)
+    metrics = analysis._metrics(raw)
+    assert metrics["mae"] == pytest.approx(1.0)
+    assert metrics["mse"] == pytest.approx(2.0)
+    assert metrics["rmse"] == pytest.approx(math.sqrt(2.0))
+    assert metrics["max"] == pytest.approx(2.0)
+    assert metrics["exact"] == pytest.approx(0.5)
+    assert len(top) == 2
+    assert {entry[1] for entry in top} == {1, 2}
+
+
+def test_cwb_derived_metrics_have_independent_expected_values(analysis):
+    a = torch.tensor([[1.0, 0.0], [1.0, 0.0]])
+    b = torch.tensor([[1.0, 0.0], [0.0, 1.0]])
+    values = analysis._compute_cwb(a, b, False).values()
+    assert values["pair_mean"] == pytest.approx(0.5)
+    assert values["pair_min"] == pytest.approx(0.0)
+    assert values["pair_max"] == pytest.approx(1.0)
+    assert values["mean_aff_a"] == pytest.approx((1.0 + 2 ** -0.5) / 2)
+    assert values["mean_aff_b"] == pytest.approx((1.0 + 2 ** -0.5) / 2)
+    assert values["median_aff_a"] == pytest.approx(0.5)
+    assert values["median_aff_b"] == pytest.approx(0.5)
+    assert values["alignment"] is None
+
+
+def test_zero_constant_and_nonfinite_metrics_are_explicit(analysis):
+    raw, _ = analysis._compute_raw(torch.zeros(3), torch.zeros(3), 0)
+    metrics = analysis._metrics(raw)
+    assert metrics["cosine"] == pytest.approx(1.0)
+    assert metrics["pearson"] is None
+    assert metrics["norm_ratio"] == pytest.approx(1.0)
+
+    raw, _ = analysis._compute_raw(
+        torch.tensor([1.0, float("nan"), float("inf")]),
+        torch.tensor([2.0, 4.0, 5.0]),
+        0,
+    )
+    metrics = analysis._metrics(raw)
+    assert raw.finite == 1
+    assert raw.nonfinite_a == 2
+    assert metrics["coverage"] == pytest.approx(1 / 3)
+
+
+def test_end_to_end_reports_and_topology_classification(monkeypatch, tmp_path, analysis):
+    path_a = tmp_path / "a.safetensors"
+    path_b = tmp_path / "b.safetensors"
+    save_file({
+        "model.blocks.0.weight": torch.tensor([[1.0, 2.0], [3.0, 4.0]]),
+        "only_a": torch.ones(1),
+        "shape": torch.ones(2),
+        "integer": torch.tensor([1, 2], dtype=torch.int64),
+    }, str(path_a), metadata={"source": "a"})
+    save_file({
+        "model.blocks.0.weight": torch.tensor([[1.0, 0.0], [5.0, 4.0]]),
+        "only_b": torch.ones(1),
+        "shape": torch.ones(3),
+        "integer": torch.tensor([1, 3], dtype=torch.int64),
+    }, str(path_b), metadata={"source": "b"})
+    paths = {"a": str(path_a), "b": str(path_b)}
+    monkeypatch.setattr(analysis.folder_paths, "get_full_path", lambda _, name: paths[name])
+    monkeypatch.setattr(analysis, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(analysis, "cleanup_after_operation", lambda: None)
+    processed = []
+    async_calls = []
+    original_mark_processed = analysis.MemoryEfficientSafeOpen.mark_processed
+    original_async_stream = analysis.MemoryEfficientSafeOpen.async_stream
+
+    def track_processed(loader, key):
+        processed.append((loader.filename, key))
+        return original_mark_processed(loader, key)
+
+    def track_async_stream(loader, keys, **kwargs):
+        async_calls.append((loader.filename, tuple(keys), kwargs))
+        yield from original_async_stream(loader, keys, **kwargs)
+
+    monkeypatch.setattr(analysis.MemoryEfficientSafeOpen, "mark_processed", track_processed)
+    monkeypatch.setattr(analysis.MemoryEfficientSafeOpen, "async_stream", track_async_stream)
+
+    comparison, cwb_report = analysis.ModelAnalysisLogic.execute(
+        "a", "b", "diffusion_models", _params()
+    )
+    assert "Comparable floating tensors: 1" in comparison
+    assert "mae: 1" in comparison
+    assert "mse: 2" in comparison
+    assert "[A ONLY] (1)" in comparison
+    assert "[B ONLY] (1)" in comparison
+    assert "[SHAPE MISMATCH] (1)" in comparison
+    assert "[NON-FLOATING] (1)" in comparison
+    assert "model.blocks.0 |" in comparison
+    assert "CWB-DERIVED SIMILARITY SUMMARY" in cwb_report
+    assert "merge weighting" in cwb_report
+    assert "PRESET-IMPLIED CONTRIBUTIONS" not in cwb_report
+    assert "Average contribution" not in cwb_report
+    assert len(processed) == 4
+    assert {key for _, key in processed} == {"model.blocks.0.weight", "integer"}
+    assert len(async_calls) == 4
+    assert all(len(call[1]) == 1 for call in async_calls)
+    assert all(call[2]["batch_size"] == 1 for call in async_calls)
+    assert all(call[2]["prefetch_batches"] == 1 for call in async_calls)
+    assert not list(tmp_path.glob("**/*analysis*"))
+
+
+def test_lora_cross_format_keys_compare_as_one_logical_pair(monkeypatch, tmp_path, analysis):
+    path_a = tmp_path / "lora_a.safetensors"
+    path_b = tmp_path / "lora_b.safetensors"
+    save_file({
+        "diffusion_model.foo.lora_A.weight": torch.tensor([[1.0, 0.0], [0.0, 1.0]]),
+        "diffusion_model.foo.lora_B.weight": torch.tensor([[1.0, 0.0], [0.0, 1.0]]),
+    }, str(path_a))
+    save_file({
+        "base_model.model.diffusion_model.foo.lora_down.weight": torch.tensor([[1.0, 0.0], [0.0, 1.0]]),
+        "base_model.model.diffusion_model.foo.lora_up.weight": torch.tensor([[1.0, 0.0], [0.0, 1.0]]),
+    }, str(path_b))
+    paths = {"a": str(path_a), "b": str(path_b)}
+    monkeypatch.setattr(analysis.folder_paths, "get_full_path", lambda _, name: paths[name])
+    monkeypatch.setattr(analysis, "prepare_for_large_operation", lambda *args, **kwargs: None)
+    monkeypatch.setattr(analysis, "cleanup_after_operation", lambda: None)
+    processed = []
+    async_calls = []
+    original_mark_processed = analysis.MemoryEfficientSafeOpen.mark_processed
+    original_async_stream = analysis.MemoryEfficientSafeOpen.async_stream
+
+    def track_processed(loader, key):
+        processed.append((loader.filename, key))
+        return original_mark_processed(loader, key)
+
+    def track_async_stream(loader, keys, **kwargs):
+        async_calls.append((loader.filename, tuple(keys), kwargs))
+        yield from original_async_stream(loader, keys, **kwargs)
+
+    monkeypatch.setattr(analysis.MemoryEfficientSafeOpen, "mark_processed", track_processed)
+    monkeypatch.setattr(analysis.MemoryEfficientSafeOpen, "async_stream", track_async_stream)
+
+    comparison, cwb_report = analysis.ModelAnalysisLogic.execute(
+        "a", "b", "loras", _params(), lora_mode=True,
+    )
+    assert "Comparable floating tensors: 2" in comparison
+    assert "[A ONLY] (0)" in comparison
+    assert "[B ONLY] (0)" in comparison
+    assert "foo.down |" in comparison
+    assert "foo.up |" in comparison
+    assert "foo.[down+up] |" in cwb_report
+    assert len(processed) == 4
+    assert len({entry for entry in processed}) == 4
+    assert len(async_calls) == 2
+    assert all(len(call[1]) == 2 for call in async_calls)
+
+
+def test_embedding_similarity_alignment_is_explicit_opt_in(analysis):
+    a = torch.tensor([[1.0, 0.0], [0.0, 1.0]])
+    b = torch.tensor([[0.0, 1.0], [1.0, 0.0]])
+    aligned = analysis._compute_cwb(a, b, True).values()
+    indexed = analysis._compute_cwb(a, b, False).values()
+    assert aligned["alignment"] == pytest.approx(1.0)
+    assert aligned["alignment_gain"] == pytest.approx(1.0)
+    assert indexed["pair_mean"] == pytest.approx(0.0)
+    assert indexed["alignment"] is None
+
+
+def test_cuda_oom_retries_once_on_cpu(monkeypatch, analysis):
+    calls = []
+
+    def fake_compute(a, b, device, allow_alignment, top_count):
+        calls.append(device)
+        if device == "cuda":
+            raise torch.cuda.OutOfMemoryError("synthetic CUDA out of memory")
+        return analysis.RawStats(finite=1, total=1), analysis.CWBStats(), []
+
+    monkeypatch.setattr(analysis, "_compute_pair_on_device", fake_compute)
+    monkeypatch.setattr(analysis, "_release_cuda_oom", lambda: None)
+    result = analysis._analyze_with_fallback(
+        "layer", torch.ones(1), torch.ones(1), "cuda", False, 1,
+    )
+    assert calls == ["cuda", "cpu"]
+    assert result[-1] is True
+
+
+def test_non_oom_error_is_not_retried(monkeypatch, analysis):
+    calls = []
+
+    def fake_compute(*args):
+        calls.append(args[2])
+        raise RuntimeError("unrelated failure")
+
+    monkeypatch.setattr(analysis, "_compute_pair_on_device", fake_compute)
+    with pytest.raises(RuntimeError, match="unrelated failure"):
+        analysis._analyze_with_fallback(
+            "layer", torch.ones(1), torch.ones(1), "cuda", False, 1,
+        )
+    assert calls == ["cuda"]
+
+
+def test_failed_cpu_retry_names_layer_and_chains_oom(monkeypatch, analysis):
+    def fake_compute(a, b, device, allow_alignment, top_count):
+        if device == "cuda":
+            raise torch.cuda.OutOfMemoryError("synthetic CUDA out of memory")
+        raise ValueError("CPU calculation failed")
+
+    monkeypatch.setattr(analysis, "_compute_pair_on_device", fake_compute)
+    monkeypatch.setattr(analysis, "_release_cuda_oom", lambda: None)
+    with pytest.raises(RuntimeError, match="CPU fallback failed while analyzing 'named.layer'") as caught:
+        analysis._analyze_with_fallback(
+            "named.layer", torch.ones(1), torch.ones(1), "cuda", False, 1,
+        )
+    assert isinstance(caught.value.__cause__, torch.cuda.OutOfMemoryError)