Refactor loading strategies in mergers and LoRA resize nodes.

Feature/dora extract

__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
@@ -24,13 +24,31 @@ 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.lora_resize import (
     LoRAResizeFixed, LoRAResizeRatio,
     LoRAResizeFrobenius, LoRAResizeCumulative,
     LoRAMergeToModel
 )
+from .nodes.downloader_nodes import (
+    CheckpointInfoMetaDownloader, DiffusionModelInfoMetaDownloader, LoRAInfoMetaDownloader, EmbeddingInfoMetaDownloader,
+    VAEInfoMetaDownloader, ControlNetInfoMetaDownloader, ManualPathInfoMetaDownloader
+)
+from .nodes.model_info_nodes import (
+    CheckpointInfoLoader, LoRAInfoLoader, EmbeddingInfoLoader,
+    VAEInfoLoader, ControlNetInfoLoader, DiffusionModelInfoLoader
+)
 
 
+from .nodes.lora_rename import AnimaLoraRename
+
 class ModelUtilsExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
@@ -53,13 +71,27 @@ class ModelUtilsExtension(ComfyExtension):
             # LoRA Extraction
             LoRAExtractFixed, LoRAExtractRatio, LoRAExtractQuantile,
             LoRAExtractKnee, LoRAExtractFrobenius,
+            # DoRA Extraction
+            DoRAExtractFixed, DoRAExtractRatio, DoRAExtractQuantile,
+            DoRAExtractKnee, DoRAExtractFrobenius,
+            # Learned DoRA Extraction
+            DoRALearnedExtractFixed, DoRALearnedExtractRatio, DoRALearnedExtractQuantile,
+            DoRALearnedExtractKnee, DoRALearnedExtractFrobenius,
             # 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,
+            # Info Loaders
+            CheckpointInfoLoader, LoRAInfoLoader, EmbeddingInfoLoader,
+            VAEInfoLoader, ControlNetInfoLoader, DiffusionModelInfoLoader,
         ]
 
 

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:

nodes/dora_extract_wd.py

@@ -0,0 +1,981 @@
+"""
+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:
+                    # 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_nodes.py

@@ -0,0 +1,341 @@
+import os
+import folder_paths
+from comfy_api.latest import io
+from .downloader_utils import scan_and_process
+
+def get_subdirectories(folder_name: str) -> list[str]:
+    """Helper to get a list of subdirectories for a given model folder name."""
+    paths = folder_paths.get_folder_paths(folder_name)
+    if not paths:
+        return ["/"]
+
+    subdirs = set()
+    for base_path in paths:
+        if not os.path.exists(base_path):
+            continue
+        for root, dirs, _ in os.walk(base_path):
+            # Ignore hidden directories directly during walk
+            dirs[:] = [d for d in dirs if not d.startswith('.')]
+
+            rel_path = os.path.relpath(root, base_path)
+            if rel_path == ".":
+                subdirs.add("/")
+            else:
+                # Store with forward slashes for consistency
+                subdirs.add("/" + rel_path.replace("\\", "/"))
+
+    return sorted(list(subdirs))
+
+class BaseInfoMetaDownloaderNode(io.ComfyNode):
+    CATEGORY = "ModelUtils/Downloader"
+    OUTPUT_NODE = True
+
+    @classmethod
+    def get_common_inputs(cls):
+        return {
+            "recursive": io.Boolean.Input("recursive", default=True, tooltip="Whether to scan subdirectories recursively."),
+            "nsfw_level": io.Combo.Input("nsfw_level", options=["None", "Soft", "Mature", "X", "XXX", "All"], default="All", tooltip="Filter models based on NSFW level in metadata."),
+            "max_examples": io.Int.Input("max_examples", default=0, min=0, max=100, tooltip="Maximum number of examples to process."),
+            "threads": io.Int.Input("threads", default=4, min=1, max=32, tooltip="Number of threads to use for processing."),
+            "api_key": io.String.Input("api_key", default="", tooltip="API key for accessing the model repository.")
+        }
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name=cls.__name__.replace("Downloader", " Downloader"),
+            category=cls.CATEGORY,
+            inputs=list(cls.get_common_inputs().values()),
+            outputs=[io.String.Output(display_name="status")],
+            is_output_node=True,
+            description=("Scans specified directories for models, extracts metadata, generates previews, and optionally compresses files.")
+        )
+
+    @classmethod
+    def execute(cls, recursive, nsfw_level, max_examples, threads, api_key, **kwargs):
+        scan_dirs = cls.get_scan_dirs(**kwargs)
+        if not scan_dirs:
+            return io.NodeOutput("No directories to scan.")
+
+        # Call the scan_and_process function
+        stats = scan_and_process(
+            scan_dirs=scan_dirs,
+            recursive=recursive,
+            nsfw_level=nsfw_level,
+            max_examples=max_examples,
+            api_key=api_key if api_key else None,
+            threads=threads
+        )
+
+        mb_saved = stats.get("space_saved", 0) / (1024 * 1024)
+
+        output_msg = (
+            f"Scan Complete for {len(scan_dirs)} directory/directories.\n"
+            f"Total Models Processed: {stats.get('total_processed', 0)}\n"
+            f"New Models Hashed: {stats.get('new_hashes', 0)}\n"
+            f"Images Processed: {stats.get('images_processed', 0)}\n"
+            f"Workflows Extracted: {stats.get('workflows_extracted', 0)}\n"
+            f"Space Saved via Compression: {mb_saved:.2f} MB"
+        )
+
+        return io.NodeOutput(output_msg)
+
+    @classmethod
+    def get_scan_dirs(cls, **kwargs) -> list[str]:
+        return []
+
+class CheckpointInfoMetaDownloader(BaseInfoMetaDownloaderNode):
+    @classmethod
+    def define_schema(cls):
+        inputs = list(cls.get_common_inputs().values())
+        inputs.insert(0, io.Combo.Input("subdirectory", options=get_subdirectories("checkpoints"), default="/"))
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="Checkpoint Downloader",
+            category=cls.CATEGORY,
+            inputs=inputs,
+            outputs=[io.String.Output(display_name="status")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, subdirectory, recursive, nsfw_level, max_examples, threads, api_key, **kwargs):
+        return super().execute(
+            recursive=recursive, nsfw_level=nsfw_level, max_examples=max_examples,
+            threads=threads, api_key=api_key, subdirectory=subdirectory, **kwargs
+        )
+
+    @classmethod
+    def get_scan_dirs(cls, **kwargs):
+        base_paths = folder_paths.get_folder_paths("checkpoints")
+        if not base_paths:
+            return []
+        sub = kwargs.get("subdirectory", "/")
+        if sub == "/":
+            return base_paths
+
+        target_paths = []
+        for bp in base_paths:
+            target_path = os.path.join(bp, sub.lstrip("/"))
+            if os.path.exists(target_path):
+                target_paths.append(target_path)
+        return target_paths
+
+class DiffusionModelInfoMetaDownloader(BaseInfoMetaDownloaderNode):
+    @classmethod
+    def define_schema(cls):
+        inputs = list(cls.get_common_inputs().values())
+        inputs.insert(0, io.Combo.Input("subdirectory", options=get_subdirectories("diffusion_models"), default="/"))
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="Diffusion Model Downloader",
+            category=cls.CATEGORY,
+            inputs=inputs,
+            outputs=[io.String.Output(display_name="status")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, subdirectory, recursive, nsfw_level, max_examples, threads, api_key, **kwargs):
+        return super().execute(
+            recursive=recursive, nsfw_level=nsfw_level, max_examples=max_examples,
+            threads=threads, api_key=api_key, subdirectory=subdirectory, **kwargs
+        )
+
+    @classmethod
+    def get_scan_dirs(cls, **kwargs):
+        base_paths = folder_paths.get_folder_paths("diffusion_models")
+        if not base_paths:
+            return []
+        sub = kwargs.get("subdirectory", "/")
+        if sub == "/":
+            return base_paths
+
+        target_paths = []
+        for bp in base_paths:
+            target_path = os.path.join(bp, sub.lstrip("/"))
+            if os.path.exists(target_path):
+                target_paths.append(target_path)
+        return target_paths
+
+class LoRAInfoMetaDownloader(BaseInfoMetaDownloaderNode):
+    @classmethod
+    def define_schema(cls):
+        inputs = list(cls.get_common_inputs().values())
+        inputs.insert(0, io.Combo.Input("subdirectory", options=get_subdirectories("loras"), default="/"))
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="LoRA Downloader",
+            category=cls.CATEGORY,
+            inputs=inputs,
+            outputs=[io.String.Output(display_name="status")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, subdirectory, recursive, nsfw_level, max_examples, threads, api_key, **kwargs):
+        return super().execute(
+            recursive=recursive, nsfw_level=nsfw_level, max_examples=max_examples,
+            threads=threads, api_key=api_key, subdirectory=subdirectory, **kwargs
+        )
+
+    @classmethod
+    def get_scan_dirs(cls, **kwargs):
+        base_paths = folder_paths.get_folder_paths("loras")
+        if not base_paths:
+            return []
+        sub = kwargs.get("subdirectory", "/")
+        if sub == "/":
+            return base_paths
+
+        target_paths = []
+        for bp in base_paths:
+            target_path = os.path.join(bp, sub.lstrip("/"))
+            if os.path.exists(target_path):
+                target_paths.append(target_path)
+        return target_paths
+
+class EmbeddingInfoMetaDownloader(BaseInfoMetaDownloaderNode):
+    @classmethod
+    def define_schema(cls):
+        inputs = list(cls.get_common_inputs().values())
+        inputs.insert(0, io.Combo.Input("subdirectory", options=get_subdirectories("embeddings"), default="/"))
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="Embedding Downloader",
+            category=cls.CATEGORY,
+            inputs=inputs,
+            outputs=[io.String.Output(display_name="status")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, subdirectory, recursive, nsfw_level, max_examples, threads, api_key, **kwargs):
+        return super().execute(
+            recursive=recursive, nsfw_level=nsfw_level, max_examples=max_examples,
+            threads=threads, api_key=api_key, subdirectory=subdirectory, **kwargs
+        )
+
+    @classmethod
+    def get_scan_dirs(cls, **kwargs):
+        base_paths = folder_paths.get_folder_paths("embeddings")
+        if not base_paths:
+            return []
+        sub = kwargs.get("subdirectory", "/")
+        if sub == "/":
+            return base_paths
+
+        target_paths = []
+        for bp in base_paths:
+            target_path = os.path.join(bp, sub.lstrip("/"))
+            if os.path.exists(target_path):
+                target_paths.append(target_path)
+        return target_paths
+
+class VAEInfoMetaDownloader(BaseInfoMetaDownloaderNode):
+    @classmethod
+    def define_schema(cls):
+        inputs = list(cls.get_common_inputs().values())
+        inputs.insert(0, io.Combo.Input("subdirectory", options=get_subdirectories("vae"), default="/"))
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="VAE Downloader",
+            category=cls.CATEGORY,
+            inputs=inputs,
+            outputs=[io.String.Output(display_name="status")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, subdirectory, recursive, nsfw_level, max_examples, threads, api_key, **kwargs):
+        return super().execute(
+            recursive=recursive, nsfw_level=nsfw_level, max_examples=max_examples,
+            threads=threads, api_key=api_key, subdirectory=subdirectory, **kwargs
+        )
+
+    @classmethod
+    def get_scan_dirs(cls, **kwargs):
+        base_paths = folder_paths.get_folder_paths("vae")
+        if not base_paths:
+            return []
+        sub = kwargs.get("subdirectory", "/")
+        if sub == "/":
+            return base_paths
+
+        target_paths = []
+        for bp in base_paths:
+            target_path = os.path.join(bp, sub.lstrip("/"))
+            if os.path.exists(target_path):
+                target_paths.append(target_path)
+        return target_paths
+
+class ControlNetInfoMetaDownloader(BaseInfoMetaDownloaderNode):
+    @classmethod
+    def define_schema(cls):
+        inputs = list(cls.get_common_inputs().values())
+        inputs.insert(0, io.Combo.Input("subdirectory", options=get_subdirectories("controlnet"), default="/"))
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="ControlNet Downloader",
+            category=cls.CATEGORY,
+            inputs=inputs,
+            outputs=[io.String.Output(display_name="status")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, subdirectory, recursive, nsfw_level, max_examples, threads, api_key, **kwargs):
+        return super().execute(
+            recursive=recursive, nsfw_level=nsfw_level, max_examples=max_examples,
+            threads=threads, api_key=api_key, subdirectory=subdirectory, **kwargs
+        )
+
+    @classmethod
+    def get_scan_dirs(cls, **kwargs):
+        base_paths = folder_paths.get_folder_paths("controlnet")
+        if not base_paths:
+            return []
+        sub = kwargs.get("subdirectory", "/")
+        if sub == "/":
+            return base_paths
+
+        target_paths = []
+        for bp in base_paths:
+            target_path = os.path.join(bp, sub.lstrip("/"))
+            if os.path.exists(target_path):
+                target_paths.append(target_path)
+        return target_paths
+
+class ManualPathInfoMetaDownloader(BaseInfoMetaDownloaderNode):
+    @classmethod
+    def define_schema(cls):
+        inputs = list(cls.get_common_inputs().values())
+        inputs.append(io.String.Input("scan_dir", default=""))
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="Manual Path Downloader",
+            category=cls.CATEGORY,
+            inputs=inputs,
+            outputs=[io.String.Output(display_name="status")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, recursive, nsfw_level, max_examples, threads, api_key, scan_dir, **kwargs):
+        return super().execute(
+            recursive=recursive,
+            nsfw_level=nsfw_level,
+            max_examples=max_examples,
+            threads=threads,
+            api_key=api_key,
+            scan_dir=scan_dir,
+            **kwargs
+        )
+
+    @classmethod
+    def get_scan_dirs(cls, **kwargs):
+        scan_dir = kwargs.get("scan_dir", "")
+        if scan_dir.strip():
+            return [scan_dir.strip()]
+        return []

nodes/downloader_utils.py

@@ -0,0 +1,695 @@
+import os
+import sys
+import hashlib
+import json
+import sqlite3
+import requests
+import glob
+from pathlib import Path
+import threading
+import concurrent.futures
+import codecs
+import shutil
+
+try:
+    from PIL import Image
+    HAS_PIL = True
+except ImportError:
+    HAS_PIL = False
+    print("Warning: Pillow not found. Image workflow extraction disabled.")
+
+try:
+    from mutagen.mp4 import MP4
+    HAS_MUTAGEN = True
+except ImportError:
+    HAS_MUTAGEN = False
+    print("Warning: mutagen not found. MP4 tag extraction will rely entirely on av fallback.")
+
+try:
+    import av
+    HAS_AV = True
+except ImportError:
+    HAS_AV = False
+    print("Warning: av (PyAV) not found. MP4 metadata extraction fallback disabled.")
+
+# Add the reference module to the Python path
+# 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'))
+from unifiedefficientloader import UnifiedSafetensorsLoader
+
+# --- Database Path Logic ---
+try:
+    import folder_paths
+    base_user_dir = folder_paths.get_public_user_directory("default")
+    if not base_user_dir:
+        base_user_dir = root_dir
+except (ImportError, AttributeError):
+    base_user_dir = root_dir
+
+PLUGIN_DATA_DIR = os.path.join(base_user_dir, "ComfyUI-ModelUtils")
+os.makedirs(PLUGIN_DATA_DIR, exist_ok=True)
+DEFAULT_DB_PATH = os.path.join(PLUGIN_DATA_DIR, "cache.db")
+
+
+# --- Workflow Extraction Logic ---
+
+def unescape_and_parse_nested_json(json_string):
+    if not isinstance(json_string, str): return None
+    stripped = json_string.strip()
+    if not (stripped.startswith('{') or stripped.startswith('[')): return None
+    try: return json.loads(stripped)
+    except json.JSONDecodeError: pass
+    try:
+        first_parse = json.loads(stripped)
+        if isinstance(first_parse, str): return json.loads(first_parse)
+    except json.JSONDecodeError: pass
+    try:
+        decoded = codecs.decode(stripped, "unicode_escape")
+        return json.loads(decoded)
+    except (json.JSONDecodeError, UnicodeDecodeError, ValueError): pass
+    try:
+        unescaped = stripped.replace('\\\\"', '\\"').replace('\\"', '"').replace("\\\\", "\\")
+        unescaped = unescaped.replace("\\n", "\n").replace("\\r", "\r").replace("\\t", "\t")
+        return json.loads(unescaped)
+    except json.JSONDecodeError: pass
+    return None
+
+def get_workflow_blob(metadata):
+    for key in ['workflow', 'prompt']:
+        val = metadata.get(key)
+        if val:
+            parsed = unescape_and_parse_nested_json(val)
+            if parsed: return parsed
+    for val in metadata.values():
+        parsed = unescape_and_parse_nested_json(val)
+        if isinstance(parsed, dict) and (any(k.isdigit() for k in parsed.keys()) or 'nodes' in parsed):
+            return parsed
+    return None
+
+def process_image_workflow_and_compress(filepath_str):
+    if not HAS_PIL: return {"images_processed": 0, "workflows_extracted": 0, "space_saved": 0}
+    p = Path(filepath_str)
+    if p.suffix.lower() not in ['.png', '.jpg', '.jpeg']: return {"images_processed": 0, "workflows_extracted": 0, "space_saved": 0}
+
+    workflow = None
+    actual_format = None
+    stats = {"images_processed": 1, "workflows_extracted": 0, "space_saved": 0}
+    try:
+        with Image.open(p) as img:
+            actual_format = img.format
+            # CivitAI sometimes serves PNGs with .jpeg extension in URL
+            if actual_format == 'PNG':
+                workflow = get_workflow_blob(img.info)
+                if workflow:
+                    # Save workflow
+                    json_path = p.with_suffix('.json')
+                    if not json_path.exists():
+                        with open(json_path, 'w', encoding='utf-8') as f:
+                            json.dump(workflow, f, indent=2, ensure_ascii=False)
+                        print(f"Extracted workflow to {json_path}")
+                        stats["workflows_extracted"] = 1
+
+                    # Compress image to WebP
+                    webp_path = p.with_suffix('.webp')
+                    if not webp_path.exists():
+                        orig_size = os.path.getsize(p)
+                        img.save(webp_path, 'WEBP', quality=85, method=4)
+                        new_size = os.path.getsize(webp_path)
+                        stats["space_saved"] = max(0, orig_size - new_size)
+                        print(f"Compressed {p.name} (was actually {actual_format}) to {webp_path.name}")
+
+        # Delete original if we created a webp replacement
+        if workflow and p.with_suffix('.webp').exists() and actual_format == 'PNG':
+            os.remove(p)
+            print(f"Deleted original uncompressed image: {p.name}")
+    except Exception as e:
+        print(f"Failed to process image {p.name}: {e}")
+    return stats
+
+# --- Video Workflow Logic ---
+def get_tag_value(mp4, keys):
+    for key in keys:
+        if key in mp4.tags: return mp4.tags[key][0] if mp4.tags[key] else None
+    for key in keys:
+        search_suffix = key.split(':')[-1].lower()
+        for tag_key, val in mp4.tags.items():
+            if tag_key.lower().endswith(f":{search_suffix}"): return val[0] if val else None
+    return None
+
+def extract_metadata_av(mp4_path):
+    if not HAS_AV: return None
+    try:
+        with av.open(str(mp4_path)) as container:
+            return {k.lower(): v for k, v in container.metadata.items()}
+    except Exception as e:
+        print(f"av extraction error for {mp4_path}: {e}")
+        return None
+
+def process_video_workflow(filepath_str):
+    p = Path(filepath_str)
+    if p.suffix.lower() != '.mp4': return {"images_processed": 0, "workflows_extracted": 0, "space_saved": 0}
+
+    json_path = p.with_suffix('.json')
+    if json_path.exists(): return {"images_processed": 0, "workflows_extracted": 0, "space_saved": 0}
+
+    workflow_data = None
+    mp4 = None
+    stats = {"images_processed": 0, "workflows_extracted": 0, "space_saved": 0}
+    if HAS_MUTAGEN:
+        try: mp4 = MP4(p)
+        except: pass
+
+    if mp4:
+        for tag_keys in [['----:QuickTime:Workflow', 'workflow'], ['----:QuickTime:Prompt', 'prompt'], ['©cmt']]:
+            val = get_tag_value(mp4, tag_keys)
+            if val:
+                if isinstance(val, bytes): val = val.decode('utf-8', errors='ignore')
+                try:
+                    parsed = json.loads(val)
+                    if isinstance(parsed, dict) and ('prompt' in parsed or any(k.isdigit() for k in parsed.keys()) or 'nodes' in parsed):
+                        if 'prompt' in parsed and isinstance(parsed['prompt'], str):
+                            try: workflow_data = json.loads(parsed['prompt'])
+                            except: workflow_data = parsed
+                        else:
+                            workflow_data = parsed
+                        break
+                except json.JSONDecodeError:
+                    # Try more aggressive parsing if it's a nested string
+                    parsed = unescape_and_parse_nested_json(val)
+                    if parsed:
+                        workflow_data = parsed
+                        break
+
+    if not workflow_data:
+        av_data = extract_metadata_av(p)
+        if av_data:
+            for key in ['workflow', 'prompt', 'comment', 'description']:
+                if key in av_data:
+                    val = av_data[key]
+                    try:
+                        parsed = json.loads(val)
+                        if isinstance(parsed, dict):
+                            if 'prompt' in parsed and isinstance(parsed['prompt'], str):
+                                try: workflow_data = json.loads(parsed['prompt'])
+                                except: workflow_data = parsed
+                            else:
+                                workflow_data = parsed
+                            break
+                    except json.JSONDecodeError:
+                        parsed = unescape_and_parse_nested_json(val)
+                        if parsed:
+                            workflow_data = parsed
+                            break
+
+    if workflow_data:
+        try:
+            with open(json_path, 'w', encoding='utf-8') as f:
+                json.dump(workflow_data, f, indent=2, ensure_ascii=False)
+            print(f"Extracted workflow to {json_path.name}")
+            stats["workflows_extracted"] = 1
+        except Exception as e:
+            print(f"Error saving workflow for {p.name}: {e}")
+    return stats
+
+# --- Constants for Civitai NSFW levels ---
+NSFW_LEVELS = {
+    "None": 1,
+    "Soft": 2,
+    "Mature": 4,
+    "X": 8,
+    "XXX": 16,
+    "All": 32  # Anything
+}
+
+class HashCache:
+    def __init__(self, db_path=None):
+        self.db_path = db_path if db_path else DEFAULT_DB_PATH
+        self.is_sqlite = self.db_path.endswith('.db') or self.db_path.endswith('.sqlite')
+        self.lock = threading.Lock()
+        if self.is_sqlite:
+            self._init_sqlite()
+        else:
+            self.cache = self._load_json()
+
+    def _init_sqlite(self):
+        with sqlite3.connect(self.db_path) as conn:
+            cursor = conn.cursor()
+            cursor.execute('''
+                CREATE TABLE IF NOT EXISTS file_hashes (
+                    filepath TEXT PRIMARY KEY,
+                    mtime REAL,
+                    size INTEGER,
+                    hash TEXT
+                )
+            ''')
+            conn.commit()
+
+    def _load_json(self):
+        if os.path.exists(self.db_path):
+            with open(self.db_path, "r", encoding="utf-8") as f:
+                return json.load(f)
+        return {}
+
+    def _save_json(self):
+        with open(self.db_path, "w", encoding="utf-8") as f:
+            json.dump(self.cache, f, indent=4)
+
+    def _get_sqlite_hash(self, filepath, mtime, size):
+        with sqlite3.connect(self.db_path) as conn:
+            cursor = conn.cursor()
+            cursor.execute("SELECT mtime, size, hash FROM file_hashes WHERE filepath=?", (filepath,))
+            row = cursor.fetchone()
+            if row:
+                db_mtime, db_size, db_hash = row
+                if db_mtime == mtime and db_size == size:
+                    return db_hash
+        return None
+
+    def _set_sqlite_hash(self, filepath, mtime, size, file_hash):
+        with sqlite3.connect(self.db_path, timeout=10.0) as conn:
+            cursor = conn.cursor()
+            cursor.execute('''
+                INSERT OR REPLACE INTO file_hashes (filepath, mtime, size, hash)
+                VALUES (?, ?, ?, ?)
+            ''', (filepath, mtime, size, file_hash))
+            conn.commit()
+
+    def get_hash(self, filepath):
+        if not os.path.exists(filepath):
+            return None
+        mtime = os.path.getmtime(filepath)
+        size = os.path.getsize(filepath)
+        key = filepath
+
+        with self.lock:
+            if self.is_sqlite:
+                cached_hash = self._get_sqlite_hash(key, mtime, size)
+                if cached_hash: return cached_hash
+            else:
+                if key in self.cache:
+                    entry = self.cache[key]
+                    if entry["mtime"] == mtime and entry["size"] == size:
+                        return entry["hash"]
+
+        # Calculate new hash (outside the lock to allow parallel hashing)
+        print(f"Calculating SHA256 for: {filepath}")
+        sha256_hash = self.calculate_sha256(filepath)
+
+        with self.lock:
+            if self.is_sqlite:
+                self._set_sqlite_hash(key, mtime, size, sha256_hash)
+            else:
+                self.cache[key] = {
+                    "mtime": mtime,
+                    "size": size,
+                    "hash": sha256_hash
+                }
+                self._save_json()
+
+        return sha256_hash
+
+    @staticmethod
+    def calculate_sha256(filename):
+        sha256_hash = hashlib.sha256()
+        with open(filename, "rb") as f:
+            # Read and update hash string value in blocks of 4K
+            for byte_block in iter(lambda: f.read(4096), b""):
+                sha256_hash.update(byte_block)
+        return sha256_hash.hexdigest()
+
+class CivitaiAPI:
+    def __init__(self, api_key=None):
+        self.api_key = api_key
+        self.base_url = "https://civitai.com/api/v1"
+
+    def get_headers(self):
+        headers = {}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+        return headers
+
+    def get_model_info_by_hash(self, file_hash):
+        url = f"{self.base_url}/model-versions/by-hash/{file_hash}"
+        response = requests.get(url, headers=self.get_headers())
+        if response.status_code == 200:
+            return response.json()
+        elif response.status_code == 404:
+            return None
+        else:
+            response.raise_for_status()
+
+class Downloader:
+    @staticmethod
+    def download_file(url, target_path, headers=None):
+        if headers is None:
+            headers = {}
+
+        temp_path = target_path + ".downloading"
+
+        downloaded_size = 0
+        if os.path.exists(temp_path):
+            downloaded_size = os.path.getsize(temp_path)
+
+        if downloaded_size > 0:
+            headers["Range"] = f"bytes={downloaded_size}-"
+
+        response = requests.get(url, headers=headers, stream=True)
+
+        if response.status_code == 416: # Range not satisfiable, file probably changed or completed
+            os.remove(temp_path)
+            downloaded_size = 0
+            if "Range" in headers:
+                del headers["Range"]
+            response = requests.get(url, headers=headers, stream=True)
+
+        if response.status_code not in [200, 206]:
+            print(f"Failed to download from {url}, status code: {response.status_code}")
+            return False
+
+        mode = "ab" if response.status_code == 206 else "wb"
+
+        total_size = int(response.headers.get('content-length', 0)) + downloaded_size
+
+        print(f"Downloading: {target_path} ({downloaded_size}/{total_size} bytes)")
+
+        with open(temp_path, mode) as f:
+            for chunk in response.iter_content(chunk_size=8192):
+                if chunk:
+                    f.write(chunk)
+
+        os.rename(temp_path, target_path)
+        return True
+
+def get_potential_preview_files(base_path):
+    exts = ["png", "jpg", "jpeg", "webp", "mp4"]
+    files = []
+    for ext in exts:
+        files.append(f"{base_path}.preview.{ext}")
+        files.append(f"{base_path}.{ext}")
+    return files
+
+def preview_exists(base_path):
+    previews = get_potential_preview_files(base_path)
+    for p in previews:
+        if os.path.exists(p):
+            return True
+    return False
+
+def next_example_image_path(base_path, ext):
+    i = 0
+    while len(glob.glob(f"{base_path}.example.{i}.*")) > 0:
+        i += 1
+    return f"{base_path}.example.{i}.{ext}"
+
+def filter_images_by_nsfw(images, threshold_level):
+    valid_images = []
+    for img in images:
+        rating = img.get("nsfwLevel", 1)
+        if rating <= threshold_level:
+            valid_images.append(img)
+    return valid_images
+
+def extract_safetensors_metadata(filepath):
+    try:
+        # Use low_memory=True to only read the header and metadata without loading tensors
+        loader = UnifiedSafetensorsLoader(filepath, low_memory=True)
+        metadata = loader.metadata()
+        loader.close()
+        return metadata if metadata else None
+    except Exception as e:
+        print(f"Failed to read safetensors metadata from {filepath} using unifiedefficientloader: {e}")
+        return None
+
+def process_file(filepath, cache, api, nsfw_threshold, max_examples):
+    filepath_str = str(filepath)
+    base_path = os.path.splitext(filepath_str)[0]
+    is_safetensors = filepath_str.lower().endswith(".safetensors")
+
+    stats = {"new_hashes": 0, "images_processed": 0, "workflows_extracted": 0, "space_saved": 0}
+
+    # Check if all targets already exist
+    info_path = f"{base_path}.civitai.info"
+    st_metadata_path = f"{base_path}.metadata.json"
+
+    has_info = os.path.exists(info_path)
+    has_preview = preview_exists(base_path)
+    has_st_metadata = not is_safetensors or os.path.exists(st_metadata_path)
+
+    # Process existing images/videos for workflow extraction & compression BEFORE we possibly skip the file
+    for p in get_potential_preview_files(base_path):
+        if os.path.exists(p):
+            ext_lower = p.lower()
+            if ext_lower.endswith(('.png', '.jpg', '.jpeg')):
+                res = process_image_workflow_and_compress(p)
+                for k in stats: stats[k] += res.get(k, 0)
+            elif ext_lower.endswith('.mp4'):
+                res = process_video_workflow(p)
+                for k in stats: stats[k] += res.get(k, 0)
+
+    for ex in glob.glob(f"{base_path}.example.*"):
+        ext_lower = ex.lower()
+        if ext_lower.endswith(('.png', '.jpg', '.jpeg')):
+            res = process_image_workflow_and_compress(ex)
+            for k in stats: stats[k] += res.get(k, 0)
+        elif ext_lower.endswith('.mp4'):
+            res = process_video_workflow(ex)
+            for k in stats: stats[k] += res.get(k, 0)
+
+    # If everything exists and we don't want examples or already have enough examples, we might skip API logic
+    if has_info and has_preview and has_st_metadata and max_examples == 0:
+        print(f"Skipping fully populated model: {filepath_str}")
+        return stats
+
+    print(f"Processing model: {filepath_str}")
+
+    # Safetensors built-in metadata extraction
+    if is_safetensors and not has_st_metadata:
+        metadata = extract_safetensors_metadata(filepath_str)
+        if metadata:
+            with open(st_metadata_path, "w", encoding="utf-8") as f:
+                json.dump(metadata, f, indent=4)
+            print(f"Saved safetensors metadata to {st_metadata_path}")
+
+    mtime = os.path.getmtime(filepath_str)
+    size = os.path.getsize(filepath_str)
+
+    # We want to know if it's a NEW hash or cached
+    is_cached = False
+    with cache.lock:
+        if cache.is_sqlite:
+            cached_hash = cache._get_sqlite_hash(filepath_str, mtime, size)
+            if cached_hash: is_cached = True
+        else:
+            if filepath_str in cache.cache:
+                entry = cache.cache[filepath_str]
+                if entry["mtime"] == mtime and entry["size"] == size:
+                    is_cached = True
+
+    file_hash = cache.get_hash(filepath_str)
+    if not file_hash:
+        return stats
+
+    if not is_cached:
+        stats["new_hashes"] += 1
+
+    try:
+        model_info = api.get_model_info_by_hash(file_hash)
+    except Exception as e:
+        print(f"API Error for {filepath_str}: {e}")
+        return stats
+
+    if not model_info:
+        print(f"Model not found on Civitai: {filepath_str}")
+        return stats
+
+    # 1. Save Info
+    if not has_info:
+        with open(info_path, "w", encoding="utf-8") as f:
+            json.dump(model_info, f, indent=4)
+        print(f"Saved metadata to {info_path}")
+
+    images = model_info.get("images", [])
+    valid_images = filter_images_by_nsfw(images, nsfw_threshold)
+
+    if not valid_images:
+        print(f"No valid images found matching NSFW criteria for {filepath_str}")
+        return stats
+
+    # 2. Save Preview
+    if not has_preview and len(valid_images) > 0:
+        preview_img = valid_images[0]
+        url = preview_img.get("url")
+        ext = "jpeg" # Default fallback
+        if ".png" in url.lower(): ext = "png"
+        elif ".mp4" in url.lower(): ext = "mp4"
+        elif ".jpg" in url.lower() or ".jpeg" in url.lower(): ext = "jpeg"
+
+        preview_path = f"{base_path}.preview.{ext}"
+        if Downloader.download_file(url, preview_path):
+            if ext in ("png", "jpg", "jpeg"):
+                res = process_image_workflow_and_compress(preview_path)
+                for k in stats: stats[k] += res.get(k, 0)
+            elif ext == "mp4":
+                res = process_video_workflow(preview_path)
+                for k in stats: stats[k] += res.get(k, 0)
+
+    # 3. Save Examples
+    if max_examples > 0:
+        existing_examples_count = len(glob.glob(f"{base_path}.example.*"))
+        # We divide by 2 roughly if there are json files, let's just count unique example indices
+        existing_indices = set()
+        for f in glob.glob(f"{base_path}.example.*"):
+            parts = f.split('.example.')
+            if len(parts) > 1:
+                idx = parts[1].split('.')[0]
+                if idx.isdigit(): existing_indices.add(idx)
+
+        needed_examples = max_examples - len(existing_indices)
+
+        if needed_examples > 0:
+            example_images = valid_images[1:max_examples+1]
+            # Skip images we probably already downloaded by skipping the first N
+            example_images_to_dl = example_images[len(existing_indices):]
+
+            for img in example_images_to_dl:
+                url = img.get("url")
+                ext = "jpeg"
+                if ".png" in url.lower(): ext = "png"
+                elif ".mp4" in url.lower(): ext = "mp4"
+                elif ".jpg" in url.lower() or ".jpeg" in url.lower(): ext = "jpeg"
+
+                example_path = next_example_image_path(base_path, ext)
+                if Downloader.download_file(url, example_path):
+                    if ext in ("png", "jpg", "jpeg"):
+                        res = process_image_workflow_and_compress(example_path)
+                        for k in stats: stats[k] += res.get(k, 0)
+                    elif ext == "mp4":
+                        res = process_video_workflow(example_path)
+                        for k in stats: stats[k] += res.get(k, 0)
+    return stats
+
+import numpy as np
+from PIL import Image, ImageOps
+import torch
+
+def load_image_tensor(path):
+    if not os.path.exists(path):
+        return None
+    try:
+        img = Image.open(path)
+        img = ImageOps.exif_transpose(img)
+        image = img.convert("RGB")
+        image = np.array(image).astype(np.float32) / 255.0
+        image = torch.from_numpy(image)[None,]
+        return image
+    except Exception as e:
+        print(f"Error loading image {path}: {e}")
+        return None
+
+def get_model_workflows(base_path):
+    workflows = []
+    # Main workflow
+    main_json = f"{base_path}.json"
+    if os.path.exists(main_json):
+        workflows.append(main_json)
+
+    # Example workflows
+    # They are named like base_path.example.N.json
+    # or base_path.preview.json (if extracted from preview)
+    preview_json = f"{base_path}.preview.json"
+    if os.path.exists(preview_json) and preview_json not in workflows:
+        workflows.append(preview_json)
+
+    example_jsons = glob.glob(f"{base_path}.example.*.json")
+    # Sort them by index
+    example_jsons.sort()
+    for ej in example_jsons:
+        if ej not in workflows:
+            workflows.append(ej)
+
+    return workflows
+
+def get_model_metadata_file(base_path):
+    # Prefer civitai.info if available
+    civitai_info = f"{base_path}.civitai.info"
+    if os.path.exists(civitai_info):
+        return civitai_info
+
+    # Fallback to metadata.json (safetensors metadata)
+    metadata_json = f"{base_path}.metadata.json"
+    if os.path.exists(metadata_json):
+        return metadata_json
+
+    return None
+
+def scan_and_process(scan_dirs, recursive=True, db_path=None, nsfw_level="All", max_examples=0, api_key=None, threads=4):
+    """
+    Scans directories for models and fetches their metadata/previews.
+    Adapted for direct Python API calls rather than argparse CLI.
+    """
+    db_path = db_path if db_path else DEFAULT_DB_PATH
+    cache = HashCache(db_path)
+    api = CivitaiAPI(api_key)
+
+    model_exts = [".safetensors", ".ckpt", ".pt"]
+
+    files_to_process = []
+    if isinstance(scan_dirs, str):
+        scan_dirs = [scan_dirs]
+
+    for scan_dir in scan_dirs:
+        path = Path(scan_dir)
+        if recursive:
+            for ext in model_exts:
+                # rglob includes all directories, so we must filter out hidden ones
+                for f in path.rglob(f"*{ext}"):
+                    if not any(part.startswith('.') for part in f.parts if part != '.' and part != '..'):
+                        files_to_process.append(f)
+        else:
+            for ext in model_exts:
+                for f in path.glob(f"*{ext}"):
+                    if not any(part.startswith('.') for part in f.parts if part != '.' and part != '..'):
+                        files_to_process.append(f)
+
+    nsfw_threshold = NSFW_LEVELS.get(nsfw_level, NSFW_LEVELS["All"])
+
+    print(f"Found {len(files_to_process)} models to process.")
+
+    global_stats = {
+        "total_processed": len(files_to_process),
+        "new_hashes": 0,
+        "images_processed": 0,
+        "workflows_extracted": 0,
+        "space_saved": 0
+    }
+
+    # Import and use ComfyUI's ProgressBar
+    try:
+        import comfy.utils
+        pbar = comfy.utils.ProgressBar(len(files_to_process))
+    except ImportError:
+        pbar = None
+
+    if threads > 1:
+        print(f"Starting {threads} threads...")
+        with concurrent.futures.ThreadPoolExecutor(max_workers=threads) as executor:
+            futures = [executor.submit(process_file, fp, cache, api, nsfw_threshold, max_examples) for fp in files_to_process]
+            for future in concurrent.futures.as_completed(futures):
+                try:
+                    res = future.result()
+                    if res:
+                        for k in global_stats: global_stats[k] += res.get(k, 0)
+                except Exception as e:
+                    print(f"Error in thread: {e}")
+                if pbar:
+                    pbar.update(1)
+    else:
+        for filepath in files_to_process:
+            res = process_file(filepath, cache, api, nsfw_threshold, max_examples)
+            if res:
+                for k in global_stats: global_stats[k] += res.get(k, 0)
+            if pbar:
+                pbar.update(1)
+
+    return global_stats

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 +13,15 @@ import folder_paths
 import comfy.utils
 from tqdm import tqdm
 from comfy_api.latest import io
-from safetensors.torch import save_file
-from .merger_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
+
+
+
 
 # =============================================================================
 # Rank Index Functions - Determine LoRA rank from singular values
@@ -46,14 +49,14 @@ def _index_sv_ratio(S: torch.Tensor, ratio: float) -> int:
 
 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
@@ -63,7 +66,7 @@ def _index_sv_cumulative(S: torch.Tensor, target: float, max_rank: int = None) -
 
 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".
     """
@@ -75,10 +78,10 @@ def _index_sv_fro(S: torch.Tensor, target: float, max_rank: int = None) -> int:
     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
@@ -172,27 +175,27 @@ def _svd_extract_linear_lowrank(
 ) -> 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 - 1))
+    rank = max(1, min(rank, max_possible_rank))
-    
+
     # Check if decomposition is worthwhile
-    if rank >= max_possible_rank // 2:
+    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)
@@ -200,7 +203,7 @@ def _svd_extract_linear_lowrank(
         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:
@@ -208,14 +211,14 @@ def _svd_extract_linear_lowrank(
             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"
 
 
@@ -239,7 +242,7 @@ def _svd_extract_linear(
     """
     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)
@@ -257,7 +260,7 @@ def _svd_extract_linear(
     rank = _compute_rank(S, mode, mode_param, max_rank)
 
     # Check if decomposition is worthwhile
-    if rank >= min(out_dim, in_dim) // 2:
+    if rank >= min(out_dim, in_dim):
         return weight, "full"
 
     # Truncate to rank
@@ -303,7 +306,7 @@ def _svd_extract_conv(
 
     # Flatten for SVD
     if is_1x1:
-        mat = weight.squeeze()  # [out_ch, in_ch]
+        mat = weight.view(out_ch, in_ch)  # [out_ch, in_ch]
     else:
         mat = weight.reshape(out_ch, -1)  # [out_ch, in_ch*k*k]
 
@@ -382,8 +385,11 @@ def _extract_chunked_layer(
             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 = torch.stack(all_lora_down, dim=0).mean(dim=0)
+    combined_lora_down = all_lora_down[0]
     rank = combined_lora_down.shape[0]
 
     return combined_lora_up, combined_lora_down, rank
@@ -393,14 +399,22 @@ def _extract_chunked_layer(
 # Pattern Matching
 # =============================================================================
 
-def _compile_patterns(pattern_string: str) -> list:
+def _compile_patterns(pattern_string: str, glob_mode: bool = False) -> list:
-    """Compile regex patterns from whitespace-separated string."""
+    """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:
@@ -408,8 +422,15 @@ def _compile_patterns(pattern_string: str) -> list:
     return patterns
 
 
-def _matches_any_pattern(key: str, patterns: list) -> bool:
+def _matches_any_pattern(key: str, patterns: list, glob_mode: bool = False) -> bool:
-    """Check if key matches any pattern."""
+    """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)
 
 
@@ -425,6 +446,7 @@ 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,
@@ -433,9 +455,12 @@ def extract_lora_from_files(
     mismatch_mode: str = "skip",
     chunk_large_layers: bool = True,
     svd_niter: int = 2,
-) -> dict[str, torch.Tensor]:
+    lazy_load: bool = True,
+    force_clear_cache: bool = True,
+    glob_skip_patterns: 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
@@ -445,6 +470,7 @@ def extract_lora_from_files(
         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
@@ -452,61 +478,62 @@ 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.
     """
-    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)
     print(f"[LoRA 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)
+    handler_a = MemoryEfficientSafeOpen(model_a_path, low_memory=lazy_load)
-    handler_b = MemoryEfficientSafeOpen(model_b_path)
+    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}
 
-        for key in tqdm(weight_keys, desc="Extracting LoRA", unit="layers"):
+        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):
-                stats["skipped"] += 1
+                return "skipped", None
-                pbar.update(1)
-                continue
 
             # Load tensors with pinned memory for CUDA
             use_pinned = device == 'cuda'
-            
+
             if key not in keys_b:
                 if mismatch_mode == "skip":
-                    stats["skipped"] += 1
+                    return "skipped", None
-                    pbar.update(1)
+                if mismatch_mode == "error":
-                    continue
+                    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
+
+                if mismatch_mode == "zeros":
+                    # For zeros mode, we treat the missing base as a zeroed tensor of same shape
+                    # which is already captured by weight_diff = tensor_a
+                    pass
             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)
@@ -517,10 +544,12 @@ def extract_lora_from_files(
 
                 if tensor_a.shape != tensor_b.shape:
                     if mismatch_mode == "skip":
-                        stats["skipped"] += 1
-                        pbar.update(1)
                         del tensor_a, tensor_b
-                        continue
+                        return "skipped", None
+                    if mismatch_mode == "error":
+                        raise ValueError(f"Shape mismatch for {key}: {tensor_a.shape} vs {tensor_b.shape}")
+
+                    # For zeros/fallback, use tensor_a as the difference
                     weight_diff = tensor_a
                     del tensor_b
                 else:
@@ -529,19 +558,16 @@ def extract_lora_from_files(
 
             # Skip small differences
             if min_diff > 0 and weight_diff.abs().max() < min_diff:
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
+                return "skipped", None
-                continue
 
             # Skip 1D tensors
             if weight_diff.ndim < 2:
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
+                return "skipped", None
-                continue
 
             is_conv = weight_diff.ndim == 4
+            layer_results = {}
 
             try:
                 if is_conv:
@@ -562,33 +588,46 @@ def extract_lora_from_files(
                             weight_diff, num_chunks, mode, linear_param, device, linear_max_rank
                         )
                         if lora_up is not None:
-                            output_sd[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.lora_B.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
-                            output_sd[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.lora_A.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
-                            output_sd[f"{lora_name}.alpha"] = torch.tensor(rank).to(save_torch_dtype)
-                            stats["chunked"] += 1
                             del weight_diff
-                            pbar.update(1)
+                            return "chunked", layer_results
-                            continue
 
                 print(f"[LoRA Extract] Failed: {key}: {e}")
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
+                return "skipped", None
-                continue
 
             # Store result
             if mode_str == "full":
-                output_sd[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
-                stats["full"] += 1
+                status = "full"
             else:
                 lora_down, lora_up, _ = result
-                output_sd[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.lora_A.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
-                output_sd[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.lora_B.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
-                output_sd[f"{lora_name}.alpha"] = torch.tensor(lora_down.shape[0]).to(save_torch_dtype)
+                status = "extracted"
-                stats["extracted"] += 1
 
             del weight_diff
-            pbar.update(1)
+            return status, layer_results
+
+        writer = IncrementalSafetensorsWriter(output_path)
+        writer.__enter__()
+        try:
+            for key in tqdm(weight_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()
+
+                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")
@@ -598,22 +637,12 @@ 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")
+    return 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
 
 
 # =============================================================================
@@ -629,86 +658,56 @@ 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:]  # strip len("model.diffusion_model.")
-        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
         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 _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,
@@ -720,7 +719,10 @@ def _get_common_inputs():
         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="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."),
     ]
 
 
@@ -736,9 +738,9 @@ class LoRAExtractFixed(io.ComfyNode):
             description="Extract LoRA with specified fixed rank for each layer type.",
             inputs=[
                 *_get_model_inputs(),
-                io.Int.Input("linear_dim", default=64, min=1, max=4096,
+                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=4096,
+                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)"),
@@ -751,18 +753,20 @@ 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) -> io.NodeOutput:
+                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)
 
-        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
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers, svd_niter,
+            lazy_load, force_clear_cache, glob_skip_patterns
         )
 
-        return io.NodeOutput(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)
 
 
 class LoRAExtractRatio(io.ComfyNode):
@@ -781,9 +785,9 @@ class LoRAExtractRatio(io.ComfyNode):
                               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=3072,
+                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=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -794,18 +798,21 @@ 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) -> io.NodeOutput:
+                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)
 
-        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
+            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(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)
 
 
 class LoRAExtractQuantile(io.ComfyNode):
@@ -824,9 +831,9 @@ class LoRAExtractQuantile(io.ComfyNode):
                               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=3072,
+                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=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -837,18 +844,21 @@ 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) -> io.NodeOutput:
+                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)
 
-        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
+            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(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)
 
 
 class LoRAExtractKnee(io.ComfyNode):
@@ -865,9 +875,9 @@ class LoRAExtractKnee(io.ComfyNode):
                 *_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=3072,
+                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=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -878,18 +888,21 @@ 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) -> io.NodeOutput:
+                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)
 
-        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
+            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(_save_lora(output_sd, output_filename))
+        return io.NodeOutput(output_path)
 
 
 class LoRAExtractFrobenius(io.ComfyNode):
@@ -908,9 +921,9 @@ class LoRAExtractFrobenius(io.ComfyNode):
                               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=3072,
+                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=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -921,15 +934,18 @@ 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) -> io.NodeOutput:
+                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)
 
-        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
+            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(_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/lora_resize.py

@@ -12,11 +12,13 @@ import folder_paths
 import comfy.utils
 from tqdm import tqdm
 from comfy_api.latest import io
-from safetensors.torch import save_file
-from typing import Optional, Dict, Tuple, List
-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 typing import Optional, Dict, Tuple, List
+
+
+
 # Reuse rank functions from extraction module
 from .lora_extract_svd import (
     _index_sv_ratio,
@@ -189,7 +191,7 @@ def detect_lora_rank(handler: MemoryEfficientSafeOpen, pairs: Dict) -> Tuple[int
         # Get dim from down weight shape
         if network_dim is None:
             down_key = block_keys["down"]
-            shape = handler.header[down_key]["shape"]
+            shape = handler.get_shape(down_key)
             # Linear: [rank, in_features] or Conv: [rank, in_ch, k, k]
             network_dim = shape[0]
 
@@ -375,6 +377,8 @@ def resize_lora_file(
     output_filename: str,
     verbose: bool = True,
     svd_niter: int = 2,
+    lazy_load: bool = True,
+    force_clear_cache: bool = False,
 ) -> str:
     """
     Resize a LoRA file to a new rank.
@@ -389,6 +393,8 @@ def resize_lora_file(
         output_filename: Output filename (without extension)
         verbose: Print progress info
         svd_niter: Power iterations for SVD accuracy
+        lazy_load: Low memory mode: load tensors from disk on demand
+        force_clear_cache: Clear CUDA cache after each layer
 
     Returns:
         Path to saved resized LoRA
@@ -397,7 +403,7 @@ def resize_lora_file(
     lora_size_gb = estimate_model_size(lora_path)
     prepare_for_large_operation(lora_size_gb * 2, torch.device(device))
 
-    handler = MemoryEfficientSafeOpen(lora_path)
+    handler = MemoryEfficientSafeOpen(lora_path, low_memory=lazy_load)
 
     try:
         metadata = handler.metadata().copy()
@@ -416,60 +422,7 @@ def resize_lora_file(
             print(f"[LoRA Resize] Original dim={network_dim}, alpha={network_alpha:.1f}, scale={scale:.3f}")
             print(f"[LoRA Resize] Resizing with method={method_str}, max_rank={new_rank}")
 
-        output_sd = {}
+        # Build metadata and output path before loop (all values known at this point)
-        fro_list = []
-        pbar = comfy.utils.ProgressBar(len(pairs))
-
-        with torch.no_grad():
-            for block_name, block_keys in tqdm(pairs.items(), desc="Resizing layers", unit="layers"):
-                if "down" not in block_keys or "up" not in block_keys:
-                    pbar.update(1)
-                    continue
-
-                lora_down = handler.get_tensor(block_keys["down"])
-                lora_up = handler.get_tensor(block_keys["up"])
-
-                is_conv = len(lora_down.shape) == 4
-
-                # Transfer to GPU with pinned memory if available
-                if device == 'cuda':
-                    lora_down = transfer_to_gpu_pinned(lora_down, device, torch.float32)
-                    lora_up = transfer_to_gpu_pinned(lora_up, device, torch.float32)
-
-                # Merge and re-extract
-                if is_conv:
-                    weight = _merge_conv(lora_down, lora_up, device)
-                    result = _extract_conv(weight, new_rank, dynamic_method, dynamic_param, scale, svd_niter)
-                else:
-                    weight = _merge_linear(lora_down, lora_up, device)
-                    result = _extract_linear(weight, new_rank, dynamic_method, dynamic_param, scale, svd_niter)
-
-                del weight, lora_down, lora_up
-
-                fro_list.append(result['fro_retained'])
-
-                # Store using same format suffixes as input
-                down_suffix = format_info["down_suffix"]
-                up_suffix = format_info["up_suffix"]
-                alpha_suffix = format_info["alpha_suffix"]
-
-                # Standardize prefix using heuristics if possible
-                new_block_name = _format_lora_key(block_name)
-
-                output_sd[f"{new_block_name}{down_suffix}"] = result["lora_down"].to(save_dtype)
-                output_sd[f"{new_block_name}{up_suffix}"] = result["lora_up"].to(save_dtype)
-                if alpha_suffix:
-                    output_sd[f"{new_block_name}{alpha_suffix}"] = torch.tensor(result["new_alpha"]).to(save_dtype)
-
-                pbar.update(1)
-
-        if verbose and fro_list:
-            import numpy as np
-            avg_fro = np.mean(fro_list)
-            std_fro = np.std(fro_list)
-            print(f"[LoRA Resize] Average Frobenius retention: {avg_fro:.1%} ± {std_fro:.3f}")
-
-        # Update metadata
         if dynamic_method:
             metadata["ss_training_comment"] = f"Dynamic resize with {dynamic_method}: {dynamic_param} from dim {network_dim}"
             metadata["ss_network_dim"] = "Dynamic"
@@ -479,14 +432,79 @@ def resize_lora_file(
             metadata["ss_network_dim"] = str(new_rank)
             metadata["ss_network_alpha"] = str(scale * new_rank)
 
-        # Save
         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, metadata)
+        fro_list = []
-        print(f"[LoRA Resize] Saved to {output_path}")
+        pbar = comfy.utils.ProgressBar(len(pairs))
 
+        writer = IncrementalSafetensorsWriter(output_path, metadata=metadata)
+        writer.__enter__()
+        try:
+            with torch.no_grad():
+                for block_name, block_keys in tqdm(pairs.items(), desc="Resizing layers", unit="layers"):
+                    if "down" not in block_keys or "up" not in block_keys:
+                        pbar.update(1)
+                        continue
+
+                    lora_down = handler.get_tensor(block_keys["down"])
+                    lora_up = handler.get_tensor(block_keys["up"])
+
+                    is_conv = len(lora_down.shape) == 4
+
+                    # Transfer to GPU with pinned memory if available
+                    if device == 'cuda':
+                        lora_down = transfer_to_gpu_pinned(lora_down, device, torch.float32)
+                        lora_up = transfer_to_gpu_pinned(lora_up, device, torch.float32)
+
+                    # Merge and re-extract
+                    if is_conv:
+                        weight = _merge_conv(lora_down, lora_up, device)
+                        result = _extract_conv(weight, new_rank, dynamic_method, dynamic_param, scale, svd_niter)
+                    else:
+                        weight = _merge_linear(lora_down, lora_up, device)
+                        result = _extract_linear(weight, new_rank, dynamic_method, dynamic_param, scale, svd_niter)
+
+                    del weight, lora_down, lora_up
+
+                    fro_list.append(result['fro_retained'])
+
+                    # Store using same format suffixes as input
+                    down_suffix = format_info["down_suffix"]
+                    up_suffix = format_info["up_suffix"]
+                    alpha_suffix = format_info["alpha_suffix"]
+
+                    # Preserve original key format from input LoRA verbatim
+                    new_block_name = block_name
+
+                    block_sd = {
+                        f"{new_block_name}{down_suffix}": result["lora_down"].to(save_dtype),
+                        f"{new_block_name}{up_suffix}": result["lora_up"].to(save_dtype),
+                    }
+                    if alpha_suffix:
+                        block_sd[f"{new_block_name}{alpha_suffix}"] = torch.tensor(result["new_alpha"]).to(save_dtype)
+
+                    writer.write_dict(block_sd)
+
+                    del result
+                    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)
+
+        if verbose and fro_list:
+            import numpy as np
+            avg_fro = np.mean(fro_list)
+            std_fro = np.std(fro_list)
+            print(f"[LoRA Resize] Average Frobenius retention: {avg_fro:.1%} ± {std_fro:.3f}")
+
+        print(f"[LoRA Resize] Saved to {output_path}")
         return output_path
 
     finally:
@@ -518,17 +536,22 @@ class LoRAResizeFixed(io.ComfyNode):
                 io.String.Input("output_filename", default="resized_lora"),
                 io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
                 io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=False, tooltip="Clear CUDA cache after each layer (slower but saves VRAM)"),
             ],
             outputs=[io.String.Output(display_name="output_path")],
             is_output_node=True,
         )
 
     @classmethod
-    def execute(cls, lora_name, new_rank, svd_niter, output_filename, save_dtype, device) -> io.NodeOutput:
+    def execute(cls, lora_name, new_rank, svd_niter, output_filename, save_dtype, device, lazy_load, force_clear_cache) -> io.NodeOutput:
         lora_path = folder_paths.get_full_path_or_raise("loras", lora_name)
         dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
 
-        path = resize_lora_file(lora_path, new_rank, None, None, device, dtype, output_filename, svd_niter=svd_niter)
+        path = resize_lora_file(
+            lora_path, new_rank, None, None, device, dtype, output_filename,
+            svd_niter=svd_niter, lazy_load=lazy_load, force_clear_cache=force_clear_cache
+        )
         return io.NodeOutput(path)
 
 
@@ -552,17 +575,22 @@ class LoRAResizeRatio(io.ComfyNode):
                 io.String.Input("output_filename", default="resized_lora_ratio"),
                 io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
                 io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=False, tooltip="Clear CUDA cache after each layer (slower but saves VRAM)"),
             ],
             outputs=[io.String.Output(display_name="output_path")],
             is_output_node=True,
         )
 
     @classmethod
-    def execute(cls, lora_name, max_rank, ratio, output_filename, save_dtype, device) -> io.NodeOutput:
+    def execute(cls, lora_name, max_rank, ratio, output_filename, save_dtype, device, lazy_load, force_clear_cache) -> io.NodeOutput:
         lora_path = folder_paths.get_full_path_or_raise("loras", lora_name)
         dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
 
-        path = resize_lora_file(lora_path, max_rank, "sv_ratio", ratio, device, dtype, output_filename)
+        path = resize_lora_file(
+            lora_path, max_rank, "sv_ratio", ratio, device, dtype, output_filename,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache
+        )
         return io.NodeOutput(path)
 
 
@@ -586,17 +614,22 @@ class LoRAResizeFrobenius(io.ComfyNode):
                 io.String.Input("output_filename", default="resized_lora_fro"),
                 io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
                 io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=False, tooltip="Clear CUDA cache after each layer (slower but saves VRAM)"),
             ],
             outputs=[io.String.Output(display_name="output_path")],
             is_output_node=True,
         )
 
     @classmethod
-    def execute(cls, lora_name, max_rank, target, output_filename, save_dtype, device) -> io.NodeOutput:
+    def execute(cls, lora_name, max_rank, target, output_filename, save_dtype, device, lazy_load, force_clear_cache) -> io.NodeOutput:
         lora_path = folder_paths.get_full_path_or_raise("loras", lora_name)
         dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
 
-        path = resize_lora_file(lora_path, max_rank, "sv_fro", target, device, dtype, output_filename)
+        path = resize_lora_file(
+            lora_path, max_rank, "sv_fro", target, device, dtype, output_filename,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache
+        )
         return io.NodeOutput(path)
 
 
@@ -620,17 +653,22 @@ class LoRAResizeCumulative(io.ComfyNode):
                 io.String.Input("output_filename", default="resized_lora_cumulative"),
                 io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
                 io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=False, tooltip="Clear CUDA cache after each layer (slower but saves VRAM)"),
             ],
             outputs=[io.String.Output(display_name="output_path")],
             is_output_node=True,
         )
 
     @classmethod
-    def execute(cls, lora_name, max_rank, target, output_filename, save_dtype, device) -> io.NodeOutput:
+    def execute(cls, lora_name, max_rank, target, output_filename, save_dtype, device, lazy_load, force_clear_cache) -> io.NodeOutput:
         lora_path = folder_paths.get_full_path_or_raise("loras", lora_name)
         dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
 
-        path = resize_lora_file(lora_path, max_rank, "sv_cumulative", target, device, dtype, output_filename)
+        path = resize_lora_file(
+            lora_path, max_rank, "sv_cumulative", target, device, dtype, output_filename,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache
+        )
         return io.NodeOutput(path)
 
 
@@ -649,6 +687,8 @@ def merge_loras_to_model(
     output_filename: str,
     skip_patterns_str: str = "",
     verbose: bool = True,
+    lazy_load: bool = True,
+    force_clear_cache: bool = False,
 ) -> str:
     """
     Merge multiple LoRAs into a base model and save the result directly.
@@ -666,6 +706,8 @@ def merge_loras_to_model(
         output_filename: Output filename (without extension)
         skip_patterns_str: Regex patterns for layers to skip
         verbose: Print progress info
+        lazy_load: Low memory mode: load tensors from disk on demand
+        force_clear_cache: Clear CUDA cache after each layer
 
     Returns:
         Path to saved merged model
@@ -681,8 +723,8 @@ def merge_loras_to_model(
     prepare_for_large_operation(total_size_gb * 1.5, torch.device(device))
 
     # Open all files - use low_memory for base model to avoid OS page caching
-    base_handler = MemoryEfficientSafeOpen(base_model_path, low_memory=True)
+    base_handler = MemoryEfficientSafeOpen(base_model_path, low_memory=lazy_load)
-    lora_handlers = [MemoryEfficientSafeOpen(lp) for lp in lora_paths]
+    lora_handlers = [MemoryEfficientSafeOpen(lp, low_memory=lazy_load) for lp in lora_paths]
 
 
     try:
@@ -706,7 +748,7 @@ def merge_loras_to_model(
                 print(f"[LoRA Merge To Model] LoRA {i+1}: {format_info['format']}, {len(pairs)} layers, dim={network_dim}")
 
         # Common prefixes
-        BASE_PREFIXES = ["model.diffusion_model.", "diffusion_model.", "transformer.", "model."]
+        BASE_PREFIXES = ["model.diffusion_model.", "diffusion_model.", "transformer.", "model.", "net."]
         LORA_PREFIXES = [
             "lora_unet_", "lora_transformer_", "lora_te1_", "lora_te2_", "lora_te_",
             "lycoris_", "diffusion_model.", "transformer.", "unet."
@@ -752,7 +794,11 @@ def merge_loras_to_model(
         base_metadata = base_handler.metadata().copy() if base_handler.metadata() else {}
         base_metadata["merge_comment"] = f"Merged {len(lora_paths)} LoRAs with weights: {lora_weights}"
 
-        output_sd = {}
+        # Build output path before loop
+        base_dir = os.path.dirname(base_model_path)
+        os.makedirs(base_dir, exist_ok=True)
+        output_path = os.path.join(base_dir, f"{output_filename.strip()}.safetensors")
+
         stats = {"merged": 0, "copied": 0, "skipped": 0}
         base_keys = list(base_handler.keys())
         pbar = comfy.utils.ProgressBar(len(base_keys))
@@ -760,114 +806,117 @@ def merge_loras_to_model(
         if verbose:
             print(f"[LoRA Merge To Model] Processing {len(base_keys)} base model keys...")
 
-        with torch.no_grad():
+        writer = IncrementalSafetensorsWriter(output_path, metadata=base_metadata)
-            for base_key in tqdm(base_keys, desc="Merging to model", unit="keys"):
+        writer.__enter__()
-                # Check skip patterns
+        try:
-                if _matches_any_pattern(base_key, skip_patterns):
+            with torch.no_grad():
-                    stats["skipped"] += 1
+                for base_key in tqdm(base_keys, desc="Merging to model", unit="keys"):
-                    pbar.update(1)
+                    # Check skip patterns
-                    continue
+                    if _matches_any_pattern(base_key, skip_patterns):
+                        stats["skipped"] += 1
+                        pbar.update(1)
+                        continue
 
-                # Load base weight
+                    # Load base weight
-                cpu_base = base_handler.get_tensor(base_key)
+                    cpu_base = base_handler.get_tensor(base_key)
 
-                # Only process weight tensors for LoRA merging
+                    # Only process weight tensors for LoRA merging
-                if base_key.endswith(".weight"):
+                    if base_key.endswith(".weight"):
-                    core = extract_core_layer_base(base_key)
+                        core = extract_core_layer_base(base_key)
-                    core_underscored = core.replace(".", "_")
+                        core_underscored = core.replace(".", "_")
 
-                    # Check if any LoRA contributes to this layer
+                        # Check if any LoRA contributes to this layer
-                    if core_underscored in lora_lookup:
+                        if core_underscored in lora_lookup:
-                        # Transfer to GPU for computation
+                            # Transfer to GPU for computation
-                        if device == 'cuda':
+                            if device == 'cuda':
-                            base_weight = transfer_to_gpu_pinned(cpu_base, device, torch.float32)
+                                base_weight = transfer_to_gpu_pinned(cpu_base, device, torch.float32)
-                        else:
-                            base_weight = cpu_base.to(device=device, dtype=torch.float32)
-                        del cpu_base
-
-                        # Accumulate deltas from all contributing LoRAs
-                        for info, block_keys in lora_lookup[core_underscored]:
-                            is_full_diff = info["format_info"].get("is_full_diff", False)
-
-                            if is_full_diff:
-                                # Full diff format
-                                if "diff" not in block_keys:
-                                    continue
-                                cpu_diff = info["handler"].get_tensor(block_keys["diff"])
-                                if device == 'cuda':
-                                    delta = transfer_to_gpu_pinned(cpu_diff, device, torch.float32)
-                                else:
-                                    delta = cpu_diff.to(device=device, dtype=torch.float32)
-                                del cpu_diff
-                                effective_scale = info["weight"]
                             else:
-                                # Standard LoRA format
+                                base_weight = cpu_base.to(device=device, dtype=torch.float32)
-                                if "down" not in block_keys or "up" not in block_keys:
+                            del cpu_base
-                                    continue
 
-                                cpu_down = info["handler"].get_tensor(block_keys["down"])
+                            # Accumulate deltas from all contributing LoRAs
-                                cpu_up = info["handler"].get_tensor(block_keys["up"])
+                            for info, block_keys in lora_lookup[core_underscored]:
-                                if device == 'cuda':
+                                is_full_diff = info["format_info"].get("is_full_diff", False)
-                                    lora_down = transfer_to_gpu_pinned(cpu_down, device, torch.float32)
+
-                                    lora_up = transfer_to_gpu_pinned(cpu_up, device, torch.float32)
+                                if is_full_diff:
+                                    # Full diff format
+                                    if "diff" not in block_keys:
+                                        continue
+                                    cpu_diff = info["handler"].get_tensor(block_keys["diff"])
+                                    if device == 'cuda':
+                                        delta = transfer_to_gpu_pinned(cpu_diff, device, torch.float32)
+                                    else:
+                                        delta = cpu_diff.to(device=device, dtype=torch.float32)
+                                    del cpu_diff
+                                    effective_scale = info["weight"]
                                 else:
-                                    lora_down = cpu_down.to(device=device, dtype=torch.float32)
+                                    # Standard LoRA format
-                                    lora_up = cpu_up.to(device=device, dtype=torch.float32)
+                                    if "down" not in block_keys or "up" not in block_keys:
-                                del cpu_down, cpu_up
+                                        continue
 
-                                # Get alpha
+                                    cpu_down = info["handler"].get_tensor(block_keys["down"])
-                                if "alpha" in block_keys:
+                                    cpu_up = info["handler"].get_tensor(block_keys["up"])
-                                    alpha_tensor = info["handler"].get_tensor(block_keys["alpha"])
+                                    if device == 'cuda':
-                                    layer_alpha = float(alpha_tensor.item())
+                                        lora_down = transfer_to_gpu_pinned(cpu_down, device, torch.float32)
-                                else:
+                                        lora_up = transfer_to_gpu_pinned(cpu_up, device, torch.float32)
-                                    layer_alpha = float(info["network_dim"])
+                                    else:
-                                layer_scale = layer_alpha / info["network_dim"] if info["network_dim"] > 0 else 1.0
+                                        lora_down = cpu_down.to(device=device, dtype=torch.float32)
-                                effective_scale = layer_scale * info["weight"]
+                                        lora_up = cpu_up.to(device=device, dtype=torch.float32)
+                                    del cpu_down, cpu_up
 
-                                # Compute delta
+                                    # Get alpha
-                                is_conv = len(lora_down.shape) == 4
+                                    if "alpha" in block_keys:
-                                if is_conv:
+                                        alpha_tensor = info["handler"].get_tensor(block_keys["alpha"])
-                                    in_rank, in_size, kernel_size, k_ = lora_down.shape
+                                        layer_alpha = float(alpha_tensor.item())
-                                    out_size, out_rank, _, _ = lora_up.shape
+                                    else:
-                                    delta = lora_up.reshape(out_size, -1) @ lora_down.reshape(in_rank, -1)
+                                        layer_alpha = float(info["network_dim"])
-                                    delta = delta.reshape(out_size, in_size, kernel_size, kernel_size)
+                                    layer_scale = layer_alpha / info["network_dim"] if info["network_dim"] > 0 else 1.0
-                                else:
+                                    effective_scale = layer_scale * info["weight"]
-                                    delta = lora_up @ lora_down
-                                del lora_down, lora_up
 
-                            # Apply delta to base weight
+                                    # Compute delta
-                            base_weight = base_weight + effective_scale * delta
+                                    is_conv = len(lora_down.shape) == 4
-                            del delta
+                                    if is_conv:
+                                        in_rank, in_size, kernel_size, k_ = lora_down.shape
+                                        out_size, out_rank, _, _ = lora_up.shape
+                                        delta = lora_up.reshape(out_size, -1) @ lora_down.reshape(in_rank, -1)
+                                        delta = delta.reshape(out_size, in_size, kernel_size, kernel_size)
+                                    else:
+                                        delta = lora_up @ lora_down
+                                    del lora_down, lora_up
 
-                        # Store merged weight
+                                # Apply delta to base weight
-                        output_sd[base_key] = base_weight.to(save_dtype).cpu().contiguous()
+                                base_weight = base_weight + effective_scale * delta
-                        del base_weight
+                                del delta
-                        stats["merged"] += 1
+
+                            # Write merged weight
+                            writer.write(base_key, base_weight.to(save_dtype).cpu().contiguous())
+                            del base_weight
+                            stats["merged"] += 1
+                        else:
+                            # No LoRA contribution, write as-is
+                            writer.write(base_key, cpu_base.to(save_dtype).contiguous())
+                            del cpu_base
+                            stats["copied"] += 1
                     else:
-                        # No LoRA contribution, copy as-is
+                        # Non-weight tensor (bias, norm, etc.), write as-is
-                        output_sd[base_key] = cpu_base.to(save_dtype).contiguous()
+                        writer.write(base_key, cpu_base.to(save_dtype).contiguous())
                         del cpu_base
                         stats["copied"] += 1
-                else:
-                    # Non-weight tensor (bias, norm, etc.), copy as-is
-                    output_sd[base_key] = cpu_base.to(save_dtype).contiguous()
-                    del cpu_base
-                    stats["copied"] += 1
 
+                    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)
 
         if verbose:
             print(f"[LoRA Merge To Model] Done: {stats['merged']} merged, {stats['copied']} copied, {stats['skipped']} skipped")
 
-        # Save to base model directory
-        base_dir = os.path.dirname(base_model_path)
-        os.makedirs(base_dir, exist_ok=True)
-        output_path = os.path.join(base_dir, f"{output_filename.strip()}.safetensors")
-
-        save_file(output_sd, output_path, base_metadata)
         print(f"[LoRA Merge To Model] Saved to {output_path}")
-
         return output_path
 
     finally:
@@ -886,11 +935,11 @@ class LoRAMergeToModel(io.ComfyNode):
             node_id="LoRAMergeToModel",
             display_name="LoRA Merge To Model",
             category="ModelUtils/LoRA/Merge",
-            description="Merge 1-4 LoRAs into a base model and save the result. Saves to base model directory.",
+            description="Merge 1-8 LoRAs into a base model and save the result. Saves to base model directory.",
             inputs=[
                 io.Combo.Input("base_model", options=folder_paths.get_filename_list("diffusion_models"),
                               tooltip="Base model the LoRAs were trained on"),
-                io.Combo.Input("lora_count", options=["1", "2", "3", "4"], default="2",
+                io.Combo.Input("lora_count", options=["1", "2", "3", "4", "5", "6", "7", "8"], default="2",
                               tooltip="Number of LoRAs to merge"),
                 # LoRA 1
                 io.Combo.Input("lora_1", options=folder_paths.get_filename_list("loras"),
@@ -912,12 +961,34 @@ class LoRAMergeToModel(io.ComfyNode):
                               default="None", tooltip="Fourth LoRA"),
                 io.Float.Input("weight_4", default=1.0, min=-10.0, max=10.0, step=0.01,
                               tooltip="Weight strength for LoRA 4"),
+                # LoRA 5
+                io.Combo.Input("lora_5", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Fifth LoRA"),
+                io.Float.Input("weight_5", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 5"),
+                # LoRA 6
+                io.Combo.Input("lora_6", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Sixth LoRA"),
+                io.Float.Input("weight_6", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 6"),
+                # LoRA 7
+                io.Combo.Input("lora_7", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Seventh LoRA"),
+                io.Float.Input("weight_7", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 7"),
+                # LoRA 8
+                io.Combo.Input("lora_8", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Eighth LoRA"),
+                io.Float.Input("weight_8", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 8"),
                 # Settings
                 io.String.Input("skip_patterns", default="", multiline=True,
                                tooltip="Regex patterns for layers to skip"),
                 io.String.Input("output_filename", default="merged_model"),
                 io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
                 io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=False, tooltip="Clear CUDA cache after each layer (slower but saves VRAM)"),
             ],
             outputs=[io.String.Output(display_name="output_path")],
             is_output_node=True,
@@ -926,12 +997,13 @@ class LoRAMergeToModel(io.ComfyNode):
     @classmethod
     def execute(cls, base_model, lora_count,
                 lora_1, weight_1, lora_2, weight_2, lora_3, weight_3, lora_4, weight_4,
-                skip_patterns, output_filename, save_dtype, device) -> io.NodeOutput:
+                lora_5, weight_5, lora_6, weight_6, lora_7, weight_7, lora_8, weight_8,
+                skip_patterns, output_filename, save_dtype, device, lazy_load, force_clear_cache) -> io.NodeOutput:
 
         # Build LoRA list based on count
         count = int(lora_count)
-        lora_names = [lora_1, lora_2, lora_3, lora_4][:count]
+        lora_names = [lora_1, lora_2, lora_3, lora_4, lora_5, lora_6, lora_7, lora_8][:count]
-        lora_weights = [weight_1, weight_2, weight_3, weight_4][:count]
+        lora_weights = [weight_1, weight_2, weight_3, weight_4, weight_5, weight_6, weight_7, weight_8][:count]
 
         # Filter out "None" entries
         valid_loras = [(name, weight) for name, weight in zip(lora_names, lora_weights) if name != "None"]
@@ -951,5 +1023,7 @@ class LoRAMergeToModel(io.ComfyNode):
             save_dtype=dtype,
             output_filename=output_filename,
             skip_patterns_str=skip_patterns,
+            lazy_load=lazy_load,
+            force_clear_cache=force_clear_cache,
         )
         return io.NodeOutput(path)

nodes/merger.py

@@ -1,3 +1,4 @@
+import fnmatch
 import os
 import re
 import torch
@@ -5,13 +6,13 @@ 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
+
 
 def load_documentation_from_file(filename):
     """Loads documentation from a markdown file in the ../docs/ directory."""
@@ -24,26 +25,38 @@ 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):
+def _compile_patterns(pattern_string, glob_mode=False):
-    """Compiles whitespace-separated regex patterns into a list of compiled patterns.
+    """Compiles whitespace-separated patterns into a list.
 
-    Returns empty list if pattern_string is empty or whitespace-only.
+    In regex mode (default): returns compiled re.Pattern objects.
-    Raises ValueError if any pattern is invalid regex.
+      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:
+        if not pattern:
-            patterns.append(re.compile(pattern))
+            continue
-        except re.error as e:
+        if glob_mode:
-            raise ValueError(f"Invalid regex pattern '{pattern}': {e}")
+            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):
+def _matches_any_pattern(key, patterns, glob_mode=False):
-    """Returns True if key matches any of the compiled regex patterns (substring search)."""
+    """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 +89,9 @@ 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():
@@ -83,11 +99,11 @@ class MergerLogic:
                 path = folder_paths.get_full_path(model_type, name)
                 if not path:
                     raise FileNotFoundError(f"Model '{name}' not found.")
-                handlers[name] = MemoryEfficientSafeOpen(path)
+                handlers[name] = MemoryEfficientSafeOpen(path, low_memory=True)
 
         primary_handler = handlers[primary_model_name]
         all_keys = primary_handler.keys()
-        metadata = primary_handler.header.get("__metadata__", {})
+        metadata = primary_handler.metadata()
 
         # Convert mismatch_mode string to enum
         mismatch_mode_str = recipe_params.get('mismatch_mode', 'skip')
@@ -96,8 +112,9 @@ class MergerLogic:
         alignment_mode = recipe_params.get('alignment_mode', 'pad/crop')
 
         # Compile filter patterns
-        exclude_patterns = _compile_patterns(recipe_params.get('exclude_patterns', ''))
+        glob_mode = recipe_params.get('glob_patterns', False)
-        discard_patterns = _compile_patterns(recipe_params.get('discard_patterns', ''))
+        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 +132,25 @@ 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)
+
+        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)
@@ -124,65 +160,134 @@ class MergerLogic:
         discarded_keys = 0
         error_keys = []
 
-        for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
+        output_folder = "loras" if calc_mode == "SVD LoRA Extraction" else model_type
-            # Check discard patterns first - skip entirely
+        # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
-            if _matches_any_pattern(key, discard_patterns):
+        output_dir = folder_paths.get_folder_paths(output_folder)[-1]
-                discarded_keys += 1
+        os.makedirs(output_dir, exist_ok=True)
-                pbar.update(1)
+        output_filename = recipe_params.get("output_filename")
-                continue
+        output_path = os.path.join(output_dir, f"{output_filename}.safetensors")
 
-            # Pre-load Model A's tensor with pinned memory for CUDA
+        writer = None
-            cpu_tensor = primary_handler.get_tensor(key)
-            if process_device == 'cuda':
-                tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
-            else:
-                tensor_a = cpu_tensor.to(device=process_device, dtype=process_dtype)
-            del cpu_tensor
 
-            # Check exclude patterns - use Model A only, no merge
+        writer = IncrementalSafetensorsWriter(output_path, metadata=metadata)
-            if _matches_any_pattern(key, exclude_patterns):
+        writer.__enter__()
-                merged_state_dict[key] = tensor_a.to(save_torch_dtype).cpu().clone()
-                excluded_keys += 1
-                pbar.update(1)
-                continue
 
-            # Pass tensor_a metadata to recipes for zeros mode and fallback
+        batch_buffer = {}
-            recipe_params['_tensor_a'] = tensor_a
+        batch_bytes = 0
-            recipe_params['_tensor_a_shape'] = tensor_a.shape
+        batch_key_count = 0
-            recipe_params['_tensor_a_dtype'] = tensor_a.dtype
 
-            try:
+        def _flush_batch():
-                recipe = calc_mode_class.create_recipe(key=key, **recipe_params)
+            nonlocal batch_buffer, batch_bytes, batch_key_count
-                result = recipe.merge()
+            if not batch_buffer:
-            except MissingTensorError as e:
+                return
-                if mismatch_mode == MissingTensorBehavior.ERROR:
+            writer.write_dict(batch_buffer)
-                    raise ValueError(f"Layer mismatch error (mismatch_mode='error'): {e}")
+            pbar.update(batch_key_count)
-                result = None
+            batch_buffer = {}
-                error_keys.append(key)
+            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()
 
-            # Handle None result (mismatch occurred with skip mode)
+        with torch.no_grad():
-            if result is None:
+            for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
-                result = tensor_a
+                # Check discard patterns first - skip entirely
-                skipped_keys += 1
+                if _matches_any_pattern(key, discard_patterns, glob_mode=glob_mode):
+                    discarded_keys += 1
+                    continue
 
-            if isinstance(result, dict):
+                # Pre-load Model A's tensor with pinned memory for CUDA
-                for r_key, r_tensor in result.items():
+                cpu_tensor = primary_handler.get_tensor(key)
-                    merged_state_dict[r_key] = r_tensor.to(save_torch_dtype).cpu().clone()
-            else:
-                # Ensure compatibility with Model A's architecture.
-                # If alignment_mode is 'pad/crop', we crop results that were padded.
-                # If alignment_mode is 'interpolate', resizing happened during operators.
-                if alignment_mode == 'pad/crop':
-                    target_shape = recipe_params['_tensor_a_shape']
-                    if result.shape != target_shape:
-                        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.to(save_torch_dtype).cpu().clone()
+                # 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))
 
-            # Clean up tensor_a reference to allow GC
+                target_dtype = determine_target_dtype(original_dtypes, save_torch_dtype, override_dtype)
-            del recipe_params['_tensor_a']
+
-            pbar.update(1)
+                if process_device == 'cuda':
+                    tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
+                else:
+                    tensor_a = cpu_tensor.to(device=process_device, dtype=process_dtype)
+                del cpu_tensor
+
+                # Check exclude patterns - use Model A only, no merge
+                if _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
+                    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
+                recipe_params['_tensor_a'] = tensor_a
+                recipe_params['_tensor_a_shape'] = tensor_a.shape
+                recipe_params['_tensor_a_dtype'] = tensor_a.dtype
+
+                try:
+                    recipe = calc_mode_class.create_recipe(key=key, **recipe_params)
+                    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)
+
+                # Handle None result (mismatch occurred with skip mode)
+                if result is None:
+                    result = tensor_a
+                    skipped_keys += 1
+
+                if isinstance(result, dict):
+                    for r_key, r_tensor in result.items():
+                        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.
+                    # If alignment_mode is 'interpolate', resizing happened during operators.
+                    if alignment_mode == 'pad/crop':
+                        target_shape = recipe_params['_tensor_a_shape']
+                        if result.shape != target_shape:
+                            slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(result.shape, target_shape))
+                            result = result[slices]
+
+                    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.
+                recipe.clean()
+                del recipe_params['_tensor_a']
+                del tensor_a
+                del recipe
+                del result
+
+                if (batch_bytes >= flush_threshold_bytes or batch_key_count >= 32):
+                    _flush_batch()
+
+
+        _flush_batch()
+        writer.__exit__(None, None, None)
 
         # Log summary
         if excluded_keys > 0:
@@ -203,10 +308,9 @@ class MergerLogic:
         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"
@@ -230,16 +334,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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_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"),
@@ -250,8 +362,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -259,10 +372,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -285,16 +400,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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_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"),
@@ -305,8 +428,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -314,10 +438,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -340,16 +466,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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_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"),
@@ -360,8 +494,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -369,10 +504,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -395,16 +532,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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_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 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"),
@@ -415,8 +560,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -424,10 +570,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -450,16 +598,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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_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"),
@@ -470,8 +626,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -479,10 +636,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -508,16 +667,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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"),
@@ -528,8 +695,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -537,10 +705,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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)
@@ -564,16 +734,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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"),
@@ -584,8 +762,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -593,10 +772,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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)
@@ -620,16 +801,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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_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"),
@@ -640,8 +829,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -649,10 +839,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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)
@@ -676,16 +868,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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 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"),
@@ -696,8 +896,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -705,10 +906,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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)
@@ -732,16 +935,24 @@ 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("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, 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=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, 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"),
@@ -752,8 +963,9 @@ 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) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                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)
@@ -761,10 +973,12 @@ 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,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "epsilon": epsilon, "zeta": zeta, "seed": seed,
-            "output_filename": output_filename, "save_dtype": save_dtype,
+            "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)

nodes/merger_ops.py

@@ -4,6 +4,8 @@ import torch.nn.functional as F
 from enum import Enum
 
 
+from unifiedefficientloader import transfer_to_gpu_pinned
+
 class MissingTensorBehavior(Enum):
     """Controls behavior when a tensor key is missing from a model."""
     ERROR = "error"      # Raise exception (strict mode)
@@ -127,12 +129,21 @@ class Operation:
         raise NotImplementedError
 
     def recurse(self, operation):
-        source_tensors = [source_oper.merge() for source_oper in operation.sources]
+        self._source_tensors = [source_oper.merge() for source_oper in operation.sources]
-        return operation.oper(*source_tensors)
+        return operation.oper(*self._source_tensors)
 
     def merge(self):
         return self.merge_func(self)
 
+    def clean(self):
+        if hasattr(self, '_source_tensors'):
+            for t in self._source_tensors:
+                del t
+            del self._source_tensors
+        for source in self.sources:
+            if hasattr(source, 'clean'):
+                source.clean()
+
 class LoadTensor(Operation):
     def __init__(self, key, model_name, handlers, device, dtype,
                  on_missing=MissingTensorBehavior.ERROR, fallback_shape=None, fallback_dtype=None):
@@ -165,7 +176,17 @@ class LoadTensor(Operation):
                 dtype = self.fallback_dtype if self.fallback_dtype else self.dtype
                 return torch.zeros(self.fallback_shape, device=self.device, dtype=dtype)
 
-        return handler.get_tensor(self.key).to(device=self.device, dtype=self.dtype)
+        cpu_tensor = handler.get_tensor(self.key)
+        if self.device == 'cuda':
+            self._tensor = transfer_to_gpu_pinned(cpu_tensor, self.device, self.dtype)
+        else:
+            self._tensor = cpu_tensor.to(device=self.device, dtype=self.dtype)
+        del cpu_tensor
+        return self._tensor
+
+    def clean(self):
+        if hasattr(self, '_tensor'):
+            del self._tensor
 
 class Multiply(Operation):
     def __init__(self, key, alpha, *sources):
@@ -298,6 +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)
@@ -409,6 +477,10 @@ class PreloadedTensor(Operation):
     def merge(self) -> torch.Tensor:
         return self.tensor
 
+    def clean(self):
+        if hasattr(self, 'tensor'):
+            del self.tensor
+
 class WeightSum(CalcMode):
     name = 'Weight-Sum'
     description = 'A * (1 - α) + B * α. Simple linear interpolation.'
@@ -685,5 +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

@@ -1,373 +1,7 @@
 """
-Memory-efficient safetensors utilities with parallel I/O and pinned memory support.
+Utilities for model merging operations.
-
-Based on benchmarking, parallel reading with ThreadPoolExecutor provides 2-4x speedup
-for large models compared to sequential reads.
-
-Worker count is automatically optimized based on device capabilities when workers=None.
-Pinned memory enables faster CPU→GPU transfers (2-3x improvement).
 """
-import os
+from typing import Dict, List
-import mmap
-import json
-import struct
-import torch
-from concurrent.futures import ThreadPoolExecutor, as_completed
-from typing import Dict, List, Optional, Any, Tuple
-
-# Import ComfyUI's memory pinning functions if available
-try:
-    from comfy.model_management import pin_memory, unpin_memory, is_device_cuda
-    PINNING_AVAILABLE = True
-except ImportError:
-    PINNING_AVAILABLE = False
-    def pin_memory(tensor): return False
-    def unpin_memory(tensor): return False
-    def is_device_cuda(device): return str(device).startswith('cuda')
-
-
-class PinnedTensor:
-    """Context manager for automatic pinned memory management.
-
-    Usage:
-        with PinnedTensor(cpu_tensor) as pinned:
-            gpu_tensor = pinned.to('cuda')
-        # Memory automatically unpinned after transfer
-    """
-
-    def __init__(self, tensor: torch.Tensor, auto_pin: bool = True):
-        self.tensor = tensor
-        self.is_pinned = False
-        self.auto_pin = auto_pin and PINNING_AVAILABLE
-
-    def __enter__(self) -> torch.Tensor:
-        if self.auto_pin and self.tensor.device.type == 'cpu':
-            self.is_pinned = pin_memory(self.tensor)
-        return self.tensor
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        if self.is_pinned:
-            unpin_memory(self.tensor)
-            self.is_pinned = False
-
-
-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.
-
-    Args:
-        tensor: CPU tensor to transfer
-        device: Target GPU device
-        dtype: Optional dtype conversion
-
-    Returns:
-        Tensor on GPU device
-    """
-    if not PINNING_AVAILABLE or not is_device_cuda(torch.device(device)):
-        # Fall back to regular transfer
-        if dtype:
-            return tensor.to(device=device, dtype=dtype)
-        return tensor.to(device=device)
-
-    # Pin memory for faster transfer
-    was_pinned = pin_memory(tensor)
-    try:
-        if dtype:
-            result = tensor.to(device=device, dtype=dtype, non_blocking=True)
-        else:
-            result = tensor.to(device=device, non_blocking=True)
-        # Sync to ensure transfer is complete before unpinning
-        if was_pinned:
-            torch.cuda.current_stream().synchronize()
-        return result
-    finally:
-        if was_pinned:
-            unpin_memory(tensor)
-
-
-class MemoryEfficientSafeOpen:
-    """Memory-efficient safetensors file reader with mmap, parallel I/O, and pinned memory support.
-
-    Features:
-    - mmap mode: Zero-copy tensor access via memory-mapped file
-    - low_memory mode: Direct file reads to minimize OS page cache usage
-    - Parallel loading: Multi-threaded tensor reads for 2-4x speedup
-    - Sorted batch reads: Keys sorted by file offset for sequential I/O
-    - Auto-optimized workers: Adjusts parallelism based on device capabilities
-    - Pinned memory: Faster CPU→GPU transfers when transferring to CUDA
-
-    Args:
-        filename: Path to safetensors file
-        device: Target device (default 'cpu')
-        mmap_mode: Use memory-mapped file for zero-copy (default True)
-        low_memory: Use direct file reads to minimize memory (overrides mmap_mode)
-    """
-
-    def __init__(self, filename: str, device: str = 'cpu', mmap_mode: bool = True, low_memory: bool = False):
-        self.filename = filename
-        self.device = device
-        # low_memory mode forces mmap off to avoid OS page caching
-        self.low_memory = low_memory
-        self.mmap_mode = mmap_mode and not low_memory
-        self.header, self.header_size = self._read_header()
-        self.file = open(filename, "rb")
-        self.mmap_obj = None
-
-        if self.mmap_mode:
-            self.mmap_obj = mmap.mmap(self.file.fileno(), 0, access=mmap.ACCESS_READ)
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        if self.mmap_obj:
-            self.mmap_obj.close()
-        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 keys_sorted_by_offset(self) -> List[str]:
-        """Return keys sorted by file offset for optimal sequential I/O."""
-        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."""
-        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 self.mmap_mode and self.mmap_obj:
-            if offset_start != offset_end:
-                start = self.header_size + 8 + offset_start
-                end = self.header_size + 8 + offset_end
-                tensor_bytes = memoryview(self.mmap_obj)[start:end]
-            else:
-                tensor_bytes = None
-        else:
-            # Non-mmap mode: use pre-allocated bytearray with readinto for minimal copies
-            tensor_bytes = None
-            if offset_start != offset_end:
-                self.file.seek(self.header_size + 8 + offset_start)
-                # Pre-allocate writable buffer and read directly into it
-                tensor_bytes = bytearray(offset_end - offset_start)
-                self.file.readinto(tensor_bytes)
-
-        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:
-        """Load a tensor and transfer to GPU using pinned memory for faster transfer.
-
-        Args:
-            key: Tensor key to load
-            device: Target GPU device (default 'cuda')
-            dtype: Optional dtype conversion
-
-        Returns:
-            Tensor on GPU device
-        """
-        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]:
-        """Load multiple tensors in parallel using ThreadPoolExecutor.
-
-        Args:
-            keys: List of keys to load (default: all keys)
-            workers: Number of worker threads (default: auto-calculated based on device)
-            sort_by_offset: Sort keys by file offset before loading (default: True)
-
-        Returns:
-            Dict mapping key -> tensor
-        """
-        if keys is None:
-            keys = self.keys()
-
-        # Auto-calculate workers based on device capabilities
-        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:
-            # Sort by offset for better I/O patterns
-            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]:
-        """Load a tensor and decode as JSON dict (for config tensors)."""
-        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 {}
-
-        if self.mmap_mode and self.mmap_obj:
-            start = self.header_size + 8 + offset_start
-            end = self.header_size + 8 + offset_end
-            tensor_bytes = bytes(self.mmap_obj[start:end])
-        else:
-            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:
-            # Avoid extra copy if already a bytearray (low_memory mode)
-            if isinstance(tensor_bytes, bytearray):
-                byte_tensor = torch.frombuffer(tensor_bytes, dtype=torch.uint8)
-            else:
-                # mmap memoryview needs copy to create writable tensor
-                byte_tensor = torch.frombuffer(bytearray(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}. Your PyTorch version may be too old.")
-
-
-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)
-
 
 def get_union_keys(*handlers) -> List[str]:
     """Returns the union of all tensor keys across multiple model handlers."""
@@ -386,4 +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/model_info_nodes.py

@@ -0,0 +1,184 @@
+import os
+import folder_paths
+from comfy_api.latest import io
+from .downloader_utils import get_potential_preview_files, load_image_tensor, get_model_workflows, get_model_metadata_file
+
+class BaseModelInfoLoader(io.ComfyNode):
+    CATEGORY = "ModelUtils/Info"
+
+    @classmethod
+    def get_info(cls, category, name, workflow_index):
+        full_path = folder_paths.get_full_path(category, name)
+        if not full_path or not os.path.exists(full_path):
+            return (None, "{}", "{}")
+
+        base_path = os.path.splitext(full_path)[0]
+
+        # 1. Image Preview
+        image_tensor = None
+        previews = get_potential_preview_files(base_path)
+        for p in previews:
+            if os.path.exists(p) and not p.lower().endswith('.mp4'):
+                image_tensor = load_image_tensor(p)
+                if image_tensor is not None:
+                    break
+
+        # 2. Workflow
+        workflow_json = "{}"
+        workflows = get_model_workflows(base_path)
+        if workflows and 0 <= workflow_index < len(workflows):
+            try:
+                with open(workflows[workflow_index], 'r', encoding='utf-8') as f:
+                    workflow_json = f.read()
+            except Exception as e:
+                print(f"Error reading workflow {workflows[workflow_index]}: {e}")
+
+        # 3. Metadata
+        metadata_json = "{}"
+        metadata_file = get_model_metadata_file(base_path)
+        if metadata_file:
+            try:
+                with open(metadata_file, 'r', encoding='utf-8') as f:
+                    metadata_json = f.read()
+            except Exception as e:
+                print(f"Error reading metadata {metadata_file}: {e}")
+
+        return (image_tensor, workflow_json, metadata_json)
+
+class CheckpointInfoLoader(BaseModelInfoLoader):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="Checkpoint Info Loader",
+            category=cls.CATEGORY,
+            inputs=[
+                io.Combo.Input("checkpoint", options=folder_paths.get_filename_list("checkpoints")),
+                io.Int.Input("workflow_index", default=0, min=0, max=100)
+            ],
+            outputs=[
+                io.Image.Output(display_name="preview"),
+                io.String.Output(display_name="workflow_json"),
+                io.String.Output(display_name="metadata_json")
+            ]
+        )
+
+    @classmethod
+    def execute(cls, checkpoint, workflow_index):
+        res = cls.get_info("checkpoints", checkpoint, workflow_index)
+        return io.NodeOutput(res[0], res[1], res[2])
+
+class LoRAInfoLoader(BaseModelInfoLoader):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="LoRA Info Loader",
+            category=cls.CATEGORY,
+            inputs=[
+                io.Combo.Input("lora", options=folder_paths.get_filename_list("loras")),
+                io.Int.Input("workflow_index", default=0, min=0, max=100)
+            ],
+            outputs=[
+                io.Image.Output(display_name="preview"),
+                io.String.Output(display_name="workflow_json"),
+                io.String.Output(display_name="metadata_json")
+            ]
+        )
+
+    @classmethod
+    def execute(cls, lora, workflow_index):
+        res = cls.get_info("loras", lora, workflow_index)
+        return io.NodeOutput(res[0], res[1], res[2])
+
+class EmbeddingInfoLoader(BaseModelInfoLoader):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="Embedding Info Loader",
+            category=cls.CATEGORY,
+            inputs=[
+                io.Combo.Input("embedding", options=folder_paths.get_filename_list("embeddings")),
+                io.Int.Input("workflow_index", default=0, min=0, max=100)
+            ],
+            outputs=[
+                io.Image.Output(display_name="preview"),
+                io.String.Output(display_name="workflow_json"),
+                io.String.Output(display_name="metadata_json")
+            ]
+        )
+
+    @classmethod
+    def execute(cls, embedding, workflow_index):
+        res = cls.get_info("embeddings", embedding, workflow_index)
+        return io.NodeOutput(res[0], res[1], res[2])
+
+class VAEInfoLoader(BaseModelInfoLoader):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="VAE Info Loader",
+            category=cls.CATEGORY,
+            inputs=[
+                io.Combo.Input("vae", options=folder_paths.get_filename_list("vae")),
+                io.Int.Input("workflow_index", default=0, min=0, max=100)
+            ],
+            outputs=[
+                io.Image.Output(display_name="preview"),
+                io.String.Output(display_name="workflow_json"),
+                io.String.Output(display_name="metadata_json")
+            ]
+        )
+
+    @classmethod
+    def execute(cls, vae, workflow_index):
+        res = cls.get_info("vae", vae, workflow_index)
+        return io.NodeOutput(res[0], res[1], res[2])
+
+class ControlNetInfoLoader(BaseModelInfoLoader):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="ControlNet Info Loader",
+            category=cls.CATEGORY,
+            inputs=[
+                io.Combo.Input("controlnet", options=folder_paths.get_filename_list("controlnet")),
+                io.Int.Input("workflow_index", default=0, min=0, max=100)
+            ],
+            outputs=[
+                io.Image.Output(display_name="preview"),
+                io.String.Output(display_name="workflow_json"),
+                io.String.Output(display_name="metadata_json")
+            ]
+        )
+
+    @classmethod
+    def execute(cls, controlnet, workflow_index):
+        res = cls.get_info("controlnet", controlnet, workflow_index)
+        return io.NodeOutput(res[0], res[1], res[2])
+
+class DiffusionModelInfoLoader(BaseModelInfoLoader):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id=cls.__name__,
+            display_name="Diffusion Model Info Loader",
+            category=cls.CATEGORY,
+            inputs=[
+                io.Combo.Input("diffusion_model", options=folder_paths.get_filename_list("diffusion_models")),
+                io.Int.Input("workflow_index", default=0, min=0, max=100)
+            ],
+            outputs=[
+                io.Image.Output(display_name="preview"),
+                io.String.Output(display_name="workflow_json"),
+                io.String.Output(display_name="metadata_json")
+            ]
+        )
+
+    @classmethod
+    def execute(cls, diffusion_model, workflow_index):
+        res = cls.get_info("diffusion_models", diffusion_model, workflow_index)
+        return io.NodeOutput(res[0], res[1], res[2])

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/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

pyproject.toml

@@ -1,9 +1,24 @@
 [project]
 name = "ComfyUI-ModelUtils"
 description = "[WIP]Custom nodes for handling, inspecting, modifying and creating various model files."
-version = "0.2.7"
+version = "0.4.4"
 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",
+]
+dependencies = [
+    "unifiedefficientloader==0.4.4"
+]
+
 [project.urls]
 Repository = "https://github.com/silveroxides/ComfyUI-ModelUtils"
 #  Used by Comfy Registry https://comfyregistry.org

requirements.txt

@@ -1 +1,5 @@
-# No external dependencies required
+requests
+Pillow
+mutagen
+unifiedefficientloader>=0.5.2
+av