Refactor loading strategies in mergers and LoRA resize nodes.

Feature/dora extract

__init__.py

@@ -24,6 +24,14 @@ 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,
@@ -39,6 +47,8 @@ from .nodes.model_info_nodes import (
 )
 
 
+from .nodes.lora_rename import AnimaLoraRename
+
 class ModelUtilsExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
@@ -61,6 +71,12 @@ 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,
@@ -68,6 +84,8 @@ class ModelUtilsExtension(ComfyExtension):
             LoRAMultiMerge, LoRAMultiMergeDARE, LoRAMultiMergeDAREEnhanced,
             # LoRA Merge To Model
             LoRAMergeToModel,
+            # LoRA Utilities
+            AnimaLoraRename,
             # Downloaders
             CheckpointInfoMetaDownloader, DiffusionModelInfoMetaDownloader, LoRAInfoMetaDownloader, EmbeddingInfoMetaDownloader,
             VAEInfoMetaDownloader, ControlNetInfoMetaDownloader, ManualPathInfoMetaDownloader,

docs/merger_2_model_modes.md

@@ -121,6 +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/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

@@ -377,6 +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.
@@ -391,6 +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
@@ -399,7 +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()
@@ -484,10 +488,11 @@ def resize_lora_file(
                     writer.write_dict(block_sd)
 
                     del result
-                    import gc
-                    gc.collect()
-                    if torch.cuda.is_available():
-                        torch.cuda.empty_cache()
+                    if force_clear_cache:
+                        import gc
+                        gc.collect()
+                        if torch.cuda.is_available():
+                            torch.cuda.empty_cache()
 
                     pbar.update(1)
         finally:
@@ -531,17 +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)
 
 
@@ -565,17 +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)
 
 
@@ -599,17 +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)
 
 
@@ -633,17 +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)
 
 
@@ -662,6 +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.
@@ -679,6 +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
@@ -694,8 +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)
-    lora_handlers = [MemoryEfficientSafeOpen(lp) for lp in lora_paths]
+    base_handler = MemoryEfficientSafeOpen(base_model_path, low_memory=lazy_load)
+    lora_handlers = [MemoryEfficientSafeOpen(lp, low_memory=lazy_load) for lp in lora_paths]
 
 
     try:
@@ -874,10 +903,11 @@ def merge_loras_to_model(
                         del cpu_base
                         stats["copied"] += 1
 
-                    import gc
-                    gc.collect()
-                    if torch.cuda.is_available():
-                        torch.cuda.empty_cache()
+                    if force_clear_cache:
+                        import gc
+                        gc.collect()
+                        if torch.cuda.is_available():
+                            torch.cuda.empty_cache()
 
                     pbar.update(1)
         finally:
@@ -957,6 +987,8 @@ class LoRAMergeToModel(io.ComfyNode):
                 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,
@@ -966,7 +998,7 @@ class LoRAMergeToModel(io.ComfyNode):
     def execute(cls, base_model, lora_count,
                 lora_1, weight_1, lora_2, weight_2, lora_3, weight_3, lora_4, weight_4,
                 lora_5, weight_5, lora_6, weight_6, lora_7, weight_7, lora_8, weight_8,
-                skip_patterns, output_filename, save_dtype, device) -> io.NodeOutput:
+                skip_patterns, output_filename, save_dtype, device, lazy_load, force_clear_cache) -> io.NodeOutput:
 
         # Build LoRA list based on count
         count = int(lora_count)
@@ -991,5 +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

@@ -132,6 +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)
@@ -181,6 +200,15 @@ class MergerLogic:
 
                 # Pre-load Model A's tensor with pinned memory for CUDA
                 cpu_tensor = primary_handler.get_tensor(key)
+
+                # Determine original dtypes across all active models for this key
+                original_dtypes = [primary_handler.get_dtype(key)]
+                for name, handler in handlers.items():
+                    if name != primary_model_name and key in handler.keys():
+                        original_dtypes.append(handler.get_dtype(key))
+
+                target_dtype = determine_target_dtype(original_dtypes, save_torch_dtype, override_dtype)
+
                 if process_device == 'cuda':
                     tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
                 else:
@@ -189,7 +217,7 @@ class MergerLogic:
 
                 # 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(save_torch_dtype).cpu()
+                    t = tensor_a.detach().to(target_dtype).cpu()
 
                     batch_buffer[key] = t
                     batch_bytes += t.numel() * t.element_size()
@@ -224,7 +252,7 @@ class MergerLogic:
 
                 if isinstance(result, dict):
                     for r_key, r_tensor in result.items():
-                        t = r_tensor.detach().to(save_torch_dtype).cpu()
+                        t = r_tensor.detach().to(target_dtype).cpu()
 
                         batch_buffer[r_key] = t
                         batch_bytes += t.numel() * t.element_size()
@@ -238,7 +266,7 @@ class MergerLogic:
                             slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(result.shape, target_shape))
                             result = result[slices]
 
-                    t = result.detach().to(save_torch_dtype).cpu()
+                    t = result.detach().to(target_dtype).cpu()
 
                     batch_buffer[key] = t
                     batch_bytes += t.numel() * t.element_size()
@@ -323,6 +351,7 @@ class CheckpointTwoMerger(io.ComfyNode):
                                          "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"),
@@ -335,7 +364,7 @@ class CheckpointTwoMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -344,7 +373,7 @@ class CheckpointTwoMerger(io.ComfyNode):
             "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, "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,
@@ -388,6 +417,7 @@ class ModelTwoMerger(io.ComfyNode):
                                          "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"),
@@ -400,7 +430,7 @@ class ModelTwoMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -409,7 +439,7 @@ class ModelTwoMerger(io.ComfyNode):
             "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, "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,
@@ -453,6 +483,7 @@ class TextEncoderTwoMerger(io.ComfyNode):
                                          "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"),
@@ -465,7 +496,7 @@ class TextEncoderTwoMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -474,7 +505,7 @@ class TextEncoderTwoMerger(io.ComfyNode):
             "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, "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,
@@ -518,6 +549,7 @@ class LoRATwoMerger(io.ComfyNode):
                                          "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"),
@@ -530,7 +562,7 @@ class LoRATwoMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -539,7 +571,7 @@ class LoRATwoMerger(io.ComfyNode):
             "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, "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,
@@ -583,6 +615,7 @@ class EmbeddingTwoMerger(io.ComfyNode):
                                          "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"),
@@ -595,7 +628,7 @@ class EmbeddingTwoMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -604,7 +637,7 @@ class EmbeddingTwoMerger(io.ComfyNode):
             "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, "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,
@@ -651,6 +684,7 @@ class CheckpointThreeMerger(io.ComfyNode):
                                          "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"),
@@ -663,7 +697,7 @@ class CheckpointThreeMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -672,7 +706,7 @@ class CheckpointThreeMerger(io.ComfyNode):
             "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, "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,
@@ -717,6 +751,7 @@ class ModelThreeMerger(io.ComfyNode):
                                          "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"),
@@ -729,7 +764,7 @@ class ModelThreeMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -738,7 +773,7 @@ class ModelThreeMerger(io.ComfyNode):
             "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, "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,
@@ -783,6 +818,7 @@ class TextEncoderThreeMerger(io.ComfyNode):
                                          "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"),
@@ -795,7 +831,7 @@ class TextEncoderThreeMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -804,7 +840,7 @@ class TextEncoderThreeMerger(io.ComfyNode):
             "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, "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,
@@ -849,6 +885,7 @@ class LoRAThreeMerger(io.ComfyNode):
                                          "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"),
@@ -861,7 +898,7 @@ class LoRAThreeMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -870,7 +907,7 @@ class LoRAThreeMerger(io.ComfyNode):
             "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, "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,
@@ -915,6 +952,7 @@ class EmbeddingThreeMerger(io.ComfyNode):
                                          "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"),
@@ -927,7 +965,7 @@ class EmbeddingThreeMerger(io.ComfyNode):
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 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,
-                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
+                glob_patterns: bool, lazy_load: bool, force_clear_cache: bool, override_dtype: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -936,7 +974,7 @@ class EmbeddingThreeMerger(io.ComfyNode):
             "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, "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,

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "ComfyUI-ModelUtils"
 description = "[WIP]Custom nodes for handling, inspecting, modifying and creating various model files."
-version = "0.4.1"
+version = "0.4.4"
 license = { file = "LICENSE" }
 
 classifiers = [

requirements.txt

@@ -1,5 +1,5 @@
 requests
 Pillow
 mutagen
-unifiedefficientloader>=0.5.1
+unifiedefficientloader>=0.5.2
 av