changes to LoRAs

- MemoryEfficientSafeOpen now supports low_memory=True which disables mmap
- Uses direct file reads with readinto() to avoid OS page caching
- _deserialize_tensor avoids bytearray copy when input is already bytearray
- merge_loras_to_model now uses low_memory=True for base model

This should reduce peak RAM from 2x model size to 1x + overhead

README.md

@@ -1,2 +1,55 @@
 # ComfyUI-ModelUtils
-[WIP]Custom nodes for handling, inspecting, modifying and creating various model files.
+
+A collection of ComfyUI custom nodes for inspecting, modifying, merging, and creating model files. Supports Models, TextEncoders, LoRAs, Checkpoints, and Embeddings.
+
+## Features
+
+- **MetaKeys** – Inspect and display metadata and tensor keys from model files
+- **RenameKeys** – Batch rename tensor keys using pattern matching
+- **PruneKeys** – Remove unwanted layers/keys from models
+- **Mergers** – Combine 2 or 3 models with configurable blend modes and ratios
+- **LoRA Extraction** – Extract LoRA adapters from model pairs using various SVD rank selection methods (Fixed, Ratio, Quantile, Knee-detection, Frobenius-norm)
+
+## Example Workflows
+
+<p align="center">
+  <img src="assets/GetMetaAndkeys.png" width="400" alt="Get Meta and Keys">
+  <br>
+  <a href="example_workflows/GetMetaAndkeys.json">📥 GetMetaAndkeys.json</a>
+</p>
+
+---
+
+<p align="center">
+  <img src="assets/LoRA_Extract_nodes.png" width="400" alt="LoRA Extraction Nodes">
+  <br>
+  <a href="example_workflows/LoRA_Extract_nodes.json">📥 LoRA_Extract_nodes.json</a>
+</p>
+
+---
+
+<p align="center">
+  <img src="assets/Merging_Examples.png" width="400" alt="Merging Examples">
+  <br>
+  <a href="example_workflows/Merging_Examples.json">📥 Merging_Examples.json</a>
+</p>
+
+---
+
+<p align="center">
+  <img src="assets/RenameKeysInModel.png" width="400" alt="Rename Keys in Model">
+  <br>
+  <a href="example_workflows/RenameKeysInModel.json">📥 RenameKeysInModel.json</a>
+</p>
+
+## Acknowledgements
+
+The LoRA extraction functionality was inspired by and references the excellent work from:
+
+- [kohya-ss/sd-scripts](https://github.com/kohya-ss/sd-scripts) – Training scripts for Stable Diffusion
+- [KohakuBlueleaf/LyCORIS](https://github.com/KohakuBlueleaf/LyCORIS) – Advanced LoRA techniques
+- [bmaltais/kohya_ss](https://github.com/bmaltais/kohya_ss) – Windows-friendly GUI for sd-scripts
+
+## License
+
+See [LICENSE](LICENSE) for details.

__init__.py

@@ -1,6 +1,7 @@
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
 
+from .nodes.lora_merger import LoRAMultiMerge, LoRAMultiMergeDARE
 from .nodes.metakeys import (
     ModelMetaKeys, TextEncoderMetaKeys, LoRAMetaKeys,
     CheckpointMetaKeys, EmbeddingMetaKeys
@@ -19,8 +20,14 @@ from .nodes.merger import (
     ModelThreeMerger, TextEncoderThreeMerger, LoRAThreeMerger,
     CheckpointThreeMerger, EmbeddingThreeMerger
 )
-from .nodes.lora_extract import (
-    LoRASaveFromFile, LoRACheckpointSaveFromFile
+from .nodes.lora_extract_svd import (
+    LoRAExtractFixed, LoRAExtractRatio, LoRAExtractQuantile,
+    LoRAExtractKnee, LoRAExtractFrobenius
+)
+from .nodes.lora_resize import (
+    LoRAResizeFixed, LoRAResizeRatio,
+    LoRAResizeFrobenius, LoRAResizeCumulative,
+    LoRAMergeToModel
 )
 
 
@@ -44,9 +51,18 @@ class ModelUtilsExtension(ComfyExtension):
             ModelThreeMerger, TextEncoderThreeMerger, LoRAThreeMerger,
             CheckpointThreeMerger, EmbeddingThreeMerger,
             # LoRA Extraction
-            LoRASaveFromFile, LoRACheckpointSaveFromFile,
+            LoRAExtractFixed, LoRAExtractRatio, LoRAExtractQuantile,
+            LoRAExtractKnee, LoRAExtractFrobenius,
+            # LoRA Resize
+            LoRAResizeFixed, LoRAResizeRatio,
+            LoRAResizeFrobenius, LoRAResizeCumulative,
+            # LoRA Multi-Merge
+            LoRAMultiMerge, LoRAMultiMergeDARE,
+            # LoRA Merge To Model
+            LoRAMergeToModel,
         ]
 
 
 async def comfy_entrypoint() -> ModelUtilsExtension:
-    return ModelUtilsExtension()
+    return ModelUtilsExtension()
+

docs/DEVELOPMENT.md

@@ -0,0 +1,321 @@
+# ComfyUI-ModelUtils - Complete Project History
+
+## Version Timeline
+
+| Version | Commits | Key Features |
+|---------|---------|--------------|
+| 0.1.1 | `70d6c13`-`d67cbca` | Initial nodes (MetaKeys) |
+| 0.1.2 | `6b662a4` | Icon assets |
+| 0.1.4 | `cbe3502` | Key renaming/pruning nodes |
+| 0.1.5 | `78b3198` | Model merging nodes |
+| 0.1.6 | `31989be` | Embedding nodes, namespace fixes |
+| 0.1.7 | `b6fae8d` | Mismatch handling, layer filtering, V3 migration |
+| **0.2.0** | `3df073f` | LoRA extraction, memory optimization, pinned memory |
+
+---
+
+## Phase 1: Initial Release (v0.1.1)
+
+**Commit:** `70d6c13` - "Add first nodes, toml and publish"
+
+### Files Created
+- `nodes/metakeys.py` - Get model metadata and layer keys
+- `nodes/utils.py` - Shared utilities
+- `__init__.py` - Node registration
+- `pyproject.toml` - Package configuration
+
+### Nodes
+- `ModelMetaKeys` - Get diffusion model metadata
+- `CheckpointMetaKeys` - Get checkpoint metadata
+- `TextEncoderMetaKeys` - Get text encoder metadata
+- `LoRAMetaKeys` - Get LoRA metadata
+- `EmbeddingMetaKeys` - Get embedding metadata
+
+---
+
+## Phase 2: Key Manipulation (v0.1.4)
+
+**Commit:** `4f08fe6` - "add key renaming and key pruning nodes"
+
+### Files Created
+- `nodes/renamekeys.py` - Rename tensor keys
+- `nodes/prunekeys.py` - Remove tensor keys by pattern
+
+### Nodes (5 variants each for model types)
+- `ModelRenameKeys`, `CheckpointRenameKeys`, `TextEncoderRenameKeys`, `LoRARenameKeys`, `EmbeddingRenameKeys`
+- `ModelPruneKeys`, `CheckpointPruneKeys`, `TextEncoderPruneKeys`, `LoRAPruneKeys`, `EmbeddingPruneKeys`
+
+---
+
+## Phase 3: Model Merging (v0.1.5)
+
+**Commit:** `7959370` - "Add model merging nodes"
+
+### Files Created
+- `nodes/merger.py` - Main merging logic
+- `nodes/merger_ops.py` - Merge operation modes
+- `nodes/merger_utils.py` - `MemoryEfficientSafeOpen` class
+- `docs/merger_2_model_modes.md` - Documentation
+- `docs/merger_3_model_modes.md` - Documentation
+
+### Nodes
+- `CheckpointTwoMerger`, `DiffusionModelTwoMerger`, `TextEncoderTwoMerger`, `LoRATwoMerger`
+- `CheckpointThreeMerger`, `DiffusionModelThreeMerger`, `TextEncoderThreeMerger`, `LoRAThreeMerger`
+
+### Merge Modes (TWO_MODEL_MODES)
+- Weighted Sum, Add Difference, Smooth Add Difference
+- Train Difference, Sum Twice, Triple Sum
+- Euclidean Add Difference, Distribution Crossover, Ties
+- Slerp, Lerp, SVD Supermerger, SVD LoRA Extraction
+
+### Merge Modes (THREE_MODEL_MODES)
+- Add Difference (A + (B - C) × α)
+
+### MemoryEfficientSafeOpen (initial)
+```python
+class MemoryEfficientSafeOpen:
+    def __init__(self, filename, device='cpu', mmap_mode=True)
+    def keys() -> List[str]
+    def get_tensor(key) -> Tensor
+    # mmap for zero-copy reads
+```
+
+---
+
+## Phase 4: Embedding Nodes (v0.1.6)
+
+**Commit:** `0d035ef` - "Add embedding nodes"
+
+### Additions
+- Added embedding variants to all node types
+- Namespace conflict fixes (`68f22fa`, `5ea63ab`)
+- Refactored names for universality (`232dc76`)
+
+---
+
+## Phase 5: Mismatch Handling (v0.1.7)
+
+**Commits:** `19bc047` → `de27186`
+
+### Features Added
+- `mismatch_mode` parameter: skip/zeros/error
+- Layer filtering with regex patterns
+- `exclude_patterns` - Keep model A only (no merge)
+- `discard_patterns` - Remove from output entirely
+
+### Files Modified
+- `merger.py` - Added mismatch handling logic
+- `merger_ops.py` - Added `MissingTensorBehavior` enum, `MissingTensorError`
+
+---
+
+## Phase 6: V3 API Migration (v0.1.7)
+
+**Commit:** `a3116cb` - "migrate to v3 schema and add LoRA extraction node"
+
+### Changes
+- Migrated all nodes from legacy API to `comfy_api.latest.io`
+- `io.Schema`, `io.ComfyNode`, `io.NodeOutput`
+- Added initial LoRA extraction node (diff-based)
+
+---
+
+## Phase 7: LyCORIS Extraction (pre-0.2.0)
+
+**Commits:** `92d3c55` → `18679cc`
+
+### Files Created
+- `nodes/lycoris_extract.py` - LyCORIS-based extraction
+- `nodes/lora_extract.py` - Simple diff extraction
+
+### Initial Implementation
+- Used `lycoris-lora` external dependency
+- `lycoris.utils.extract_linear`, `extract_conv`
+- Added `requirements.txt` with `lycoris-lora`
+
+### Nodes Split (`18679cc`)
+- `LyCORISExtractFixed`, `LyCORISExtractThreshold`
+- `LyCORISExtractRatio`, `LyCORISExtractQuantile`
+
+---
+
+## Phase 8: Native SVD Port (0.2.0)
+
+**Commits:** `6d7b2d1` → `eef7377`
+
+### Removed
+- `lycoris-lora` dependency
+- `lycoris_extract.py`
+- `lora_extract.py`
+
+### NEW: [lora_extract_svd.py](file:///f:/ComfyUI/custom_nodes/ComfyUI-ModelUtils/nodes/lora_extract_svd.py)
+
+**8 Rank Selection Functions:**
+```python
+_index_sv_fixed(S, dim)           # Fixed rank
+_index_sv_threshold(S, threshold) # Absolute threshold
+_index_sv_ratio(S, ratio)         # Relative to max
+_index_sv_cumulative(S, target)   # Cumulative sum %
+_index_sv_fro(S, target)          # Frobenius norm
+_index_sv_knee(S)                 # Knee detection
+_index_sv_cumulative_knee(S)      # Cumsum knee
+_index_sv_rel_decrease(S, tau)    # Ratio drop
+```
+
+**SVD Functions:**
+```python
+_svd_extract_linear_lowrank(weight, rank, device, clamp_quantile)
+_svd_extract_linear(weight, mode, mode_param, device, max_rank, clamp_quantile)
+_svd_extract_conv(weight, mode, mode_param, device, max_rank, clamp_quantile)
+_detect_fused_layer(key, shape)  # QKV/MLP detection
+_extract_chunked_layer(...)      # Chunked extraction
+```
+
+**Nodes:**
+- `LoRAExtractFixed` - Uses `svd_lowrank` (10x faster)
+- `LoRAExtractRatio` - Relative threshold
+- `LoRAExtractQuantile` - Cumulative target
+- `LoRAExtractKnee` - Auto knee detection
+- `LoRAExtractFrobenius` - Norm preservation
+
+---
+
+## Phase 9: Memory Management (0.2.0)
+
+**Commit:** `6d7b2d1` - "refactor nodes and implement memory management"
+
+### NEW: [device_utils.py](file:///f:/ComfyUI/custom_nodes/ComfyUI-ModelUtils/nodes/device_utils.py)
+
+Integrates with `comfy.model_management`:
+
+```python
+@dataclass
+class DeviceCapabilities:
+    device_type, total_vram_gb, free_vram_gb, total_ram_gb, free_ram_gb
+    cpu_cores, recommended_workers, max_tensor_size_gb
+    supports_pinned_memory, supports_async_streams
+
+get_device_capabilities(device) -> DeviceCapabilities
+estimate_model_size(path) -> float  # GB
+can_fit_in_memory(paths, device, safety_factor) -> (bool, str)
+prepare_for_large_operation(gb, device) -> bool  # Calls free_memory()
+cleanup_after_operation()  # gc.collect() + soft_empty_cache()
+get_optimal_workers(size_gb) -> int
+```
+
+### Updates to Existing Nodes
+- `metakeys.py` - Header-only reading (no tensor loading)
+- `renamekeys.py` - `MemoryEfficientSafeOpen` + memory mgmt
+- `prunekeys.py` - `MemoryEfficientSafeOpen` + memory mgmt
+- `merger.py` - Memory prep/cleanup
+
+---
+
+## Phase 10: Pinned Memory (0.2.0)
+
+**Commit:** `b770786` - "Add pinned memory for faster merging and extraction"
+
+### [merger_utils.py](file:///f:/ComfyUI/custom_nodes/ComfyUI-ModelUtils/nodes/merger_utils.py) Enhancements
+
+```python
+# Context manager
+class PinnedTensor:
+    def __enter__(self) -> Tensor  # Pins memory
+    def __exit__(...)              # Unpins
+
+# Transfer helper (2-3x faster)
+def transfer_to_gpu_pinned(tensor, device='cuda', dtype=None) -> Tensor
+
+# MemoryEfficientSafeOpen additions
+def get_tensor_to_gpu(key, device, dtype)  # Pinned transfer
+def keys_sorted_by_offset() -> List[str]   # Sequential I/O
+def get_tensors_parallel(keys, workers, sort_by_offset) -> Dict  # Parallel
+def get_tensor_as_dict(key) -> Dict        # JSON config parsing
+```
+
+---
+
+## Phase 11: SVD Optimization (0.2.0)
+
+**Commit:** `65c1ae3` - "add svd_lowrank for fixed rank extraction"
+
+### Changes
+- Added `_svd_extract_linear_lowrank()` using `torch.svd_lowrank`
+- Routes fixed mode to lowrank for **10x speedup**
+- Changed `chunk_large_layers` default to `False`
+
+---
+
+## Phase 12: Bug Fixes (0.2.0)
+
+**Commit:** `eef7377` - "Added .contiguous() to saves"
+
+### Fix
+Non-contiguous tensor error from SVD slices:
+```python
+tensor.to(dtype).cpu().contiguous()  # Added .contiguous()
+```
+
+---
+
+## Current File Structure
+
+```
+ComfyUI-ModelUtils/
+├── __init__.py
+├── pyproject.toml
+├── requirements.txt (empty - no external deps)
+├── README.md
+├── docs/
+│   ├── merger_2_model_modes.md
+│   └── merger_3_model_modes.md
+├── nodes/
+│   ├── utils.py
+│   ├── metakeys.py
+│   ├── renamekeys.py
+│   ├── prunekeys.py
+│   ├── merger.py
+│   ├── merger_ops.py
+│   ├── merger_utils.py
+│   ├── device_utils.py
+│   └── lora_extract_svd.py
+└── assets/
+    └── icon.png
+```
+
+---
+
+## Performance Summary
+
+| Feature | Improvement |
+|---------|-------------|
+| Fixed-rank SVD | **10x** (svd_lowrank) |
+| CPU→GPU transfer | **2-3x** (pinned memory) |
+| Metadata reading | **100x+** (header only) |
+| Parallel I/O | **2-4x** (ThreadPoolExecutor) |
+| Memory management | Auto-unload models before heavy ops |
+| **LoRA Resize Fast** | **~91x** (merge-to-base extraction) |
+
+---
+
+## Phase 13: Fast LoRA Resize (2026-01-11)
+
+### Summary
+Added fast LoRA resize via base model extraction path, achieving ~91x speedup.
+
+### Changes
+- Added `resize_lora_via_base()` function using extraction code path
+- Added 4 new fast resize nodes requiring base model input
+
+### Files Modified
+
+| File | Changes |
+|------|---------|
+| `nodes/lora_resize.py` | Added `resize_lora_via_base()`, 4 new node classes |
+| `__init__.py` | Registered new nodes |
+
+### New Nodes
+- `LoRAResizeViaBaseFixed` - Fast fixed rank resize
+- `LoRAResizeViaBaseRatio` - Fast SV ratio resize  
+- `LoRAResizeViaBaseFrobenius` - Fast Frobenius resize
+- `LoRAResizeViaBaseCumulative` - Fast cumulative resize

docs/merger_2_model_modes.md

@@ -23,12 +23,47 @@
 ---
 
 ## Power-Up (DARE)
-> Adds the unique capabilities of Model B to Model A using the Drop and Rescale (DARE) technique, which often preserves the knowledge of the base model better than simple additions.
+> Adds the unique capabilities of Model B to Model A using the Drop and Rescale (DARE) technique. This implementation handles shape mismatches between models by padding and uses a randomized dropout mask.
 
 **Models Used:** A, B
 **Parameters:**
-- **Alpha:** The dropout rate. This is the proportion of unique weights from Model B that are *dropped* before merging. A higher value means less of B is merged.
+- **Alpha:** The dropout rate ($p$). This is the proportion of delta parameters from Model B that are randomly set to zero.
 - **Beta:** A final multiplier for the rescaled difference before it's added to Model A.
+- **Rescaling Logic:** Remaining weights are automatically rescaled by $1/(1-p)$ as per the DARE paper to approximate the original embeddings.
+
+---
+
+## Enhanced Man Interp
+> Sophisticated interpolation between values from A and B depending on their difference relative to other values, with manual threshold control.
+
+**Models Used:** A, B
+**Parameters:**
+- **Alpha:** Interpolation strength.
+- **Beta:** Lower mean threshold for filtering differences.
+- **Gamma:** Upper mean threshold for filtering differences.
+- **Delta:** Smoothness factor (mix between randomized mask and powered differences).
+
+---
+
+## Enhanced Auto Interp
+> Automated version of the enhanced interpolation mode that dynamically calculates thresholds based on mean differences.
+
+**Models Used:** A, B
+**Parameters:**
+- **Alpha:** Interpolation strength.
+- **Beta:** Threshold adjustment factor.
+- **Gamma:** Smoothness factor.
+
+---
+
+## Weight-Sum Cutoff
+> A linear interpolation mode that only applies the merge to weights whose differences fall within a specific threshold range.
+
+**Models Used:** A, B
+**Parameters:**
+- **Alpha:** Interpolation weight (multiplier for the difference).
+- **Beta:** Upper threshold for the difference cutoff.
+- **Gamma:** Lower threshold for the difference cutoff.
 
 ---
 
@@ -84,4 +119,4 @@ Layers matching any pattern will be **removed entirely** from the output.
 | `text_model` | All text encoder layers |
 | `\.norm` | All normalization layers |
 | `attn\.(q\|k\|v)` | Query, key, value attention weights |
-| `block\.[0-9]\.` | Blocks 0-9 |
+| `block\.[0-9]\.` | Blocks 0-9 |

docs/merger_3_model_modes.md

@@ -22,19 +22,19 @@
 ---
 
 ## Extract-Features
-> A powerful mode that identifies features present in both `(B - A)` and `(C - A)` and adds them to A. Allows for fine-grained control over combining aspects based on their similarity.
+> A powerful mode that identifies features present in both `(B - A)` and `(C - A)` and adds them to A. It uses per-vector cosine similarity to decide how much of each feature to keep, allowing for fine-grained control over combining aspects.
 
 **Models Used:** A, B, C
 **Parameters:**
 - **Alpha:** Weights the merge between Model B (`0.0`) and Model C (`1.0`).
 - **Beta:** Controls the focus on similarity (`0.0`) versus dissimilarity (`1.0`).
-- **Gamma:** A bias exponent for similarity. Higher values increase the bias.
+- **Gamma:** A bias exponent for similarity calculation.
 - **Delta:** A final multiplier for the extracted features before they are added to Model A.
 
 ---
 
 ## Add-Dissimilarities
-> Identifies features that are dissimilar between Model B and Model C and adds them to Model A. Useful for combining unique aspects of two different models.
+> Identifies features that are dissimilar between Model B and Model C (relative to A) and adds them to Model A. Useful for combining unique aspects of two different models.
 
 **Models Used:** A, B, C
 **Parameters:**
@@ -80,4 +80,4 @@ Layers matching any pattern will be **removed entirely** from the output.
 **Pattern format:**
 - Whitespace-separated regex patterns
 - Patterns use **substring matching** (not full match)
-- Example: `text_model lora` matches any key containing "text_model" OR "lora"
+- Example: `text_model lora` matches any key containing "text_model" OR "lora"

example_workflows/GetMetaAndkeys.json

@@ -0,0 +1,600 @@
+{
+  "id": "e07ecc9b-30b9-45e4-9ed6-5fbb7e510257",
+  "revision": 0,
+  "last_node_id": 31,
+  "last_link_id": 19,
+  "nodes": [
+    {
+      "id": 16,
+      "type": "PreviewAny",
+      "pos": [
+        -3608,
+        200
+      ],
+      "size": [
+        832,
+        232
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 18
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 17,
+      "type": "ModelMetaKeys",
+      "pos": [
+        -2800,
+        504
+      ],
+      "size": [
+        346.253515625,
+        78
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "metadata",
+          "type": "STRING",
+          "links": [
+            13
+          ]
+        },
+        {
+          "name": "keys",
+          "type": "STRING",
+          "links": [
+            17
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "ModelMetaKeys"
+      },
+      "widgets_values": [
+        "flux2-dev.safetensors"
+      ]
+    },
+    {
+      "id": 18,
+      "type": "TextEncoderMetaKeys",
+      "pos": [
+        -3176,
+        680
+      ],
+      "size": [
+        329.4234375,
+        78
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "metadata",
+          "type": "STRING",
+          "links": [
+            15
+          ]
+        },
+        {
+          "name": "keys",
+          "type": "STRING",
+          "links": [
+            16
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "TextEncoderMetaKeys"
+      },
+      "widgets_values": [
+        "clip_l.safetensors"
+      ]
+    },
+    {
+      "id": 19,
+      "type": "CheckpointMetaKeys",
+      "pos": [
+        -3168,
+        512
+      ],
+      "size": [
+        317.6724609375,
+        78
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "metadata",
+          "type": "STRING",
+          "links": [
+            12
+          ]
+        },
+        {
+          "name": "keys",
+          "type": "STRING",
+          "links": [
+            18
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "CheckpointMetaKeys"
+      },
+      "widgets_values": [
+        "unsorted\\v1-5-pruned-emaonly.safetensors"
+      ]
+    },
+    {
+      "id": 21,
+      "type": "LoRAMetaKeys",
+      "pos": [
+        -2800,
+        680
+      ],
+      "size": [
+        277.0943359375,
+        78
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "metadata",
+          "type": "STRING",
+          "links": [
+            14
+          ]
+        },
+        {
+          "name": "keys",
+          "type": "STRING",
+          "links": [
+            19
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRAMetaKeys"
+      },
+      "widgets_values": [
+        "Chroma\\FlatAnimeStyle_Chroma.safetensors"
+      ]
+    },
+    {
+      "id": 24,
+      "type": "PreviewAny",
+      "pos": [
+        -2728,
+        216
+      ],
+      "size": [
+        512,
+        200
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 17
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 25,
+      "type": "PreviewAny",
+      "pos": [
+        -2816,
+        832
+      ],
+      "size": [
+        736,
+        248
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 19
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 27,
+      "type": "PreviewAny",
+      "pos": [
+        -3496,
+        848
+      ],
+      "size": [
+        632,
+        224
+      ],
+      "flags": {},
+      "order": 11,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 16
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 28,
+      "type": "PreviewAny",
+      "pos": [
+        -3600,
+        664
+      ],
+      "size": [
+        352,
+        136
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 15
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 29,
+      "type": "PreviewAny",
+      "pos": [
+        -2408,
+        648
+      ],
+      "size": [
+        360,
+        120
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 14
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 30,
+      "type": "PreviewAny",
+      "pos": [
+        -3576,
+        480
+      ],
+      "size": [
+        336,
+        128
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 12
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 31,
+      "type": "PreviewAny",
+      "pos": [
+        -2368,
+        472
+      ],
+      "size": [
+        328,
+        120
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 13
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    }
+  ],
+  "links": [
+    [
+      12,
+      19,
+      0,
+      30,
+      0,
+      "STRING"
+    ],
+    [
+      13,
+      17,
+      0,
+      31,
+      0,
+      "STRING"
+    ],
+    [
+      14,
+      21,
+      0,
+      29,
+      0,
+      "STRING"
+    ],
+    [
+      15,
+      18,
+      0,
+      28,
+      0,
+      "STRING"
+    ],
+    [
+      16,
+      18,
+      1,
+      27,
+      0,
+      "STRING"
+    ],
+    [
+      17,
+      17,
+      1,
+      24,
+      0,
+      "STRING"
+    ],
+    [
+      18,
+      19,
+      1,
+      16,
+      0,
+      "STRING"
+    ],
+    [
+      19,
+      21,
+      1,
+      25,
+      0,
+      "STRING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "workflowRendererVersion": "LG",
+    "ds": {
+      "scale": 0.9070294784580497,
+      "offset": [
+        3855.323424397858,
+        -56.106573378634465
+      ]
+    },
+    "frontendVersion": "1.37.2",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true
+  },
+  "version": 0.4
+}

example_workflows/LoRA_Extract_nodes.json

@@ -0,0 +1,522 @@
+{
+  "id": "e07ecc9b-30b9-45e4-9ed6-5fbb7e510257",
+  "revision": 0,
+  "last_node_id": 10,
+  "last_link_id": 5,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "LoRAExtractFixed",
+      "pos": [
+        -1848,
+        16
+      ],
+      "size": [
+        296,
+        352
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_path",
+          "type": "STRING",
+          "links": [
+            3
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRAExtractFixed"
+      },
+      "widgets_values": [
+        "13499-bf16.safetensors",
+        "13499-bf16.safetensors",
+        64,
+        32,
+        false,
+        0.99,
+        0,
+        "skip",
+        "extracted_lora",
+        "fp16",
+        "cuda",
+        ""
+      ]
+    },
+    {
+      "id": 2,
+      "type": "LoRAExtractFrobenius",
+      "pos": [
+        -2432,
+        24
+      ],
+      "size": [
+        296,
+        352
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_path",
+          "type": "STRING",
+          "links": [
+            1
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRAExtractFrobenius"
+      },
+      "widgets_values": [
+        "13499-bf16.safetensors",
+        "13499-bf16.safetensors",
+        0.9,
+        0.9,
+        false,
+        0.99,
+        0,
+        "skip",
+        "extracted_lora",
+        "fp16",
+        "cuda",
+        ""
+      ]
+    },
+    {
+      "id": 3,
+      "type": "LoRAExtractKnee",
+      "pos": [
+        -2416,
+        440
+      ],
+      "size": [
+        280,
+        328
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_path",
+          "type": "STRING",
+          "links": [
+            2
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRAExtractKnee"
+      },
+      "widgets_values": [
+        "13499-bf16.safetensors",
+        "13499-bf16.safetensors",
+        "sv_knee",
+        false,
+        0.99,
+        0,
+        "skip",
+        "extracted_lora",
+        "fp16",
+        "cuda",
+        ""
+      ]
+    },
+    {
+      "id": 4,
+      "type": "LoRAExtractQuantile",
+      "pos": [
+        -1248,
+        128
+      ],
+      "size": [
+        288,
+        352
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_path",
+          "type": "STRING",
+          "links": [
+            5
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRAExtractQuantile"
+      },
+      "widgets_values": [
+        "13499-bf16.safetensors",
+        "13499-bf16.safetensors",
+        0.9,
+        0.9,
+        false,
+        0.99,
+        0,
+        "skip",
+        "extracted_lora",
+        "fp16",
+        "cuda",
+        ""
+      ]
+    },
+    {
+      "id": 5,
+      "type": "LoRAExtractRatio",
+      "pos": [
+        -1824,
+        424
+      ],
+      "size": [
+        296,
+        352
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_path",
+          "type": "STRING",
+          "links": [
+            4
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRAExtractRatio"
+      },
+      "widgets_values": [
+        "13499-bf16.safetensors",
+        "13499-bf16.safetensors",
+        0.5,
+        0.5,
+        false,
+        0.99,
+        0,
+        "skip",
+        "extracted_lora",
+        "fp16",
+        "cuda",
+        ""
+      ]
+    },
+    {
+      "id": 6,
+      "type": "PreviewAny",
+      "pos": [
+        -2096,
+        136
+      ],
+      "size": [
+        210,
+        168
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 1
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 7,
+      "type": "PreviewAny",
+      "pos": [
+        -2096,
+        448
+      ],
+      "size": [
+        210,
+        168
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 2
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 8,
+      "type": "PreviewAny",
+      "pos": [
+        -1504,
+        96
+      ],
+      "size": [
+        210,
+        168
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 3
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 9,
+      "type": "PreviewAny",
+      "pos": [
+        -1472,
+        432
+      ],
+      "size": [
+        210,
+        168
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 4
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 10,
+      "type": "PreviewAny",
+      "pos": [
+        -904,
+        184
+      ],
+      "size": [
+        210,
+        168
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 5
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    }
+  ],
+  "links": [
+    [
+      1,
+      2,
+      0,
+      6,
+      0,
+      "STRING"
+    ],
+    [
+      2,
+      3,
+      0,
+      7,
+      0,
+      "STRING"
+    ],
+    [
+      3,
+      1,
+      0,
+      8,
+      0,
+      "STRING"
+    ],
+    [
+      4,
+      5,
+      0,
+      9,
+      0,
+      "STRING"
+    ],
+    [
+      5,
+      4,
+      0,
+      10,
+      0,
+      "STRING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "workflowRendererVersion": "LG",
+    "ds": {
+      "scale": 1.05,
+      "offset": [
+        2532,
+        84
+      ]
+    },
+    "frontendVersion": "1.37.2",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true
+  },
+  "version": 0.4
+}

example_workflows/Merging_Examples.json

@@ -0,0 +1,460 @@
+{
+  "id": "e07ecc9b-30b9-45e4-9ed6-5fbb7e510257",
+  "revision": 0,
+  "last_node_id": 16,
+  "last_link_id": 11,
+  "nodes": [
+    {
+      "id": 6,
+      "type": "PreviewAny",
+      "pos": [
+        -3552,
+        136
+      ],
+      "size": [
+        336,
+        160
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 6
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 7,
+      "type": "PreviewAny",
+      "pos": [
+        -2792,
+        144
+      ],
+      "size": [
+        456,
+        464
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 9
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        false
+      ]
+    },
+    {
+      "id": 11,
+      "type": "ModelThreeMerger",
+      "pos": [
+        -3120,
+        112
+      ],
+      "size": [
+        296,
+        522
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_filename",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "documentation",
+          "type": "STRING",
+          "links": [
+            9
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "ModelThreeMerger"
+      },
+      "widgets_values": [
+        "DOCUMENTATION ONLY",
+        "Chroma1-HD.safetensors",
+        "Chroma-DC-2K.safetensors",
+        "None",
+        "Add-Difference",
+        "skip",
+        0.5,
+        0.5,
+        0.5,
+        0.5,
+        996811747788113,
+        "randomize",
+        "merged_3_model",
+        "fp32",
+        "cuda",
+        "",
+        ""
+      ]
+    },
+    {
+      "id": 12,
+      "type": "ModelTwoMerger",
+      "pos": [
+        -3120,
+        704
+      ],
+      "size": [
+        304,
+        474
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_filename",
+          "type": "STRING",
+          "links": [
+            11
+          ]
+        },
+        {
+          "name": "documentation",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "ModelTwoMerger"
+      },
+      "widgets_values": [
+        "MERGE",
+        "None",
+        "None",
+        "Weight-Sum",
+        "skip",
+        0.5,
+        0.5,
+        0.99,
+        388546580517523,
+        "randomize",
+        "merged_2_model",
+        "fp32",
+        "cuda",
+        "",
+        ""
+      ]
+    },
+    {
+      "id": 13,
+      "type": "LoRAThreeMerger",
+      "pos": [
+        -3872,
+        128
+      ],
+      "size": [
+        296,
+        522
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_filename",
+          "type": "STRING",
+          "links": [
+            6
+          ]
+        },
+        {
+          "name": "documentation",
+          "type": "STRING",
+          "links": []
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRAThreeMerger"
+      },
+      "widgets_values": [
+        "MERGE",
+        "None",
+        "None",
+        "None",
+        "Add-Difference",
+        "skip",
+        0.5,
+        0.5,
+        0.5,
+        0.5,
+        1120102695669269,
+        "randomize",
+        "merged_3_lora",
+        "fp32",
+        "cuda",
+        "",
+        ""
+      ]
+    },
+    {
+      "id": 14,
+      "type": "LoRATwoMerger",
+      "pos": [
+        -3872,
+        720
+      ],
+      "size": [
+        296,
+        474
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "output_filename",
+          "type": "STRING",
+          "links": []
+        },
+        {
+          "name": "documentation",
+          "type": "STRING",
+          "links": [
+            10
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRATwoMerger"
+      },
+      "widgets_values": [
+        "DOCUMENTATION ONLY",
+        "Hyper-FLUX.1-dev-16steps-lora-rescale_z1.safetensors",
+        "Hyper-FLUX.1-dev-16steps-lora-rescale_z2.safetensors",
+        "Weight-Sum",
+        "skip",
+        0.5,
+        0.5,
+        0.99,
+        1020143314396572,
+        "randomize",
+        "merged_2_lora",
+        "fp32",
+        "cuda",
+        "",
+        ""
+      ]
+    },
+    {
+      "id": 15,
+      "type": "PreviewAny",
+      "pos": [
+        -2768,
+        720
+      ],
+      "size": [
+        432,
+        160
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 11
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 16,
+      "type": "PreviewAny",
+      "pos": [
+        -3544,
+        712
+      ],
+      "size": [
+        400,
+        448
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 10
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        false
+      ]
+    }
+  ],
+  "links": [
+    [
+      6,
+      13,
+      0,
+      6,
+      0,
+      "STRING"
+    ],
+    [
+      9,
+      11,
+      1,
+      7,
+      0,
+      "STRING"
+    ],
+    [
+      10,
+      14,
+      1,
+      16,
+      0,
+      "STRING"
+    ],
+    [
+      11,
+      12,
+      0,
+      15,
+      0,
+      "STRING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "workflowRendererVersion": "LG",
+    "ds": {
+      "scale": 0.9070294784580494,
+      "offset": [
+        3972,
+        -12
+      ]
+    },
+    "frontendVersion": "1.37.2",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true
+  },
+  "version": 0.4
+}

example_workflows/RenameKeysInModel.json

@@ -0,0 +1,583 @@
+{
+  "id": "e07ecc9b-30b9-45e4-9ed6-5fbb7e510257",
+  "revision": 0,
+  "last_node_id": 46,
+  "last_link_id": 31,
+  "nodes": [
+    {
+      "id": 35,
+      "type": "LoRAMetaKeys",
+      "pos": [
+        -3136,
+        512
+      ],
+      "size": [
+        376,
+        78
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "metadata",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "keys",
+          "type": "STRING",
+          "links": [
+            21,
+            22
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRAMetaKeys"
+      },
+      "widgets_values": [
+        "qwen\\gmqwen_k8_piid_4step.safetensors"
+      ]
+    },
+    {
+      "id": 36,
+      "type": "LoRARenameKeys",
+      "pos": [
+        -2856,
+        864
+      ],
+      "size": [
+        320,
+        190
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "old_keys",
+          "type": "STRING",
+          "widget": {
+            "name": "old_keys"
+          },
+          "link": 27
+        },
+        {
+          "name": "new_keys",
+          "type": "STRING",
+          "widget": {
+            "name": "new_keys"
+          },
+          "link": 28
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output_path",
+          "type": "STRING",
+          "links": [
+            31
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfyui_modelutils",
+        "ver": "eef737706127fe4a2c28172ab706e92318e0dc60",
+        "Node name for S&R": "LoRARenameKeys"
+      },
+      "widgets_values": [
+        "qwen\\gmqwen_k8_piid_4step.safetensors",
+        "",
+        "",
+        "gmqwen_k8_piid_4step"
+      ]
+    },
+    {
+      "id": 37,
+      "type": "PreviewAny",
+      "pos": [
+        -3792,
+        920
+      ],
+      "size": [
+        592,
+        384
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 21
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 40,
+      "type": "RegexReplace",
+      "pos": [
+        -3400,
+        568
+      ],
+      "size": [
+        210,
+        292
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "string",
+          "type": "STRING",
+          "widget": {
+            "name": "string"
+          },
+          "link": 22
+        }
+      ],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": [
+            23,
+            25
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "RegexReplace"
+      },
+      "widgets_values": [
+        "",
+        ",\\s\\[.+$",
+        "",
+        true,
+        true,
+        false,
+        0
+      ]
+    },
+    {
+      "id": 41,
+      "type": "PreviewAny",
+      "pos": [
+        -3784,
+        528
+      ],
+      "size": [
+        360,
+        328
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 23
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 42,
+      "type": "RegexReplace",
+      "pos": [
+        -3176,
+        1016
+      ],
+      "size": [
+        288,
+        292
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "string",
+          "type": "STRING",
+          "widget": {
+            "name": "string"
+          },
+          "link": 26
+        }
+      ],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": [
+            28,
+            30
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "RegexReplace"
+      },
+      "widgets_values": [
+        "",
+        "^(\\w+)",
+        "diffusion_model.\\1",
+        true,
+        true,
+        false,
+        0
+      ]
+    },
+    {
+      "id": 43,
+      "type": "RegexReplace",
+      "pos": [
+        -3152,
+        648
+      ],
+      "size": [
+        288,
+        292
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "string",
+          "type": "STRING",
+          "widget": {
+            "name": "string"
+          },
+          "link": 25
+        }
+      ],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": [
+            26,
+            27,
+            29
+          ]
+        }
+      ],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "RegexReplace"
+      },
+      "widgets_values": [
+        "",
+        "^(proj_out.+|norm_out.+)$\\n",
+        "",
+        true,
+        true,
+        false,
+        0
+      ]
+    },
+    {
+      "id": 44,
+      "type": "PreviewAny",
+      "pos": [
+        -2832,
+        632
+      ],
+      "size": [
+        640,
+        192
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 29
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 45,
+      "type": "PreviewAny",
+      "pos": [
+        -2848,
+        1096
+      ],
+      "size": [
+        640,
+        192
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 30
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    },
+    {
+      "id": 46,
+      "type": "PreviewAny",
+      "pos": [
+        -2520,
+        872
+      ],
+      "size": [
+        336,
+        168
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 31
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65,
+        "cnr_id": "comfy-core",
+        "ver": "0.7.0",
+        "Node name for S&R": "PreviewAny"
+      },
+      "widgets_values": [
+        null,
+        null,
+        true
+      ]
+    }
+  ],
+  "links": [
+    [
+      21,
+      35,
+      1,
+      37,
+      0,
+      "STRING"
+    ],
+    [
+      22,
+      35,
+      1,
+      40,
+      0,
+      "STRING"
+    ],
+    [
+      23,
+      40,
+      0,
+      41,
+      0,
+      "STRING"
+    ],
+    [
+      25,
+      40,
+      0,
+      43,
+      0,
+      "STRING"
+    ],
+    [
+      26,
+      43,
+      0,
+      42,
+      0,
+      "STRING"
+    ],
+    [
+      27,
+      43,
+      0,
+      36,
+      0,
+      "STRING"
+    ],
+    [
+      28,
+      42,
+      0,
+      36,
+      1,
+      "STRING"
+    ],
+    [
+      29,
+      43,
+      0,
+      44,
+      0,
+      "STRING"
+    ],
+    [
+      30,
+      42,
+      0,
+      45,
+      0,
+      "STRING"
+    ],
+    [
+      31,
+      36,
+      0,
+      46,
+      0,
+      "STRING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "workflowRendererVersion": "LG",
+    "ds": {
+      "scale": 1,
+      "offset": [
+        4191.258750150775,
+        -356.1917593715161
+      ]
+    },
+    "frontendVersion": "1.37.2",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true
+  },
+  "version": 0.4
+}

nodes/device_utils.py

@@ -0,0 +1,268 @@
+"""
+Device capabilities and memory management utilities.
+
+Integrates with ComfyUI's model_management for safe memory operations.
+"""
+import os
+import gc
+import psutil
+import torch
+from typing import Optional, Tuple
+from dataclasses import dataclass
+
+# Import ComfyUI's model management functions
+try:
+    from comfy.model_management import (
+        get_torch_device,
+        get_free_memory,
+        get_total_memory,
+        free_memory,
+        soft_empty_cache,
+        unload_all_models,
+        is_device_cuda,
+    )
+    COMFY_AVAILABLE = True
+except ImportError:
+    COMFY_AVAILABLE = False
+
+
+@dataclass
+class DeviceCapabilities:
+    """Device capabilities for I/O and compute optimization."""
+    device_type: str  # 'cuda', 'cpu', 'mps', 'xpu'
+    total_vram_gb: float
+    free_vram_gb: float
+    total_ram_gb: float
+    free_ram_gb: float
+    cpu_cores: int
+    recommended_workers: int
+    max_tensor_size_gb: float  # Largest tensor we can safely load
+    supports_pinned_memory: bool
+    supports_async_streams: bool
+
+
+def get_device_capabilities(device: Optional[torch.device] = None) -> DeviceCapabilities:
+    """
+    Analyze device capabilities for optimal I/O configuration.
+
+    Args:
+        device: Target device (default: ComfyUI's torch device)
+
+    Returns:
+        DeviceCapabilities dataclass with recommended settings
+    """
+    if device is None:
+        if COMFY_AVAILABLE:
+            device = get_torch_device()
+        else:
+            device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+    device_type = device.type if hasattr(device, 'type') else 'cpu'
+
+    # Get RAM info
+    ram_info = psutil.virtual_memory()
+    total_ram_gb = ram_info.total / (1024**3)
+    free_ram_gb = ram_info.available / (1024**3)
+
+    # Get VRAM info
+    if device_type == 'cuda' and torch.cuda.is_available():
+        if COMFY_AVAILABLE:
+            total_vram = get_total_memory(device)
+            free_vram = get_free_memory(device)
+        else:
+            free_vram, total_vram = torch.cuda.mem_get_info(device)
+        total_vram_gb = total_vram / (1024**3)
+        free_vram_gb = free_vram / (1024**3)
+        supports_pinned = True
+        supports_async = True
+    elif device_type == 'mps':
+        total_vram_gb = total_ram_gb * 0.5  # MPS shares RAM
+        free_vram_gb = free_ram_gb * 0.5
+        supports_pinned = False
+        supports_async = False
+    else:
+        total_vram_gb = 0
+        free_vram_gb = 0
+        supports_pinned = False
+        supports_async = False
+
+    # CPU cores
+    cpu_cores = os.cpu_count() or 4
+
+    # Recommended workers based on hardware
+    if device_type == 'cuda':
+        # GPU: fewer workers since I/O is not the bottleneck
+        recommended_workers = min(4, cpu_cores // 2)
+    else:
+        # CPU: more workers for parallel I/O
+        recommended_workers = min(8, cpu_cores)
+
+    # Max tensor size: leave headroom for operations
+    if device_type == 'cuda':
+        max_tensor_gb = min(free_vram_gb * 0.8, free_ram_gb * 0.5)
+    else:
+        max_tensor_gb = free_ram_gb * 0.6
+
+    return DeviceCapabilities(
+        device_type=device_type,
+        total_vram_gb=total_vram_gb,
+        free_vram_gb=free_vram_gb,
+        total_ram_gb=total_ram_gb,
+        free_ram_gb=free_ram_gb,
+        cpu_cores=cpu_cores,
+        recommended_workers=recommended_workers,
+        max_tensor_size_gb=max_tensor_gb,
+        supports_pinned_memory=supports_pinned,
+        supports_async_streams=supports_async,
+    )
+
+
+def estimate_model_size(file_path: str) -> float:
+    """Estimate model size in GB from file size."""
+    try:
+        return os.path.getsize(file_path) / (1024**3)
+    except OSError:
+        return 0.0
+
+
+def can_fit_in_memory(
+    model_paths: list[str],
+    device: Optional[torch.device] = None,
+    safety_factor: float = 0.8
+) -> Tuple[bool, str]:
+    """
+    Check if models can fit in available memory.
+
+    Args:
+        model_paths: List of model file paths
+        device: Target device
+        safety_factor: Fraction of free memory to use (default 0.8)
+
+    Returns:
+        (can_fit, reason_message)
+    """
+    caps = get_device_capabilities(device)
+
+    total_size = sum(estimate_model_size(p) for p in model_paths)
+
+    if caps.device_type == 'cuda':
+        available = caps.free_vram_gb * safety_factor
+        mem_type = "VRAM"
+    else:
+        available = caps.free_ram_gb * safety_factor
+        mem_type = "RAM"
+
+    if total_size <= available:
+        return True, f"OK: {total_size:.2f}GB needed, {available:.2f}GB {mem_type} available"
+    else:
+        return False, f"Insufficient memory: {total_size:.2f}GB needed but only {available:.2f}GB {mem_type} available"
+
+
+def prepare_for_large_operation(
+    estimated_memory_gb: float,
+    device: Optional[torch.device] = None
+) -> bool:
+    """
+    Prepare device for a large memory operation by freeing resources.
+
+    Args:
+        estimated_memory_gb: Estimated memory needed in GB
+        device: Target device
+
+    Returns:
+        True if preparation succeeded
+    """
+    if not COMFY_AVAILABLE:
+        gc.collect()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        return True
+
+    if device is None:
+        device = get_torch_device()
+
+    memory_bytes = int(estimated_memory_gb * 1024**3)
+
+    # Use ComfyUI's free_memory which unloads models intelligently
+    try:
+        free_memory(memory_bytes, device)
+        return True
+    except Exception as e:
+        print(f"[DeviceUtils] Warning: Could not free memory: {e}")
+        return False
+
+
+def cleanup_after_operation():
+    """Clean up memory after a large operation."""
+    gc.collect()
+
+    if COMFY_AVAILABLE:
+        soft_empty_cache()
+    elif torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+
+
+def get_optimal_workers(
+    model_size_gb: float,
+    caps: Optional[DeviceCapabilities] = None
+) -> int:
+    """
+    Calculate optimal number of I/O workers based on model size and device.
+
+    Args:
+        model_size_gb: Model size in GB
+        caps: Pre-calculated device capabilities (optional)
+
+    Returns:
+        Recommended number of workers
+    """
+    if caps is None:
+        caps = get_device_capabilities()
+
+    # For very large models, fewer workers to avoid memory pressure
+    if model_size_gb > caps.free_ram_gb * 0.5:
+        return max(1, caps.recommended_workers // 2)
+
+    # For small models, use more workers
+    if model_size_gb < 2.0:
+        return min(8, caps.cpu_cores)
+
+    return caps.recommended_workers
+
+
+def get_processing_device_string(prefer_cuda: bool = True) -> str:
+    """
+    Get the best available processing device as a string.
+
+    Args:
+        prefer_cuda: Prefer CUDA over CPU if available
+
+    Returns:
+        Device string ('cuda', 'mps', 'cpu')
+    """
+    if prefer_cuda:
+        if torch.cuda.is_available():
+            return "cuda"
+        if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+            return "mps"
+    return "cpu"
+
+
+def print_device_info():
+    """Print device capabilities for debugging."""
+    caps = get_device_capabilities()
+    print(f"\n{'='*50}")
+    print("  Device Capabilities")
+    print(f"{'='*50}")
+    print(f"  Device Type: {caps.device_type}")
+    print(f"  Total VRAM: {caps.total_vram_gb:.2f} GB")
+    print(f"  Free VRAM: {caps.free_vram_gb:.2f} GB")
+    print(f"  Total RAM: {caps.total_ram_gb:.2f} GB")
+    print(f"  Free RAM: {caps.free_ram_gb:.2f} GB")
+    print(f"  CPU Cores: {caps.cpu_cores}")
+    print(f"  Recommended Workers: {caps.recommended_workers}")
+    print(f"  Max Tensor Size: {caps.max_tensor_size_gb:.2f} GB")
+    print(f"  Pinned Memory: {caps.supports_pinned_memory}")
+    print(f"  Async Streams: {caps.supports_async_streams}")
+    print(f"{'='*50}\n")

nodes/lora_extract.py

@@ -1,350 +0,0 @@
-"""
-Memory-efficient LoRA extraction node.
-
-This node loads two models from files and extracts a LoRA by computing the difference,
-without requiring the full models to be loaded into GPU memory via ComfyUI's model management.
-"""
-import os
-import torch
-import folder_paths
-import comfy.utils
-from comfy_api.latest import io
-from safetensors.torch import save_file
-from .merger_utils import MemoryEfficientSafeOpen
-
-
-CLAMP_QUANTILE = 0.99
-
-
-def extract_lora(diff, rank):
-    """Extract LoRA weights from a weight difference tensor using SVD."""
-    conv2d = (len(diff.shape) == 4)
-    kernel_size = None if not conv2d else diff.size()[2:4]
-    conv2d_3x3 = conv2d and kernel_size != (1, 1)
-    out_dim, in_dim = diff.size()[0:2]
-    rank = min(rank, in_dim, out_dim)
-
-    if conv2d:
-        if conv2d_3x3:
-            diff = diff.flatten(start_dim=1)
-        else:
-            diff = diff.squeeze()
-
-    U, S, Vh = torch.linalg.svd(diff.float())
-    U = U[:, :rank]
-    S = S[:rank]
-    U = U @ torch.diag(S)
-    Vh = Vh[:rank, :]
-
-    dist = torch.cat([U.flatten(), Vh.flatten()])
-    hi_val = torch.quantile(dist, CLAMP_QUANTILE)
-    low_val = -hi_val
-
-    U = U.clamp(low_val, hi_val)
-    Vh = Vh.clamp(low_val, hi_val)
-    if conv2d:
-        U = U.reshape(out_dim, rank, 1, 1)
-        Vh = Vh.reshape(rank, in_dim, kernel_size[0], kernel_size[1])
-    return (U, Vh)
-
-
-def extract_lora_from_files(
-    model_a_path: str,
-    model_b_path: str,
-    rank: int,
-    lora_type: str,
-    bias_diff: bool,
-    prefix_lora: str,
-    process_device: str,
-    save_dtype: str,
-) -> dict[str, torch.Tensor]:
-    """
-    Memory-efficient LoRA extraction by streaming tensors from safetensors files.
-    
-    Args:
-        model_a_path: Path to the finetuned/modified model
-        model_b_path: Path to the base/original model  
-        rank: LoRA rank for SVD decomposition
-        lora_type: "standard" for LoRA, "full_diff" for full difference
-        bias_diff: Whether to include bias differences
-        prefix_lora: Prefix for LoRA keys (e.g., "diffusion_model.")
-        process_device: Device for processing ("cuda" or "cpu")
-        save_dtype: Output dtype ("fp16", "bf16", "fp32")
-    
-    Returns:
-        Dictionary of LoRA tensors
-    """
-    output_sd = {}
-    
-    # Get dtype for saving
-    save_torch_dtype = {
-        "fp32": torch.float32, 
-        "fp16": torch.float16, 
-        "bf16": torch.bfloat16
-    }.get(save_dtype, torch.float16)
-    
-    # Open both files with memory-efficient handler
-    handler_a = MemoryEfficientSafeOpen(model_a_path)
-    handler_b = MemoryEfficientSafeOpen(model_b_path)
-    
-    try:
-        keys_a = set(handler_a.keys())
-        keys_b = set(handler_b.keys())
-        
-        # Only process keys that exist in both models
-        common_keys = keys_a & keys_b
-        weight_keys = [k for k in common_keys if k.endswith(".weight")]
-        bias_keys = [k for k in common_keys if k.endswith(".bias")] if bias_diff else []
-        
-        # Log key differences
-        missing_in_b = keys_a - keys_b
-        extra_in_b = keys_b - keys_a
-        if missing_in_b:
-            print(f"[LoRA Extract] {len(missing_in_b)} keys in model A not in model B (skipped)")
-        if extra_in_b:
-            print(f"[LoRA Extract] {len(extra_in_b)} keys in model B not in model A (ignored)")
-        
-        pbar = comfy.utils.ProgressBar(len(weight_keys) + len(bias_keys))
-        
-        for key in weight_keys:
-            # Load tensors one at a time
-            tensor_a = handler_a.get_tensor(key).to(device=process_device, dtype=torch.float32)
-            tensor_b = handler_b.get_tensor(key).to(device=process_device, dtype=torch.float32)
-            
-            # Compute difference: finetuned - base
-            weight_diff = tensor_a - tensor_b
-            
-            # Clean up immediately
-            del tensor_a, tensor_b
-            
-            # Extract LoRA key name
-            lora_key = key[:-7]  # Remove ".weight"
-            
-            if lora_type == "standard":
-                if weight_diff.ndim < 2:
-                    # 1D weights (e.g., LayerNorm) - store as diff if bias_diff enabled
-                    if bias_diff:
-                        output_sd[f"{prefix_lora}{lora_key}.diff"] = weight_diff.to(save_torch_dtype).cpu()
-                else:
-                    try:
-                        lora_up, lora_down = extract_lora(weight_diff, rank)
-                        output_sd[f"{prefix_lora}{lora_key}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu()
-                        output_sd[f"{prefix_lora}{lora_key}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu()
-                    except Exception as e:
-                        print(f"[LoRA Extract] Could not extract LoRA for {key}: {e}")
-            
-            elif lora_type == "full_diff":
-                output_sd[f"{prefix_lora}{lora_key}.diff"] = weight_diff.to(save_torch_dtype).cpu()
-            
-            del weight_diff
-            pbar.update(1)
-        
-        # Process bias keys
-        for key in bias_keys:
-            tensor_a = handler_a.get_tensor(key).to(device=process_device, dtype=torch.float32)
-            tensor_b = handler_b.get_tensor(key).to(device=process_device, dtype=torch.float32)
-            
-            bias_diff_tensor = tensor_a - tensor_b
-            del tensor_a, tensor_b
-            
-            lora_key = key[:-5]  # Remove ".bias"
-            output_sd[f"{prefix_lora}{lora_key}.diff_b"] = bias_diff_tensor.to(save_torch_dtype).cpu()
-            
-            del bias_diff_tensor
-            pbar.update(1)
-    
-    finally:
-        handler_a.__exit__(None, None, None)
-        handler_b.__exit__(None, None, None)
-    
-    return output_sd
-
-
-class LoRASaveFromFile(io.ComfyNode):
-    """
-    Memory-efficient LoRA extraction from safetensors files.
-    
-    Unlike the standard LoraSave node which requires pre-loaded and subtracted models,
-    this node loads models directly from files one tensor at a time, making it much
-    more memory efficient for extracting LoRAs from large models.
-    """
-    
-    @classmethod
-    def define_schema(cls):
-        return io.Schema(
-            node_id="LoRASaveFromFile",
-            display_name="Extract LoRA from Files (Memory Efficient)",
-            category="ModelUtils/LoRA",
-            description="Extracts a LoRA by computing the difference between two model files. "
-                        "Much more memory efficient than using ModelSubtract + LoraSave.",
-            inputs=[
-                io.Combo.Input(
-                    "model_a", 
-                    options=folder_paths.get_filename_list("diffusion_models"),
-                    tooltip="The finetuned/modified model (A - B = LoRA)"
-                ),
-                io.Combo.Input(
-                    "model_b", 
-                    options=folder_paths.get_filename_list("diffusion_models"),
-                    tooltip="The base/original model (A - B = LoRA)"
-                ),
-                io.Int.Input("rank", default=64, min=1, max=4096, step=1,
-                            tooltip="LoRA rank - higher preserves more detail but increases size"),
-                io.Combo.Input("lora_type", options=["standard", "full_diff"], default="standard",
-                              tooltip="standard: SVD decomposition to lora_up/lora_down, full_diff: raw weight differences"),
-                io.Boolean.Input("bias_diff", default=True,
-                                tooltip="Include bias differences in the LoRA"),
-                io.String.Input("output_filename", default="extracted_lora",
-                               tooltip="Output filename (without extension)"),
-                io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
-                io.Combo.Input("process_device", options=["cuda", "cpu"], default="cuda",
-                              tooltip="Device for SVD computation"),
-            ],
-            outputs=[
-                io.String.Output(display_name="output_path"),
-            ],
-            is_output_node=True,
-        )
-
-    @classmethod
-    def execute(
-        cls,
-        model_a: str,
-        model_b: str,
-        rank: int,
-        lora_type: str,
-        bias_diff: bool,
-        output_filename: str,
-        save_dtype: str,
-        process_device: str,
-    ) -> io.NodeOutput:
-        # Get full paths
-        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)
-        
-        # Extract LoRA
-        output_sd = extract_lora_from_files(
-            model_a_path=model_a_path,
-            model_b_path=model_b_path,
-            rank=rank,
-            lora_type=lora_type,
-            bias_diff=bias_diff,
-            prefix_lora="diffusion_model.",
-            process_device=process_device,
-            save_dtype=save_dtype,
-        )
-        
-        if not output_sd:
-            raise ValueError("No LoRA weights extracted - models may be identical or incompatible")
-        
-        # Save to loras folder
-        output_dir = folder_paths.get_folder_paths("loras")[0]
-        os.makedirs(output_dir, exist_ok=True)
-        output_path = os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
-        
-        save_file(output_sd, output_path)
-        print(f"[LoRA Extract] Saved {len(output_sd)} tensors to {output_path}")
-        
-        return io.NodeOutput(output_path)
-
-
-class LoRACheckpointSaveFromFile(io.ComfyNode):
-    """
-    Memory-efficient LoRA extraction from checkpoint files (which may contain both model and text encoder).
-    """
-    
-    @classmethod
-    def define_schema(cls):
-        return io.Schema(
-            node_id="LoRACheckpointSaveFromFile",
-            display_name="Extract LoRA from Checkpoints (Memory Efficient)",
-            category="ModelUtils/LoRA",
-            description="Extracts a LoRA by computing the difference between two checkpoint files. "
-                        "Handles both diffusion model and text encoder weights.",
-            inputs=[
-                io.Combo.Input(
-                    "checkpoint_a", 
-                    options=folder_paths.get_filename_list("checkpoints"),
-                    tooltip="The finetuned/modified checkpoint (A - B = LoRA)"
-                ),
-                io.Combo.Input(
-                    "checkpoint_b", 
-                    options=folder_paths.get_filename_list("checkpoints"),
-                    tooltip="The base/original checkpoint (A - B = LoRA)"
-                ),
-                io.Int.Input("rank", default=64, min=1, max=4096, step=1,
-                            tooltip="LoRA rank - higher preserves more detail but increases size"),
-                io.Combo.Input("lora_type", options=["standard", "full_diff"], default="standard",
-                              tooltip="standard: SVD decomposition, full_diff: raw weight differences"),
-                io.Boolean.Input("bias_diff", default=True,
-                                tooltip="Include bias differences in the LoRA"),
-                io.String.Input("output_filename", default="extracted_lora",
-                               tooltip="Output filename (without extension)"),
-                io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
-                io.Combo.Input("process_device", options=["cuda", "cpu"], default="cuda",
-                              tooltip="Device for SVD computation"),
-            ],
-            outputs=[
-                io.String.Output(display_name="output_path"),
-            ],
-            is_output_node=True,
-        )
-
-    @classmethod
-    def execute(
-        cls,
-        checkpoint_a: str,
-        checkpoint_b: str,
-        rank: int,
-        lora_type: str,
-        bias_diff: bool,
-        output_filename: str,
-        save_dtype: str,
-        process_device: str,
-    ) -> io.NodeOutput:
-        # Get full paths
-        checkpoint_a_path = folder_paths.get_full_path_or_raise("checkpoints", checkpoint_a)
-        checkpoint_b_path = folder_paths.get_full_path_or_raise("checkpoints", checkpoint_b)
-        
-        # Open handlers to detect prefixes
-        handler_a = MemoryEfficientSafeOpen(checkpoint_a_path)
-        keys = handler_a.keys()
-        handler_a.__exit__(None, None, None)
-        
-        # Auto-detect key prefixes
-        has_model = any(k.startswith("model.") for k in keys)
-        has_diffusion = any(k.startswith("diffusion_model.") for k in keys)
-        has_text_encoders = any(k.startswith("text_encoders.") or k.startswith("cond_stage_model.") for k in keys)
-        
-        output_sd = {}
-        
-        # Extract diffusion model LoRA
-        if has_model or has_diffusion:
-            prefix = "model.diffusion_model." if has_model else "diffusion_model."
-            model_lora = extract_lora_from_files(
-                model_a_path=checkpoint_a_path,
-                model_b_path=checkpoint_b_path,
-                rank=rank,
-                lora_type=lora_type,
-                bias_diff=bias_diff,
-                prefix_lora="diffusion_model.",
-                process_device=process_device,
-                save_dtype=save_dtype,
-            )
-            output_sd.update(model_lora)
-        
-        # Note: Text encoder extraction would need additional prefix handling
-        # For full checkpoint support, this would need to be expanded
-        
-        if not output_sd:
-            raise ValueError("No LoRA weights extracted - checkpoints may be identical or incompatible")
-        
-        # Save to loras folder
-        output_dir = folder_paths.get_folder_paths("loras")[0]
-        os.makedirs(output_dir, exist_ok=True)
-        output_path = os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
-        
-        save_file(output_sd, output_path)
-        print(f"[LoRA Extract] Saved {len(output_sd)} tensors to {output_path}")
-        
-        return io.NodeOutput(output_path)

nodes/lora_extract_svd.py

@@ -0,0 +1,935 @@
+"""
+LoRA Extraction via SVD decomposition.
+
+Native implementation of Low-Rank Adapter extraction with multiple rank selection modes.
+No external dependencies required.
+"""
+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 safetensors.torch import save_file
+from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
+from .device_utils import (
+    estimate_model_size, prepare_for_large_operation,
+    cleanup_after_operation, get_device_capabilities
+)
+
+
+# =============================================================================
+# 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 - 1))
+    
+    # Check if decomposition is worthwhile
+    if rank >= max_possible_rank // 2:
+        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) // 2:
+        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.squeeze()  # [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
+    combined_lora_up = torch.cat(all_lora_up, dim=0)
+    combined_lora_down = torch.stack(all_lora_down, dim=0).mean(dim=0)
+    rank = combined_lora_down.shape[0]
+
+    return combined_lora_up, combined_lora_down, rank
+
+
+# =============================================================================
+# Pattern Matching
+# =============================================================================
+
+def _compile_patterns(pattern_string: str) -> list:
+    """Compile regex patterns from whitespace-separated string."""
+    if not pattern_string or not pattern_string.strip():
+        return []
+    patterns = []
+    for p in pattern_string.split():
+        p = p.strip()
+        if p:
+            try:
+                patterns.append(re.compile(p))
+            except re.error:
+                pass
+    return patterns
+
+
+def _matches_any_pattern(key: str, patterns: list) -> bool:
+    """Check if key matches any pattern."""
+    return any(p.search(key) for p in patterns)
+
+
+# =============================================================================
+# Main Extraction Function
+# =============================================================================
+
+def extract_lora_from_files(
+    model_a_path: str,
+    model_b_path: str,
+    mode: str,
+    linear_param: float,
+    conv_param: float,
+    device: str,
+    save_dtype: 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,
+) -> dict[str, torch.Tensor]:
+    """
+    Extract LoRA from difference between two models.
+
+    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
+        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
+    """
+    output_sd = {}
+
+    save_torch_dtype = {
+        "fp32": torch.float32,
+        "fp16": torch.float16,
+        "bf16": torch.bfloat16,
+    }.get(save_dtype, torch.float16)
+
+    skip_patterns = _compile_patterns(skip_patterns_str)
+
+    # Prepare memory before heavy operation
+    total_size_gb = estimate_model_size(model_a_path) + estimate_model_size(model_b_path)
+    print(f"[LoRA Extract] Preparing memory for {total_size_gb:.2f}GB operation...")
+    prepare_for_large_operation(total_size_gb * 1.5, torch.device(device))  # 1.5x for SVD headroom
+
+    handler_a = MemoryEfficientSafeOpen(model_a_path)
+    handler_b = MemoryEfficientSafeOpen(model_b_path)
+
+    try:
+        keys_a = set(handler_a.keys())
+        keys_b = set(handler_b.keys())
+
+        weight_keys = [k for k in keys_a if k.endswith(".weight")]
+
+        pbar = comfy.utils.ProgressBar(len(weight_keys))
+
+        stats = {"extracted": 0, "full": 0, "skipped": 0, "chunked": 0}
+
+        for key in tqdm(weight_keys, desc="Extracting LoRA", unit="layers"):
+            lora_name = _format_lora_key(key)
+
+            if _matches_any_pattern(key, skip_patterns):
+                stats["skipped"] += 1
+                pbar.update(1)
+                continue
+
+            # Load tensors with pinned memory for CUDA
+            use_pinned = device == 'cuda'
+            
+            if key not in keys_b:
+                if mismatch_mode == "skip":
+                    stats["skipped"] += 1
+                    pbar.update(1)
+                    continue
+                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
+            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":
+                        stats["skipped"] += 1
+                        pbar.update(1)
+                        del tensor_a, tensor_b
+                        continue
+                    weight_diff = tensor_a
+                    del tensor_b
+                else:
+                    weight_diff = tensor_a - tensor_b
+                    del tensor_a, tensor_b
+
+            # Skip small differences
+            if min_diff > 0 and weight_diff.abs().max() < min_diff:
+                stats["skipped"] += 1
+                del weight_diff
+                pbar.update(1)
+                continue
+
+            # Skip 1D tensors
+            if weight_diff.ndim < 2:
+                stats["skipped"] += 1
+                del weight_diff
+                pbar.update(1)
+                continue
+
+            is_conv = weight_diff.ndim == 4
+
+            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"[LoRA 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:
+                            output_sd[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                            output_sd[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                            output_sd[f"{lora_name}.alpha"] = torch.tensor(rank).to(save_torch_dtype)
+                            stats["chunked"] += 1
+                            del weight_diff
+                            pbar.update(1)
+                            continue
+
+                print(f"[LoRA Extract] Failed: {key}: {e}")
+                stats["skipped"] += 1
+                del weight_diff
+                pbar.update(1)
+                continue
+
+            # Store result
+            if mode_str == "full":
+                output_sd[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
+                stats["full"] += 1
+            else:
+                lora_down, lora_up, _ = result
+                output_sd[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                output_sd[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                output_sd[f"{lora_name}.alpha"] = torch.tensor(lora_down.shape[0]).to(save_torch_dtype)
+                stats["extracted"] += 1
+
+            del weight_diff
+            pbar.update(1)
+
+        print(f"[LoRA 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()
+
+    return output_sd
+
+
+def _save_lora(output_sd: dict, output_filename: str) -> str:
+    """Save LoRA to file."""
+    if not output_sd:
+        raise ValueError("No LoRA weights extracted")
+
+    output_dir = folder_paths.get_folder_paths("loras")[0]
+    os.makedirs(output_dir, exist_ok=True)
+    output_path = os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
+
+    save_file(output_sd, output_path)
+    print(f"[LoRA Extract] Saved to {output_path}")
+
+    return output_path
+
+
+# =============================================================================
+# 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 _reconstruct_dots(key: str) -> str:
+    """Reconstruct dot structure from underscored key using heuristics."""
+    # Block indices: input_blocks_1 -> input_blocks.1
+    blocks = [
+        "input_blocks", "output_blocks", "middle_block",
+        "transformer_blocks", "single_transformer_blocks",
+        "double_blocks", "single_blocks",
+        "down_blocks", "up_blocks", "mid_block",
+        "attentions", "resnets", "upsamplers", "downsamplers"
+    ]
+    for b in blocks:
+        key = re.sub(f"{b}_(\\d+)", f"{b}.\\1", key)
+
+    # Sequential numbers: .1_1 -> .1.1
+    key = re.sub(r"(\.\d+)_(\d+)", r"\1.\2", key)
+
+    # Separators that should be preceded by dot
+    separators = [
+        "transformer_blocks", "single_transformer_blocks",
+        "attn1", "attn2", "attn",
+        "img_attn", "txt_attn",
+        "to_q", "to_k", "to_v", "to_out",
+        "q_proj", "k_proj", "v_proj", "out_proj",
+        "qkv", "proj", "ff", "net",
+        "norm1", "norm2", "norm3", "norm",
+        "time_emb_proj"
+    ]
+    for s in separators:
+        key = key.replace(f"_{s}", f".{s}")
+
+    # Special case: to_out_0 -> to_out.0
+    key = re.sub(r"to_out_(\d+)", r"to_out.\1", key)
+
+    return key
+
+
+def _format_lora_key(key: str) -> str:
+    """
+    Format the key for LoRA saving.
+    Standardizes to 'diffusion_model.' or 'transformer.' prefix.
+    """
+    if key.endswith(".weight"):
+        key = key[:-7]
+
+    # Handle ComfyUI Checkpoint wrapper
+    if key.startswith("model.diffusion_model."):
+        # Check if it's a Diffusers-format transformer inside a checkpoint
+        inner_key = key[22:]  # len("model.diffusion_model.")
+        if inner_key.startswith("transformer_blocks") or inner_key.startswith("single_transformer_blocks"):
+            return f"transformer.{inner_key}"
+        # Standard checkpoint or Original Flux
+        return f"diffusion_model.{inner_key}"
+
+    # Handle direct Diffusers keys
+    if key.startswith("transformer_blocks") or key.startswith("single_transformer_blocks"):
+        return f"transformer.{key}"
+
+    # Handle legacy lora_unet_ prefix (for resizing without base)
+    if key.startswith("lora_unet_"):
+        core = key[10:]
+        dotted = _reconstruct_dots(core)
+        # Check if reconstructed key implies transformer format
+        if dotted.startswith("transformer_blocks") or dotted.startswith("single_transformer_blocks"):
+            return f"transformer.{dotted}"
+        return f"diffusion_model.{dotted}"
+
+    # Handle already correct prefixes
+    if key.startswith("diffusion_model.") or key.startswith("transformer."):
+        return key
+    
+    # Handle known Unet blocks (Diffusers format UNet)
+    if any(key.startswith(p) for p in ["down_blocks", "up_blocks", "mid_block", "conv_in", "conv_out", "time_embedding", "class_embedding"]):
+        return f"diffusion_model.{key}"
+
+    # Default fallback to diffusion_model
+    return f"diffusion_model.{key}"
+
+
+def _get_common_inputs():
+    return [
+        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="Regex patterns for layers to skip"),
+    ]
+
+
+class LoRAExtractFixed(io.ComfyNode):
+    """Extract LoRA with fixed rank."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAExtractFixed",
+            display_name="LoRA Extract (Fixed Rank)",
+            category="ModelUtils/LoRA",
+            description="Extract LoRA with specified fixed rank for each layer type.",
+            inputs=[
+                *_get_model_inputs(),
+                io.Int.Input("linear_dim", default=64, min=1, max=4096,
+                            tooltip="Rank for linear/attention layers"),
+                io.Int.Input("conv_dim", default=32, min=1, max=4096,
+                            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) -> 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_sd = extract_lora_from_files(
+            model_a_path, model_b_path, "fixed", linear_dim, conv_dim,
+            device, save_dtype, linear_dim, conv_dim,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers, svd_niter
+        )
+
+        return io.NodeOutput(_save_lora(output_sd, output_filename))
+
+
+class LoRAExtractRatio(io.ComfyNode):
+    """Extract LoRA with singular value ratio threshold."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAExtractRatio",
+            display_name="LoRA Extract (Ratio)",
+            category="ModelUtils/LoRA",
+            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=3072,
+                            tooltip="Maximum rank for linear layers"),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                            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) -> 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_sd = extract_lora_from_files(
+            model_a_path, model_b_path, "ratio", linear_ratio, conv_ratio,
+            device, save_dtype, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers
+        )
+
+        return io.NodeOutput(_save_lora(output_sd, output_filename))
+
+
+class LoRAExtractQuantile(io.ComfyNode):
+    """Extract LoRA with cumulative singular value quantile."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAExtractQuantile",
+            display_name="LoRA Extract (Quantile)",
+            category="ModelUtils/LoRA",
+            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=3072,
+                            tooltip="Maximum rank for linear layers"),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                            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) -> 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_sd = extract_lora_from_files(
+            model_a_path, model_b_path, "quantile", linear_quantile, conv_quantile,
+            device, save_dtype, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers
+        )
+
+        return io.NodeOutput(_save_lora(output_sd, output_filename))
+
+
+class LoRAExtractKnee(io.ComfyNode):
+    """Extract LoRA with automatic knee detection."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAExtractKnee",
+            display_name="LoRA Extract (Knee Detection)",
+            category="ModelUtils/LoRA",
+            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=3072,
+                            tooltip="Maximum rank for linear layers"),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                            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) -> 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_sd = extract_lora_from_files(
+            model_a_path, model_b_path, knee_method, 0, 0,
+            device, save_dtype, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers
+        )
+
+        return io.NodeOutput(_save_lora(output_sd, output_filename))
+
+
+class LoRAExtractFrobenius(io.ComfyNode):
+    """Extract LoRA preserving target fraction of Frobenius norm."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAExtractFrobenius",
+            display_name="LoRA Extract (Frobenius)",
+            category="ModelUtils/LoRA",
+            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=3072,
+                            tooltip="Maximum rank for linear layers"),
+                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                            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) -> 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_sd = extract_lora_from_files(
+            model_a_path, model_b_path, "sv_fro", linear_target, conv_target,
+            device, save_dtype, linear_max_rank, conv_max_rank,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers
+        )
+
+        return io.NodeOutput(_save_lora(output_sd, output_filename))

nodes/lora_merger.py

@@ -0,0 +1,549 @@
+"""
+LoRA Multi-Merge - Merge multiple LoRAs into a single LoRA file.
+Resolves different naming conventions and ranks via concatenation.
+"""
+import os
+import torch
+import folder_paths
+import comfy.utils
+from tqdm import tqdm
+from comfy_api.latest import io
+from safetensors.torch import save_file
+from typing import List, Dict, Tuple, Optional
+from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
+from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
+from .lora_resize import (
+    detect_lora_format, 
+    extract_lora_pairs, 
+    detect_lora_rank,
+    _compile_patterns,
+    _matches_any_pattern,
+    _format_lora_key
+)
+
+def merge_multi_loras(
+    lora_paths: List[str],
+    lora_weights: List[float],
+    merge_mode: str,
+    device: str,
+    save_dtype: torch.dtype,
+    output_filename: str,
+    verbose: bool = True,
+) -> str:
+    """
+    Merge multiple LoRAs into a single LoRA file.
+    
+    Modes:
+    - concatenate: Mathematically sound merge by stacking ranks.
+                   New rank = sum(input ranks). Alpha set to new rank.
+    - weighted_sum: Weighted sum of A and B weights (Kohya style).
+                    If ranks differ, smaller ones are zero-padded to match the largest rank.
+                    New rank = max(input ranks).
+    """
+    # Estimate memory
+    total_size_gb = sum(estimate_model_size(p) for p in lora_paths)
+    if verbose:
+        print(f"[LoRA Multi-Merge] Preparing memory for {total_size_gb:.2f}GB operation...")
+        print(f"[LoRA Multi-Merge] Mode: {merge_mode}")
+    
+    prepare_for_large_operation(total_size_gb * 2.5, torch.device(device))
+
+    handlers = [MemoryEfficientSafeOpen(p) for p in lora_paths]
+    
+    try:
+        # 1. Analyze all LoRAs
+        layer_map = {}
+        lora_infos = []
+        
+        for i, handler in enumerate(handlers):
+            keys = handler.keys()
+            fmt = detect_lora_format(keys)
+            pairs = extract_lora_pairs(keys, fmt)
+            rank, alpha = detect_lora_rank(handler, pairs)
+            
+            info = {
+                "name": os.path.basename(lora_paths[i]),
+                "format": fmt,
+                "pairs": pairs,
+                "rank": rank,
+                "alpha": alpha,
+                "weight": lora_weights[i],
+                "scale": alpha / rank if rank > 0 else 1.0
+            }
+            lora_infos.append(info)
+            
+            if verbose:
+                print(f"[LoRA Multi-Merge] LoRA {i+1} ({info['name']}): {len(pairs)} layers, dim={rank}, format={fmt['format']}")
+            
+            for block_name in pairs.keys():
+                core = _format_lora_key(block_name)
+                if core not in layer_map:
+                    layer_map[core] = []
+                layer_map[core].append(i)
+
+        if verbose:
+            print(f"[LoRA Multi-Merge] Total unique layers after resolving naming: {len(layer_map)}")
+
+        output_sd = {}
+        pbar = comfy.utils.ProgressBar(len(layer_map))
+
+        # 2. Merge each core layer
+        with torch.no_grad():
+            for core, indices in tqdm(layer_map.items(), desc="Merging layers", unit="layers"):
+                downs = []
+                ups = []
+                is_conv = False
+                max_rank = 0
+                
+                # First pass: load and determine max rank
+                for idx in indices:
+                    info = lora_infos[idx]
+                    handler = handlers[idx]
+                    orig_block = next(b for b in info["pairs"].keys() if _format_lora_key(b) == core)
+                    block_keys = info["pairs"][orig_block]
+                    
+                    if "down" not in block_keys or "up" not in block_keys:
+                        continue
+                        
+                    t_down = handler.get_tensor(block_keys["down"])
+                    t_up = handler.get_tensor(block_keys["up"])
+                    
+                    if device == 'cuda':
+                        t_down = transfer_to_gpu_pinned(t_down, device, torch.float32)
+                        t_up = transfer_to_gpu_pinned(t_up, device, torch.float32)
+                    else:
+                        t_down = t_down.to(device=device, dtype=torch.float32)
+                        t_up = t_up.to(device=device, dtype=torch.float32)
+                    
+                    # Store with original info for padding/weighting
+                    is_conv = len(t_down.shape) == 4
+                    current_rank = t_down.shape[0]
+                    max_rank = max(max_rank, current_rank)
+                    
+                    # Apply scale immediately for concatenate mode, or keep for later
+                    if merge_mode == "concatenate":
+                        effective_weight = info["weight"] * info["scale"]
+                        t_up = t_up * effective_weight
+                    
+                    downs.append((t_down, info))
+                    ups.append((t_up, info))
+
+                if not downs:
+                    pbar.update(1)
+                    continue
+
+                if merge_mode == "concatenate":
+                    # Stack ranks
+                    merged_down = torch.cat([d[0] for d in downs], dim=0)
+                    merged_up = torch.cat([u[0] for u in ups], dim=1)
+                    new_rank = merged_down.shape[0]
+                    new_alpha = float(new_rank)
+                else:
+                    # weighted_sum (Kohya style)
+                    # Result = sum( weight_i * Padded(B_i) ), sum( weight_i * Padded(A_i) )
+                    # Note: We apply scale to weights here to normalize different alphas
+                    merged_down = torch.zeros_like(downs[0][0])
+                    # Need to handle different ranks via padding
+                    target_down_shape = list(downs[0][0].shape)
+                    target_down_shape[0] = max_rank
+                    target_up_shape = list(ups[0][0].shape)
+                    target_up_shape[1] = max_rank
+                    
+                    merged_down = torch.zeros(target_down_shape, device=device, dtype=torch.float32)
+                    merged_up = torch.zeros(target_up_shape, device=device, dtype=torch.float32)
+                    
+                    for (t_d, info_d), (t_u, info_u) in zip(downs, ups):
+                        r = t_d.shape[0]
+                        w = info_d["weight"]
+                        # We also apply sqrt(scale) to both to distribute the scale factor?
+                        # Kohya just uses weights. But to be safe with different alphas,
+                        # we apply scale to the final delta.
+                        # For direct weight merge, we'll just use weights.
+                        
+                        if r < max_rank:
+                            # Pad down: [r, in...] -> [max_rank, in...]
+                            pad_d = [0] * (len(t_d.shape) * 2)
+                            pad_d[-1] = max_rank - r # last dim in pad is first dim in tensor (reversed)
+                            # Wait, F.pad uses reverse order of dims.
+                            # For [r, in], padding is (0,0, 0, max_rank-r)
+                            padding_d = [0, 0] * (len(t_d.shape) - 1) + [0, max_rank - r]
+                            t_d = torch.nn.functional.pad(t_d, tuple(padding_d))
+                            
+                            # Pad up: [out, r...] -> [out, max_rank...]
+                            # For [out, r], padding is (0, max_rank-r, 0, 0)
+                            padding_u = [0, max_rank - r] + [0, 0] * (len(t_u.shape) - 1)
+                            t_u = torch.nn.functional.pad(t_u, tuple(padding_u))
+                        
+                        merged_down += w * t_d
+                        merged_up += w * t_u
+                    
+                    new_rank = max_rank
+                    # Alpha is usually max of alphas or same as rank
+                    new_alpha = float(max_rank)
+
+                # Store
+                output_sd[f"{core}.lora_down.weight"] = merged_down.to(save_dtype).cpu().contiguous()
+                output_sd[f"{core}.lora_up.weight"] = merged_up.to(save_dtype).cpu().contiguous()
+                output_sd[f"{core}.alpha"] = torch.tensor(new_alpha, dtype=save_dtype)
+                
+                # Cleanup
+                del merged_down, merged_up
+                for d, _ in downs: del d
+                for u, _ in ups: del u
+                
+                pbar.update(1)
+
+        # 3. Final Summary
+        if verbose:
+            matched_stats = {i: 0 for i in range(len(lora_paths))}
+            for indices in layer_map.values():
+                for idx in indices:
+                    matched_stats[idx] += 1
+            
+            print(f"[LoRA Multi-Merge] --- Merge Summary ---")
+            for i, info in enumerate(lora_infos):
+                print(f"[LoRA Multi-Merge] LoRA {i+1}: {matched_stats[i]}/{len(info['pairs'])} layers used in merge")
+            print(f"[LoRA Multi-Merge] Output state dict has {len(output_sd)} tensors")
+
+        # 4. Metadata
+        metadata = {
+            "ss_training_comment": f"Merged {len(lora_paths)} LoRAs via concatenation",
+            "ss_network_module": "networks.lora",
+        }
+        
+        # 4. Save
+        output_dir = folder_paths.get_folder_paths("loras")[0]
+        os.makedirs(output_dir, exist_ok=True)
+        output_path = os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
+        
+        save_file(output_sd, output_path, metadata)
+        if verbose:
+            print(f"[LoRA Multi-Merge] Saved merged LoRA to {output_path}")
+            
+        return output_path
+
+    finally:
+        for h in handlers:
+            h.__exit__(None, None, None)
+        cleanup_after_operation()
+
+
+class LoRAMultiMerge(io.ComfyNode):
+    """Merge multiple LoRAs into a single LoRA file."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAMultiMerge",
+            display_name="LoRA Multi-Merge",
+            category="ModelUtils/LoRA/Merge",
+            description="Merge 1-8 LoRAs into a single LoRA file. This resolves different ranks and naming conventions.",
+            inputs=[
+                io.Combo.Input("lora_count", options=[str(i) for i in range(1, 9)], default="2",
+                              tooltip="Number of LoRAs to merge"),
+                # LoRA 1
+                io.Combo.Input("lora_1", options=folder_paths.get_filename_list("loras"), tooltip="First LoRA"),
+                io.Float.Input("weight_1", default=1.0, min=-10.0, max=10.0, step=0.01),
+                # LoRA 2
+                io.Combo.Input("lora_2", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_2", default=1.0, min=-10.0, max=10.0, step=0.01),
+                # LoRA 3
+                io.Combo.Input("lora_3", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_3", default=1.0, min=-10.0, max=10.0, step=0.01),
+                # LoRA 4
+                io.Combo.Input("lora_4", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_4", default=1.0, min=-10.0, max=10.0, step=0.01),
+                # LoRA 5
+                io.Combo.Input("lora_5", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_5", default=1.0, min=-10.0, max=10.0, step=0.01),
+                # LoRA 6
+                io.Combo.Input("lora_6", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_6", default=1.0, min=-10.0, max=10.0, step=0.01),
+                # LoRA 7
+                io.Combo.Input("lora_7", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_7", default=1.0, min=-10.0, max=10.0, step=0.01),
+                # LoRA 8
+                io.Combo.Input("lora_8", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_8", default=1.0, min=-10.0, max=10.0, step=0.01),
+                
+                io.Combo.Input("merge_mode", options=["concatenate", "weighted_sum"], default="concatenate",
+                              tooltip="concatenate: safe, increases rank. weighted_sum: fixed rank (to max input rank), mathematically lossy but standard in Kohya."),
+                io.String.Input("output_filename", default="merged_lora"),
+                io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
+                io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, 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,
+                merge_mode, output_filename, save_dtype, device) -> io.NodeOutput:
+        
+        count = int(lora_count)
+        names = [lora_1, lora_2, lora_3, lora_4, lora_5, lora_6, lora_7, lora_8]
+        weights = [weight_1, weight_2, weight_3, weight_4, weight_5, weight_6, weight_7, weight_8]
+        
+        valid_paths = []
+        valid_weights = []
+        
+        for i in range(count):
+            if names[i] and names[i] != "None":
+                path = folder_paths.get_full_path_or_raise("loras", names[i])
+                valid_paths.append(path)
+                valid_weights.append(weights[i])
+                
+        if not valid_paths:
+            raise ValueError("No LoRAs selected for merging.")
+            
+        dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
+        
+        path = merge_multi_loras(
+            lora_paths=valid_paths,
+            lora_weights=valid_weights,
+            merge_mode=merge_mode,
+            device=device,
+            save_dtype=dtype,
+            output_filename=output_filename
+        )
+        
+        return io.NodeOutput(path)
+
+
+class LoRAMultiMergeDARE(io.ComfyNode):
+    """Merge multiple LoRAs into a single LoRA file using DARE-Ties method."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAMultiMergeDARE",
+            display_name="LoRA Multi-Merge (DARE-Ties)",
+            category="ModelUtils/LoRA/Merge",
+            description="Merge 1-8 LoRAs into a single LoRA file using DARE-Ties. Drops small values and resolves sign conflicts.",
+            inputs=[
+                io.Combo.Input("lora_count", options=[str(i) for i in range(1, 9)], default="2"),
+                # LoRA 1
+                io.Combo.Input("lora_1", options=folder_paths.get_filename_list("loras"), tooltip="First LoRA"),
+                io.Float.Input("weight_1", default=1.0, min=-10.0, max=10.0, step=0.01),
+                # ...
+                io.Combo.Input("lora_2", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_2", default=1.0, min=-10.0, max=10.0, step=0.01),
+                io.Combo.Input("lora_3", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_3", default=1.0, min=-10.0, max=10.0, step=0.01),
+                io.Combo.Input("lora_4", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_4", default=1.0, min=-10.0, max=10.0, step=0.01),
+                io.Combo.Input("lora_5", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_5", default=1.0, min=-10.0, max=10.0, step=0.01),
+                io.Combo.Input("lora_6", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_6", default=1.0, min=-10.0, max=10.0, step=0.01),
+                io.Combo.Input("lora_7", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_7", default=1.0, min=-10.0, max=10.0, step=0.01),
+                io.Combo.Input("lora_8", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
+                io.Float.Input("weight_8", default=1.0, min=-10.0, max=10.0, step=0.01),
+
+                io.Float.Input("drop_rate", default=0.1, min=0.0, max=1.0, step=0.01, tooltip="DARE drop rate"),
+                io.Float.Input("trim_quantile", default=0.2, min=0.0, max=1.0, step=0.01, tooltip="TIES trim quantile (drops smallest values)"),
+                io.Int.Input("seed", default=42, min=0, max=0xffffffffffffffff),
+                io.String.Input("output_filename", default="merged_lora_dare"),
+                io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
+                io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, 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,
+                drop_rate, trim_quantile, seed, output_filename, save_dtype, device) -> io.NodeOutput:
+        
+        count = int(lora_count)
+        names = [lora_1, lora_2, lora_3, lora_4, lora_5, lora_6, lora_7, lora_8]
+        weights = [weight_1, weight_2, weight_3, weight_4, weight_5, weight_6, weight_7, weight_8]
+        
+        valid_paths = []
+        valid_weights = []
+        
+        for i in range(count):
+            if names[i] and names[i] != "None":
+                path = folder_paths.get_full_path_or_raise("loras", names[i])
+                valid_paths.append(path)
+                valid_weights.append(weights[i])
+                
+        if not valid_paths:
+            raise ValueError("No LoRAs selected for merging.")
+            
+        dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
+        
+        path = merge_multi_loras_dare(
+            lora_paths=valid_paths,
+            lora_weights=valid_weights,
+            drop_rate=drop_rate,
+            trim_quantile=trim_quantile,
+            seed=seed,
+            device=device,
+            save_dtype=dtype,
+            output_filename=output_filename
+        )
+        
+        return io.NodeOutput(path)
+
+def merge_multi_loras_dare(
+    lora_paths: List[str],
+    lora_weights: List[float],
+    drop_rate: float,
+    trim_quantile: float,
+    seed: int,
+    device: str,
+    save_dtype: torch.dtype,
+    output_filename: str,
+    verbose: bool = True,
+) -> str:
+    """
+    Merge multiple LoRAs using DARE-Ties method applied to weights.
+    Different ranks are handled via zero-padding to the max rank.
+    """
+    total_size_gb = sum(estimate_model_size(p) for p in lora_paths)
+    if verbose:
+        print(f"[LoRA Multi-Merge DARE] Preparing memory for {total_size_gb:.2f}GB operation...")
+        print(f"[LoRA Multi-Merge DARE] Drop rate: {drop_rate}, Trim quantile: {trim_quantile}")
+
+    prepare_for_large_operation(total_size_gb * 2.5, torch.device(device))
+    handlers = [MemoryEfficientSafeOpen(p) for p in lora_paths]
+    
+    rng = torch.Generator(device=device).manual_seed(seed)
+
+    try:
+        # 1. Analyze
+        layer_map = {}
+        lora_infos = []
+        for i, handler in enumerate(handlers):
+            keys = handler.keys()
+            fmt = detect_lora_format(keys)
+            pairs = extract_lora_pairs(keys, fmt)
+            rank, alpha = detect_lora_rank(handler, pairs)
+            info = {
+                "name": os.path.basename(lora_paths[i]),
+                "pairs": pairs,
+                "rank": rank,
+                "alpha": alpha,
+                "weight": lora_weights[i]
+            }
+            lora_infos.append(info)
+            
+            if verbose:
+                print(f"[LoRA Multi-Merge DARE] LoRA {i+1} ({info['name']}): {len(pairs)} layers, dim={rank}")
+
+            for block_name in pairs.keys():
+                core = _format_lora_key(block_name)
+                if core not in layer_map: layer_map[core] = []
+                layer_map[core].append(i)
+
+        if verbose:
+            print(f"[LoRA Multi-Merge DARE] Total unique layers after resolving naming: {len(layer_map)}")
+
+        output_sd = {}
+        pbar = comfy.utils.ProgressBar(len(layer_map))
+
+        # 2. Merge
+        with torch.no_grad():
+            for core, indices in tqdm(layer_map.items(), desc="Merging layers (DARE)", unit="layers"):
+                downs = []
+                ups = []
+                max_rank = 0
+                
+                for idx in indices:
+                    info = lora_infos[idx]
+                    orig_block = next(b for b in info["pairs"].keys() if _format_lora_key(b) == core)
+                    block_keys = info["pairs"][orig_block]
+                    if "down" not in block_keys or "up" not in block_keys: continue
+                    
+                    t_d = handlers[idx].get_tensor(block_keys["down"]).to(device=device, dtype=torch.float32)
+                    t_u = handlers[idx].get_tensor(block_keys["up"]).to(device=device, dtype=torch.float32)
+                    
+                    max_rank = max(max_rank, t_d.shape[0])
+                    downs.append((t_d, info["weight"]))
+                    ups.append((t_u, info["weight"]))
+
+                if not downs:
+                    pbar.update(1)
+                    continue
+
+                def process_ties_dare(tensors, weights, dim_to_pad):
+                    # Pad all to max_rank
+                    padded = []
+                    for t, w in tensors:
+                        if t.shape[dim_to_pad] < max_rank:
+                            padding = [0] * (len(t.shape) * 2)
+                            # dim_to_pad 0 (down) -> last pair in pad
+                            # dim_to_pad 1 (up) -> second to last pair in pad
+                            rev_dim = len(t.shape) - 1 - dim_to_pad
+                            padding[rev_dim*2 + 1] = max_rank - t.shape[dim_to_pad]
+                            t = torch.nn.functional.pad(t, tuple(padding))
+                        padded.append(t * w)
+                    
+                    # DARE
+                    if drop_rate > 0:
+                        for i in range(len(padded)):
+                            mask = (torch.rand(padded[i].shape, generator=rng, device=device) > drop_rate).float()
+                            padded[i] = (padded[i] * mask) / (1 - drop_rate)
+                    
+                    # TIES
+                    # 1. Trim
+                    if trim_quantile > 0:
+                        for i in range(len(padded)):
+                            flat = padded[i].abs().flatten()
+                            k = int(len(flat) * trim_quantile)
+                            if k > 0:
+                                threshold = torch.kthvalue(flat, k).values
+                                padded[i] = torch.where(padded[i].abs() < threshold, torch.zeros_like(padded[i]), padded[i])
+                    
+                    # 2. Elect & Merge
+                    stacked = torch.stack(padded) # [N, ...]
+                    signs = torch.sign(stacked)
+                    sum_signs = signs.sum(dim=0)
+                    dominant_sign = torch.sign(sum_signs)
+                    
+                    # Filter those matching dominant sign
+                    mask = (signs == dominant_sign) & (dominant_sign != 0)
+                    filtered = torch.where(mask, stacked, torch.zeros_like(stacked))
+                    
+                    # Average matching signs
+                    count = mask.sum(dim=0)
+                    result = filtered.sum(dim=0) / torch.clamp(count, min=1.0)
+                    return result
+
+                merged_down = process_ties_dare(downs, [1.0]*len(downs), 0)
+                merged_up = process_ties_dare(ups, [1.0]*len(ups), 1)
+
+                output_sd[f"{core}.lora_down.weight"] = merged_down.to(save_dtype).cpu().contiguous()
+                output_sd[f"{core}.lora_up.weight"] = merged_up.to(save_dtype).cpu().contiguous()
+                output_sd[f"{core}.alpha"] = torch.tensor(float(max_rank), dtype=save_dtype)
+                
+                pbar.update(1)
+
+        # Final Summary
+        if verbose:
+            matched_stats = {i: 0 for i in range(len(lora_paths))}
+            for indices in layer_map.values():
+                for idx in indices:
+                    matched_stats[idx] += 1
+            
+            print(f"[LoRA Multi-Merge DARE] --- Merge Summary ---")
+            for i, info in enumerate(lora_infos):
+                print(f"[LoRA Multi-Merge DARE] LoRA {i+1}: {matched_stats[i]}/{len(info['pairs'])} layers used in merge")
+            print(f"[LoRA Multi-Merge DARE] Output state dict has {len(output_sd)} tensors")
+
+        # Save
+        output_dir = folder_paths.get_folder_paths("loras")[0]
+        os.makedirs(output_dir, exist_ok=True)
+        output_path = os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
+        save_file(output_sd, output_path, {"ss_training_comment": "Merged via DARE-Ties"})
+        return output_path
+
+    finally:
+        for h in handlers: h.__exit__(None, None, None)
+        cleanup_after_operation()

nodes/lora_resize.py

@@ -0,0 +1,955 @@
+"""
+LoRA Resize - Resize existing LoRAs to different ranks.
+
+Based on kohya_ss resize_lora.py. Merges LoRA weights and re-extracts via SVD.
+Supports fixed rank and dynamic methods (sv_ratio, sv_fro, sv_cumulative).
+"""
+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 safetensors.torch import save_file
+from typing import Optional, Dict, Tuple, List
+from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
+from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
+
+# Reuse rank functions from extraction module
+from .lora_extract_svd import (
+    _index_sv_ratio,
+    _index_sv_cumulative,
+    _index_sv_fro,
+    _svd_extract_linear,
+    _svd_extract_conv,
+    _compile_patterns,
+    _matches_any_pattern,
+    _format_lora_key,
+)
+
+
+MIN_SV = 1e-6
+
+
+# =============================================================================
+# LoRA Format Detection - Support various LoRA sources
+# =============================================================================
+
+def detect_lora_format(keys: List[str]) -> Dict:
+    """
+    Detect LoRA format and key patterns from various sources.
+
+    Supports:
+    - Kohya/A1111: lora_unet_*, lora_te_*
+    - Diffusers: *.lora_A.weight, *.lora_B.weight
+    - PEFT: base_model.model.*.lora_A/B
+    - ComfyUI native: *.lora_down.weight, *.lora_up.weight
+    - HuggingFace: *.lora.down.weight, *.lora.up.weight (dots instead of underscores)
+    - Full diff: *.diff, *.diff_b (full weight differences, not low-rank)
+    - Flux/SDXL: Various transformer key patterns
+
+    Returns:
+        Dict with format info and key patterns
+    """
+    format_info = {
+        "format": "unknown",
+        "down_suffix": ".lora_down.weight",
+        "up_suffix": ".lora_up.weight",
+        "alpha_suffix": ".alpha",
+        "has_alpha": False,
+        "key_count": 0,
+        "is_full_diff": False,  # New flag for full diff format
+    }
+
+    # Check for alpha keys
+    alpha_keys = [k for k in keys if '.alpha' in k or k.endswith('alpha')]
+    format_info["has_alpha"] = len(alpha_keys) > 0
+
+    # Detect format by key patterns
+    sample_keys = keys[:50]  # Check more keys for full_diff detection
+
+    # Full diff format: *.diff (NOT .lora_down, just straight weight diff)
+    if any(k.endswith('.diff') for k in sample_keys):
+        format_info["format"] = "full_diff"
+        format_info["down_suffix"] = ".diff"  # Used for pairing, but no up
+        format_info["up_suffix"] = None  # No up weight in full diff
+        format_info["alpha_suffix"] = None
+        format_info["is_full_diff"] = True
+
+    # Kohya/A1111 format: lora_unet_down_blocks_0_*.lora_down.weight
+    elif any('lora_unet_' in k or 'lora_te' in k for k in sample_keys):
+        format_info["format"] = "kohya"
+        format_info["down_suffix"] = ".lora_down.weight"
+        format_info["up_suffix"] = ".lora_up.weight"
+
+    # Diffusers format: transformer.*.lora_A.weight
+    elif any('.lora_A.weight' in k for k in sample_keys):
+        format_info["format"] = "diffusers"
+        format_info["down_suffix"] = ".lora_A.weight"
+        format_info["up_suffix"] = ".lora_B.weight"
+        format_info["alpha_suffix"] = ".alpha"
+
+    # PEFT format: base_model.model.*.lora_A.default.weight
+    elif any('base_model.model.' in k and 'lora_' in k for k in sample_keys):
+        format_info["format"] = "peft"
+        # Handle nested structure
+        format_info["down_suffix"] = ".lora_A.default.weight"
+        format_info["up_suffix"] = ".lora_B.default.weight"
+
+    # HuggingFace format: transformer.*.lora.down.weight (dots instead of underscores)
+    elif any('.lora.down.weight' in k for k in sample_keys):
+        format_info["format"] = "huggingface"
+        format_info["down_suffix"] = ".lora.down.weight"
+        format_info["up_suffix"] = ".lora.up.weight"
+        format_info["alpha_suffix"] = ".alpha"
+
+    # ComfyUI native / standard
+    elif any('.lora_down.weight' in k for k in sample_keys):
+        format_info["format"] = "comfy"
+        format_info["down_suffix"] = ".lora_down.weight"
+        format_info["up_suffix"] = ".lora_up.weight"
+
+    # Count LoRA layers
+    down_keys = [k for k in keys if format_info["down_suffix"] in k]
+    format_info["key_count"] = len(down_keys)
+
+    return format_info
+
+
+
+def extract_lora_pairs(keys: List[str], format_info: Dict) -> Dict[str, Dict[str, str]]:
+    """
+    Group LoRA keys into down/up/alpha pairs.
+
+    For full_diff format, groups .diff and .diff_b keys.
+
+    Returns:
+        Dict[block_name, {"down": key, "up": key, "alpha": key}]
+        For full_diff: {"diff": key, "diff_b": key}
+    """
+    down_suffix = format_info["down_suffix"]
+    up_suffix = format_info["up_suffix"]
+    alpha_suffix = format_info["alpha_suffix"]
+    is_full_diff = format_info.get("is_full_diff", False)
+
+    pairs = {}
+
+    if is_full_diff:
+        # Full diff format: group .diff and .diff_b
+        for key in keys:
+            if key.endswith('.diff'):
+                block_name = key[:-5]  # Remove .diff
+                if block_name not in pairs:
+                    pairs[block_name] = {}
+                pairs[block_name]["diff"] = key
+            elif key.endswith('.diff_b'):
+                block_name = key[:-7]  # Remove .diff_b
+                if block_name not in pairs:
+                    pairs[block_name] = {}
+                pairs[block_name]["diff_b"] = key
+    else:
+        # Standard LoRA format
+        for key in keys:
+            if down_suffix and down_suffix in key:
+                block_name = key.replace(down_suffix, "")
+                if block_name not in pairs:
+                    pairs[block_name] = {}
+                pairs[block_name]["down"] = key
+            elif up_suffix and up_suffix in key:
+                block_name = key.replace(up_suffix, "")
+                if block_name not in pairs:
+                    pairs[block_name] = {}
+                pairs[block_name]["up"] = key
+            elif alpha_suffix and (alpha_suffix in key or key.endswith('.alpha')):
+                # Handle various alpha key formats
+                block_name = key.replace(alpha_suffix, "").replace(".alpha", "")
+                if block_name not in pairs:
+                    pairs[block_name] = {}
+                pairs[block_name]["alpha"] = key
+
+    return pairs
+
+
+def detect_lora_rank(handler: MemoryEfficientSafeOpen, pairs: Dict) -> Tuple[int, float]:
+    """
+    Detect the rank and alpha of an existing LoRA.
+
+    Returns:
+        (network_dim, network_alpha)
+    """
+    network_dim = None
+    network_alpha = None
+
+    for block_name, block_keys in pairs.items():
+        if "down" not in block_keys:
+            continue
+
+        # Get dim from down weight shape
+        if network_dim is None:
+            down_key = block_keys["down"]
+            shape = handler.header[down_key]["shape"]
+            # Linear: [rank, in_features] or Conv: [rank, in_ch, k, k]
+            network_dim = shape[0]
+
+        # Get alpha if present
+        if network_alpha is None and "alpha" in block_keys:
+            alpha_tensor = handler.get_tensor(block_keys["alpha"])
+            network_alpha = float(alpha_tensor.item())
+
+        if network_dim is not None and network_alpha is not None:
+            break
+
+    # Default alpha to dim if not found
+    if network_alpha is None:
+        network_alpha = float(network_dim) if network_dim else 1.0
+    if network_dim is None:
+        network_dim = 1
+
+    return network_dim, network_alpha
+
+
+# =============================================================================
+# Merge Functions
+# =============================================================================
+
+def _merge_linear(lora_down: torch.Tensor, lora_up: torch.Tensor, device: str) -> torch.Tensor:
+    """Merge linear LoRA weights: lora_up @ lora_down."""
+    lora_down = lora_down.to(device=device, dtype=torch.float32)
+    lora_up = lora_up.to(device=device, dtype=torch.float32)
+    return lora_up @ lora_down
+
+
+def _merge_conv(lora_down: torch.Tensor, lora_up: torch.Tensor, device: str) -> torch.Tensor:
+    """Merge conv LoRA weights."""
+    in_rank, in_size, kernel_size, k_ = lora_down.shape
+    out_size, out_rank, _, _ = lora_up.shape
+    assert in_rank == out_rank, f"rank mismatch: {in_rank} vs {out_rank}"
+
+    lora_down = lora_down.to(device=device, dtype=torch.float32)
+    lora_up = lora_up.to(device=device, dtype=torch.float32)
+
+    merged = lora_up.reshape(out_size, -1) @ lora_down.reshape(in_rank, -1)
+    return merged.reshape(out_size, in_size, kernel_size, kernel_size)
+
+
+# =============================================================================
+# Extract Functions (after merge)
+# =============================================================================
+
+def _extract_linear(
+    weight: torch.Tensor,
+    max_rank: int,
+    dynamic_method: Optional[str],
+    dynamic_param: Optional[float],
+    scale: float,
+    niter: int = 2,
+) -> Dict:
+    """Extract LoRA from merged linear weight."""
+    out_size, in_size = weight.shape
+
+    if dynamic_method is None:
+        # Fixed rank - use svd_lowrank for 10x speedup
+        rank = min(max_rank, min(out_size, in_size) - 1)
+        U, S, Vh = torch.svd_lowrank(weight, q=rank, niter=niter)
+        Vh = Vh.T  # svd_lowrank returns V, not Vh
+        new_rank = rank
+        new_alpha = float(scale * new_rank)
+        stats = {"sum_retained": 1.0, "fro_retained": 1.0}  # Not computed for lowrank
+    else:
+        # Dynamic methods need full SVD to compute rank from all singular values
+        U, S, Vh = linalg.svd(weight, full_matrices=False)
+        new_rank, new_alpha, stats = _compute_resize(S, max_rank, dynamic_method, dynamic_param, scale)
+        U = U[:, :new_rank]
+        S = S[:new_rank]
+        Vh = Vh[:new_rank, :]
+
+    lora_up = U @ torch.diag(S)
+    lora_down = Vh
+
+    return {
+        "lora_down": lora_down.cpu().contiguous(),
+        "lora_up": lora_up.cpu().contiguous(),
+        "new_rank": new_rank,
+        "new_alpha": new_alpha,
+        **stats
+    }
+
+
+def _extract_conv(
+    weight: torch.Tensor,
+    max_rank: int,
+    dynamic_method: Optional[str],
+    dynamic_param: Optional[float],
+    scale: float,
+    niter: int = 2,
+) -> Dict:
+    """Extract LoRA from merged conv weight."""
+    out_ch, in_ch, kh, kw = weight.shape
+    mat = weight.reshape(out_ch, -1)
+
+    if dynamic_method is None:
+        # Fixed rank - use svd_lowrank for 10x speedup
+        rank = min(max_rank, min(mat.shape) - 1)
+        U, S, Vh = torch.svd_lowrank(mat, q=rank, niter=niter)
+        Vh = Vh.T  # svd_lowrank returns V, not Vh
+        new_rank = rank
+        new_alpha = float(scale * new_rank)
+        stats = {"sum_retained": 1.0, "fro_retained": 1.0}  # Not computed for lowrank
+    else:
+        # Dynamic methods need full SVD to compute rank from all singular values
+        U, S, Vh = linalg.svd(mat, full_matrices=False)
+        new_rank, new_alpha, stats = _compute_resize(S, max_rank, dynamic_method, dynamic_param, scale)
+        U = U[:, :new_rank]
+        S = S[:new_rank]
+        Vh = Vh[:new_rank, :]
+
+    lora_up = (U @ torch.diag(S)).reshape(out_ch, new_rank, 1, 1)
+    lora_down = Vh.reshape(new_rank, in_ch, kh, kw)
+
+    return {
+        "lora_down": lora_down.cpu().contiguous(),
+        "lora_up": lora_up.cpu().contiguous(),
+        "new_rank": new_rank,
+        "new_alpha": new_alpha,
+        **stats
+    }
+
+
+def _compute_resize(
+    S: torch.Tensor,
+    max_rank: int,
+    dynamic_method: Optional[str],
+    dynamic_param: Optional[float],
+    scale: float
+) -> Tuple[int, float, Dict]:
+    """Compute new rank and alpha based on resize method."""
+
+    if dynamic_method == "sv_ratio" and dynamic_param is not None:
+        # Use kohya convention: S[0]/ratio
+        min_sv = S[0] / dynamic_param
+        new_rank = max(1, int(torch.sum(S > min_sv).item()))
+    elif dynamic_method == "sv_cumulative" and dynamic_param is not None:
+        new_rank = _index_sv_cumulative(S, dynamic_param, max_rank)
+    elif dynamic_method == "sv_fro" and dynamic_param is not None:
+        new_rank = _index_sv_fro(S, dynamic_param, max_rank)
+    else:
+        new_rank = max_rank
+
+    # Clamp rank
+    if S[0] < MIN_SV:
+        new_rank = 1
+    else:
+        new_rank = max(1, min(new_rank, max_rank, len(S) - 1))
+
+    new_alpha = float(scale * new_rank)
+
+    # Compute retention stats
+    s_sum = float(torch.sum(torch.abs(S)))
+    s_rank = float(torch.sum(torch.abs(S[:new_rank]))) if new_rank <= len(S) else s_sum
+
+    S_sq = S.pow(2)
+    s_fro = float(torch.sqrt(torch.sum(S_sq)))
+    s_red_fro = float(torch.sqrt(torch.sum(S_sq[:new_rank]))) if new_rank <= len(S) else s_fro
+
+    stats = {
+        "sum_retained": s_rank / s_sum if s_sum > 0 else 1.0,
+        "fro_retained": s_red_fro / s_fro if s_fro > 0 else 1.0,
+    }
+
+    return new_rank, new_alpha, stats
+
+
+# =============================================================================
+# Main Resize Function
+# =============================================================================
+
+def resize_lora_file(
+    lora_path: str,
+    new_rank: int,
+    dynamic_method: Optional[str],
+    dynamic_param: Optional[float],
+    device: str,
+    save_dtype: torch.dtype,
+    output_filename: str,
+    verbose: bool = True,
+    svd_niter: int = 2,
+) -> str:
+    """
+    Resize a LoRA file to a new rank.
+
+    Args:
+        lora_path: Path to input LoRA
+        new_rank: Target rank (max rank for dynamic methods)
+        dynamic_method: None, "sv_ratio", "sv_fro", "sv_cumulative"
+        dynamic_param: Parameter for dynamic method
+        device: Processing device
+        save_dtype: Output dtype
+        output_filename: Output filename (without extension)
+        verbose: Print progress info
+        svd_niter: Power iterations for SVD accuracy
+
+    Returns:
+        Path to saved resized LoRA
+    """
+    # Prepare memory
+    lora_size_gb = estimate_model_size(lora_path)
+    prepare_for_large_operation(lora_size_gb * 2, torch.device(device))
+
+    handler = MemoryEfficientSafeOpen(lora_path)
+
+    try:
+        metadata = handler.metadata().copy()
+        all_keys = handler.keys()
+
+        # Detect format and extract pairs
+        format_info = detect_lora_format(all_keys)
+        pairs = extract_lora_pairs(all_keys, format_info)
+        network_dim, network_alpha = detect_lora_rank(handler, pairs)
+
+        scale = network_alpha / network_dim if network_dim > 0 else 1.0
+
+        if verbose:
+            method_str = f"{dynamic_method}: {dynamic_param}" if dynamic_method else "fixed"
+            print(f"[LoRA Resize] Format: {format_info['format']}, layers: {format_info['key_count']}")
+            print(f"[LoRA Resize] Original dim={network_dim}, alpha={network_alpha:.1f}, scale={scale:.3f}")
+            print(f"[LoRA Resize] Resizing with method={method_str}, max_rank={new_rank}")
+
+        output_sd = {}
+        fro_list = []
+        pbar = comfy.utils.ProgressBar(len(pairs))
+
+        with torch.no_grad():
+            for block_name, block_keys in tqdm(pairs.items(), desc="Resizing layers", unit="layers"):
+                if "down" not in block_keys or "up" not in block_keys:
+                    pbar.update(1)
+                    continue
+
+                lora_down = handler.get_tensor(block_keys["down"])
+                lora_up = handler.get_tensor(block_keys["up"])
+
+                is_conv = len(lora_down.shape) == 4
+
+                # Transfer to GPU with pinned memory if available
+                if device == 'cuda':
+                    lora_down = transfer_to_gpu_pinned(lora_down, device, torch.float32)
+                    lora_up = transfer_to_gpu_pinned(lora_up, device, torch.float32)
+
+                # Merge and re-extract
+                if is_conv:
+                    weight = _merge_conv(lora_down, lora_up, device)
+                    result = _extract_conv(weight, new_rank, dynamic_method, dynamic_param, scale, svd_niter)
+                else:
+                    weight = _merge_linear(lora_down, lora_up, device)
+                    result = _extract_linear(weight, new_rank, dynamic_method, dynamic_param, scale, svd_niter)
+
+                del weight, lora_down, lora_up
+
+                fro_list.append(result['fro_retained'])
+
+                # Store using same format suffixes as input
+                down_suffix = format_info["down_suffix"]
+                up_suffix = format_info["up_suffix"]
+                alpha_suffix = format_info["alpha_suffix"]
+
+                # Standardize prefix using heuristics if possible
+                new_block_name = _format_lora_key(block_name)
+
+                output_sd[f"{new_block_name}{down_suffix}"] = result["lora_down"].to(save_dtype)
+                output_sd[f"{new_block_name}{up_suffix}"] = result["lora_up"].to(save_dtype)
+                if alpha_suffix:
+                    output_sd[f"{new_block_name}{alpha_suffix}"] = torch.tensor(result["new_alpha"]).to(save_dtype)
+
+                pbar.update(1)
+
+        if verbose and fro_list:
+            import numpy as np
+            avg_fro = np.mean(fro_list)
+            std_fro = np.std(fro_list)
+            print(f"[LoRA Resize] Average Frobenius retention: {avg_fro:.1%} ± {std_fro:.3f}")
+
+        # Update metadata
+        if dynamic_method:
+            metadata["ss_training_comment"] = f"Dynamic resize with {dynamic_method}: {dynamic_param} from dim {network_dim}"
+            metadata["ss_network_dim"] = "Dynamic"
+            metadata["ss_network_alpha"] = "Dynamic"
+        else:
+            metadata["ss_training_comment"] = f"Resized from dim {network_dim} to {new_rank}"
+            metadata["ss_network_dim"] = str(new_rank)
+            metadata["ss_network_alpha"] = str(scale * new_rank)
+
+        # Save
+        output_dir = folder_paths.get_folder_paths("loras")[0]
+        os.makedirs(output_dir, exist_ok=True)
+        output_path = os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
+
+        save_file(output_sd, output_path, metadata)
+        print(f"[LoRA Resize] Saved to {output_path}")
+
+        return output_path
+
+    finally:
+        handler.__exit__(None, None, None)
+        cleanup_after_operation()
+
+
+# =============================================================================
+# Node Definitions
+# =============================================================================
+
+class LoRAResizeFixed(io.ComfyNode):
+    """Resize LoRA to a fixed rank."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAResizeFixed",
+            display_name="LoRA Resize (Fixed Rank)",
+            category="ModelUtils/LoRA/Resize",
+            description="Resize existing LoRA to a specific rank by merging and re-extracting via SVD.",
+            inputs=[
+                io.Combo.Input("lora_name", options=folder_paths.get_filename_list("loras"),
+                              tooltip="LoRA to resize"),
+                io.Int.Input("new_rank", default=64, min=1, max=3072,
+                            tooltip="Target rank"),
+                io.Int.Input("svd_niter", default=2, min=0, max=10,
+                            tooltip="SVD power iterations (higher = more accurate but slower)"),
+                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"),
+            ],
+            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:
+        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)
+        return io.NodeOutput(path)
+
+
+class LoRAResizeRatio(io.ComfyNode):
+    """Resize LoRA keeping SVs above ratio threshold."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAResizeRatio",
+            display_name="LoRA Resize (SV Ratio)",
+            category="ModelUtils/LoRA/Resize",
+            description="Dynamically resize LoRA, keeping singular values where S[i] > S[0]/ratio.",
+            inputs=[
+                io.Combo.Input("lora_name", options=folder_paths.get_filename_list("loras"),
+                              tooltip="LoRA to resize"),
+                io.Int.Input("max_rank", default=128, min=1, max=3072,
+                            tooltip="Maximum allowed rank"),
+                io.Float.Input("ratio", default=2.0, min=1.0, max=100.0, step=0.1,
+                              tooltip="Keep SVs where S[i] > S[0]/ratio"),
+                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"),
+            ],
+            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:
+        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)
+        return io.NodeOutput(path)
+
+
+class LoRAResizeFrobenius(io.ComfyNode):
+    """Resize LoRA to preserve Frobenius norm target."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAResizeFrobenius",
+            display_name="LoRA Resize (Frobenius)",
+            category="ModelUtils/LoRA/Resize",
+            description="Dynamically resize LoRA to preserve target fraction of Frobenius norm.",
+            inputs=[
+                io.Combo.Input("lora_name", options=folder_paths.get_filename_list("loras"),
+                              tooltip="LoRA to resize"),
+                io.Int.Input("max_rank", default=128, min=1, max=3072,
+                            tooltip="Maximum allowed rank"),
+                io.Float.Input("target", default=0.9, min=0.1, max=1.0, step=0.01,
+                              tooltip="Target Frobenius norm retention (0.9 = 90%)"),
+                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"),
+            ],
+            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:
+        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)
+        return io.NodeOutput(path)
+
+
+class LoRAResizeCumulative(io.ComfyNode):
+    """Resize LoRA to preserve cumulative SV target."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAResizeCumulative",
+            display_name="LoRA Resize (Cumulative)",
+            category="ModelUtils/LoRA/Resize",
+            description="Dynamically resize LoRA to preserve target fraction of cumulative singular values.",
+            inputs=[
+                io.Combo.Input("lora_name", options=folder_paths.get_filename_list("loras"),
+                              tooltip="LoRA to resize"),
+                io.Int.Input("max_rank", default=128, min=1, max=3072,
+                            tooltip="Maximum allowed rank"),
+                io.Float.Input("target", default=0.9, min=0.1, max=1.0, step=0.01,
+                              tooltip="Target cumulative SV retention (0.9 = 90%)"),
+                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"),
+            ],
+            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:
+        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)
+        return io.NodeOutput(path)
+
+
+
+
+# =============================================================================
+# LoRA Merge To Model (Save merged model, skip extraction)
+# =============================================================================
+
+def merge_loras_to_model(
+    lora_paths: List[str],
+    lora_weights: List[float],
+    base_model_path: str,
+    device: str,
+    save_dtype: torch.dtype,
+    output_filename: str,
+    skip_patterns_str: str = "",
+    verbose: bool = True,
+) -> str:
+    """
+    Merge multiple LoRAs into a base model and save the result directly.
+
+    Unlike merge_multi_loras_via_base, this function:
+    - Does NOT extract the result back to a LoRA
+    - Saves the merged full model to the base model's directory
+
+    Args:
+        lora_paths: List of paths to LoRA files
+        lora_weights: List of weight strengths (0.0-2.0) for each LoRA
+        base_model_path: Path to base model
+        device: Processing device
+        save_dtype: Output dtype
+        output_filename: Output filename (without extension)
+        skip_patterns_str: Regex patterns for layers to skip
+        verbose: Print progress info
+
+    Returns:
+        Path to saved merged model
+    """
+    # Estimate memory and prepare
+    total_size_gb = estimate_model_size(base_model_path)
+    for lp in lora_paths:
+        total_size_gb += estimate_model_size(lp)
+
+    if verbose:
+        print(f"[LoRA Merge To Model] Preparing memory for {total_size_gb:.2f}GB operation...")
+        print(f"[LoRA Merge To Model] Merging {len(lora_paths)} LoRAs with weights: {lora_weights}")
+    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]
+
+
+    try:
+        # Detect format and extract pairs for each LoRA
+        lora_infos = []
+
+        for i, handler in enumerate(lora_handlers):
+            keys = handler.keys()
+            format_info = detect_lora_format(keys)
+            pairs = extract_lora_pairs(keys, format_info)
+            network_dim, network_alpha = detect_lora_rank(handler, pairs)
+            lora_infos.append({
+                "handler": handler,
+                "format_info": format_info,
+                "pairs": pairs,
+                "network_dim": network_dim,
+                "network_alpha": network_alpha,
+                "weight": lora_weights[i],
+            })
+            if verbose:
+                print(f"[LoRA Merge To Model] LoRA {i+1}: {format_info['format']}, {len(pairs)} layers, dim={network_dim}")
+
+        # Common prefixes
+        BASE_PREFIXES = ["model.diffusion_model.", "diffusion_model.", "transformer.", "model."]
+        LORA_PREFIXES = [
+            "lora_unet_", "lora_transformer_", "lora_te1_", "lora_te2_", "lora_te_",
+            "lycoris_", "diffusion_model.", "transformer.", "unet."
+        ]
+
+        def extract_core_layer_base(key: str) -> str:
+            result = key
+            if result.endswith(".weight"):
+                result = result[:-7]
+            elif result.endswith(".bias"):
+                result = result[:-5]
+            for prefix in BASE_PREFIXES:
+                if result.startswith(prefix):
+                    result = result[len(prefix):]
+                    break
+            return result
+
+        def extract_core_layer_lora(block_name: str) -> str:
+            result = block_name
+            for prefix in LORA_PREFIXES:
+                if result.startswith(prefix):
+                    result = result[len(prefix):]
+                    break
+            # Strip trailing .lora suffix (HuggingFace format leaves this after suffix removal)
+            if result.endswith(".lora"):
+                result = result[:-5]
+            return result.replace(".", "_")
+
+
+        # Build LoRA lookup: core layer name (underscored) -> list of (info, block_keys)
+        lora_lookup = {}
+        for info in lora_infos:
+            for block_name, block_keys in info["pairs"].items():
+                core = extract_core_layer_lora(block_name)
+                if core not in lora_lookup:
+                    lora_lookup[core] = []
+                lora_lookup[core].append((info, block_keys))
+
+        # Compile skip patterns
+        skip_patterns = _compile_patterns(skip_patterns_str)
+
+        # Preserve metadata from base model
+        base_metadata = base_handler.metadata().copy() if base_handler.metadata() else {}
+        base_metadata["merge_comment"] = f"Merged {len(lora_paths)} LoRAs with weights: {lora_weights}"
+
+        output_sd = {}
+        stats = {"merged": 0, "copied": 0, "skipped": 0}
+        base_keys = list(base_handler.keys())
+        pbar = comfy.utils.ProgressBar(len(base_keys))
+
+        if verbose:
+            print(f"[LoRA Merge To Model] Processing {len(base_keys)} base model keys...")
+
+        with torch.no_grad():
+            for base_key in tqdm(base_keys, desc="Merging to model", unit="keys"):
+                # Check skip patterns
+                if _matches_any_pattern(base_key, skip_patterns):
+                    stats["skipped"] += 1
+                    pbar.update(1)
+                    continue
+
+                # Load base weight
+                cpu_base = base_handler.get_tensor(base_key)
+
+                # Only process weight tensors for LoRA merging
+                if base_key.endswith(".weight"):
+                    core = extract_core_layer_base(base_key)
+                    core_underscored = core.replace(".", "_")
+
+                    # Check if any LoRA contributes to this layer
+                    if core_underscored in lora_lookup:
+                        # Transfer to GPU for computation
+                        if device == 'cuda':
+                            base_weight = transfer_to_gpu_pinned(cpu_base, device, torch.float32)
+                        else:
+                            base_weight = cpu_base.to(device=device, dtype=torch.float32)
+                        del cpu_base
+
+                        # Accumulate deltas from all contributing LoRAs
+                        for info, block_keys in lora_lookup[core_underscored]:
+                            is_full_diff = info["format_info"].get("is_full_diff", False)
+
+                            if is_full_diff:
+                                # Full diff format
+                                if "diff" not in block_keys:
+                                    continue
+                                cpu_diff = info["handler"].get_tensor(block_keys["diff"])
+                                if device == 'cuda':
+                                    delta = transfer_to_gpu_pinned(cpu_diff, device, torch.float32)
+                                else:
+                                    delta = cpu_diff.to(device=device, dtype=torch.float32)
+                                del cpu_diff
+                                effective_scale = info["weight"]
+                            else:
+                                # Standard LoRA format
+                                if "down" not in block_keys or "up" not in block_keys:
+                                    continue
+
+                                cpu_down = info["handler"].get_tensor(block_keys["down"])
+                                cpu_up = info["handler"].get_tensor(block_keys["up"])
+                                if device == 'cuda':
+                                    lora_down = transfer_to_gpu_pinned(cpu_down, device, torch.float32)
+                                    lora_up = transfer_to_gpu_pinned(cpu_up, device, torch.float32)
+                                else:
+                                    lora_down = cpu_down.to(device=device, dtype=torch.float32)
+                                    lora_up = cpu_up.to(device=device, dtype=torch.float32)
+                                del cpu_down, cpu_up
+
+                                # Get alpha
+                                if "alpha" in block_keys:
+                                    alpha_tensor = info["handler"].get_tensor(block_keys["alpha"])
+                                    layer_alpha = float(alpha_tensor.item())
+                                else:
+                                    layer_alpha = float(info["network_dim"])
+                                layer_scale = layer_alpha / info["network_dim"] if info["network_dim"] > 0 else 1.0
+                                effective_scale = layer_scale * info["weight"]
+
+                                # Compute delta
+                                is_conv = len(lora_down.shape) == 4
+                                if is_conv:
+                                    in_rank, in_size, kernel_size, k_ = lora_down.shape
+                                    out_size, out_rank, _, _ = lora_up.shape
+                                    delta = lora_up.reshape(out_size, -1) @ lora_down.reshape(in_rank, -1)
+                                    delta = delta.reshape(out_size, in_size, kernel_size, kernel_size)
+                                else:
+                                    delta = lora_up @ lora_down
+                                del lora_down, lora_up
+
+                            # Apply delta to base weight
+                            base_weight = base_weight + effective_scale * delta
+                            del delta
+
+                        # Store merged weight
+                        output_sd[base_key] = base_weight.to(save_dtype).cpu().contiguous()
+                        del base_weight
+                        stats["merged"] += 1
+                    else:
+                        # No LoRA contribution, copy as-is
+                        output_sd[base_key] = cpu_base.to(save_dtype).contiguous()
+                        del cpu_base
+                        stats["copied"] += 1
+                else:
+                    # Non-weight tensor (bias, norm, etc.), copy as-is
+                    output_sd[base_key] = cpu_base.to(save_dtype).contiguous()
+                    del cpu_base
+                    stats["copied"] += 1
+
+
+                pbar.update(1)
+
+        if verbose:
+            print(f"[LoRA Merge To Model] Done: {stats['merged']} merged, {stats['copied']} copied, {stats['skipped']} skipped")
+
+        # Save to base model directory
+        base_dir = os.path.dirname(base_model_path)
+        os.makedirs(base_dir, exist_ok=True)
+        output_path = os.path.join(base_dir, f"{output_filename.strip()}.safetensors")
+
+        save_file(output_sd, output_path, base_metadata)
+        print(f"[LoRA Merge To Model] Saved to {output_path}")
+
+        return output_path
+
+    finally:
+        base_handler.__exit__(None, None, None)
+        for handler in lora_handlers:
+            handler.__exit__(None, None, None)
+        cleanup_after_operation()
+
+
+class LoRAMergeToModel(io.ComfyNode):
+    """Merge multiple LoRAs into base model and save as full model."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRAMergeToModel",
+            display_name="LoRA Merge To Model",
+            category="ModelUtils/LoRA/Merge",
+            description="Merge 1-4 LoRAs into a base model and save the result. Saves to base model directory.",
+            inputs=[
+                io.Combo.Input("base_model", options=folder_paths.get_filename_list("diffusion_models"),
+                              tooltip="Base model the LoRAs were trained on"),
+                io.Combo.Input("lora_count", options=["1", "2", "3", "4"], default="2",
+                              tooltip="Number of LoRAs to merge"),
+                # LoRA 1
+                io.Combo.Input("lora_1", options=folder_paths.get_filename_list("loras"),
+                              tooltip="First LoRA"),
+                io.Float.Input("weight_1", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 1"),
+                # LoRA 2
+                io.Combo.Input("lora_2", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Second LoRA"),
+                io.Float.Input("weight_2", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 2"),
+                # LoRA 3
+                io.Combo.Input("lora_3", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Third LoRA"),
+                io.Float.Input("weight_3", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 3"),
+                # LoRA 4
+                io.Combo.Input("lora_4", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Fourth LoRA"),
+                io.Float.Input("weight_4", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 4"),
+                # Settings
+                io.String.Input("skip_patterns", default="", multiline=True,
+                               tooltip="Regex patterns for layers to skip"),
+                io.String.Input("output_filename", default="merged_model"),
+                io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
+                io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, base_model, lora_count,
+                lora_1, weight_1, lora_2, weight_2, lora_3, weight_3, lora_4, weight_4,
+                skip_patterns, output_filename, save_dtype, device) -> io.NodeOutput:
+
+        # Build LoRA list based on count
+        count = int(lora_count)
+        lora_names = [lora_1, lora_2, lora_3, lora_4][:count]
+        lora_weights = [weight_1, weight_2, weight_3, weight_4][:count]
+
+        # Filter out "None" entries
+        valid_loras = [(name, weight) for name, weight in zip(lora_names, lora_weights) if name != "None"]
+        if not valid_loras:
+            raise ValueError("At least one LoRA must be selected")
+
+        lora_names, lora_weights = zip(*valid_loras)
+        lora_paths = [folder_paths.get_full_path_or_raise("loras", name) for name in lora_names]
+        base_path = folder_paths.get_full_path_or_raise("diffusion_models", base_model)
+        dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
+
+        path = merge_loras_to_model(
+            lora_paths=list(lora_paths),
+            lora_weights=list(lora_weights),
+            base_model_path=base_path,
+            device=device,
+            save_dtype=dtype,
+            output_filename=output_filename,
+            skip_patterns_str=skip_patterns,
+        )
+        return io.NodeOutput(path)

nodes/merger.py

@@ -3,10 +3,14 @@ import re
 import torch
 import folder_paths
 import comfy.utils
+from tqdm import tqdm
 from comfy_api.latest import io
 from safetensors.torch import save_file
 from .merger_ops import TWO_MODEL_MODES, THREE_MODEL_MODES, MissingTensorBehavior, MissingTensorError
-from .merger_utils import MemoryEfficientSafeOpen
+from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
+from .device_utils import (
+    estimate_model_size, prepare_for_large_operation, cleanup_after_operation
+)
 
 
 def load_documentation_from_file(filename):
@@ -22,13 +26,13 @@ def load_documentation_from_file(filename):
 
 def _compile_patterns(pattern_string):
     """Compiles whitespace-separated regex patterns into a list of compiled patterns.
-    
+
     Returns empty list if pattern_string is empty or whitespace-only.
     Raises ValueError if any pattern is invalid regex.
     """
     if not pattern_string or not pattern_string.strip():
         return []
-    
+
     patterns = []
     for pattern in pattern_string.split():
         try:
@@ -48,7 +52,7 @@ def _matches_any_pattern(key, patterns):
 
 class MergerLogic:
     """Shared merge execution logic for all merger nodes."""
-    
+
     @staticmethod
     def execute_merge(model_names, calc_mode, all_modes, recipe_params, model_type):
         calc_mode_class = next((m for m in all_modes if m.name == calc_mode), None)
@@ -57,7 +61,22 @@ class MergerLogic:
         primary_model_name = model_names.get('model_a')
         if not primary_model_name or primary_model_name == "None":
             raise ValueError("Model A is required to run the merge.")
-        
+
+        # Prepare memory before heavy operation
+        total_size_gb = 0
+        model_paths = []
+        for name in model_names.values():
+            if name and name != "None":
+                path = folder_paths.get_full_path(model_type, name)
+                if path:
+                    total_size_gb += estimate_model_size(path)
+                    model_paths.append(path)
+
+        if total_size_gb > 0:
+            process_device = recipe_params.get('device', 'cpu')
+            print(f"[Merger] Preparing memory for {total_size_gb:.2f}GB merge operation...")
+            prepare_for_large_operation(total_size_gb * 1.2, torch.device(process_device))
+
         handlers = {}
         for name in model_names.values():
             if name and name != "None":
@@ -65,20 +84,21 @@ class MergerLogic:
                 if not path:
                     raise FileNotFoundError(f"Model '{name}' not found.")
                 handlers[name] = MemoryEfficientSafeOpen(path)
-        
+
         primary_handler = handlers[primary_model_name]
         all_keys = primary_handler.keys()
         metadata = primary_handler.header.get("__metadata__", {})
-        
+
         # Convert mismatch_mode string to enum
         mismatch_mode_str = recipe_params.get('mismatch_mode', 'skip')
         mismatch_mode = MissingTensorBehavior(mismatch_mode_str)
         recipe_params['mismatch_mode'] = mismatch_mode
-        
+        alignment_mode = recipe_params.get('alignment_mode', 'pad/crop')
+
         # Compile filter patterns
         exclude_patterns = _compile_patterns(recipe_params.get('exclude_patterns', ''))
         discard_patterns = _compile_patterns(recipe_params.get('discard_patterns', ''))
-        
+
         # Pre-compute key differences for logging
         primary_keys = set(all_keys)
         for name, handler in handlers.items():
@@ -90,7 +110,7 @@ class MergerLogic:
                     print(f"[Merger] {name} is missing {len(missing)} keys present in Model A")
                 if extra:
                     print(f"[Merger] {name} has {len(extra)} extra keys not in Model A (ignored)")
-        
+
         merged_state_dict = {}
         pbar = comfy.utils.ProgressBar(len(all_keys))
         save_dtype = recipe_params.pop('save_dtype')
@@ -98,34 +118,39 @@ class MergerLogic:
         recipe_params.update({"handlers": handlers})
         process_device = recipe_params.get('device', 'cpu')
         process_dtype = recipe_params.get('dtype', torch.float32)
-        
+
         skipped_keys = 0
         excluded_keys = 0
         discarded_keys = 0
         error_keys = []
-        
-        for key in all_keys:
+
+        for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
             # Check discard patterns first - skip entirely
             if _matches_any_pattern(key, discard_patterns):
                 discarded_keys += 1
                 pbar.update(1)
                 continue
-            
-            # Pre-load Model A's tensor
-            tensor_a = primary_handler.get_tensor(key).to(device=process_device, dtype=process_dtype)
-            
+
+            # Pre-load Model A's tensor with pinned memory for CUDA
+            cpu_tensor = primary_handler.get_tensor(key)
+            if process_device == 'cuda':
+                tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
+            else:
+                tensor_a = cpu_tensor.to(device=process_device, dtype=process_dtype)
+            del cpu_tensor
+
             # Check exclude patterns - use Model A only, no merge
             if _matches_any_pattern(key, exclude_patterns):
                 merged_state_dict[key] = tensor_a.to(save_torch_dtype).cpu().clone()
                 excluded_keys += 1
                 pbar.update(1)
                 continue
-            
+
             # Pass tensor_a metadata to recipes for zeros mode and fallback
             recipe_params['_tensor_a'] = tensor_a
             recipe_params['_tensor_a_shape'] = tensor_a.shape
             recipe_params['_tensor_a_dtype'] = tensor_a.dtype
-            
+
             try:
                 recipe = calc_mode_class.create_recipe(key=key, **recipe_params)
                 result = recipe.merge()
@@ -134,22 +159,31 @@ class MergerLogic:
                     raise ValueError(f"Layer mismatch error (mismatch_mode='error'): {e}")
                 result = None
                 error_keys.append(key)
-            
+
             # Handle None result (mismatch occurred with skip mode)
             if result is None:
                 result = tensor_a
                 skipped_keys += 1
-            
+
             if isinstance(result, dict):
                 for r_key, r_tensor in result.items():
                     merged_state_dict[r_key] = r_tensor.to(save_torch_dtype).cpu().clone()
             else:
+                # Ensure compatibility with Model A's architecture.
+                # If alignment_mode is 'pad/crop', we crop results that were padded.
+                # If alignment_mode is 'interpolate', resizing happened during operators.
+                if alignment_mode == 'pad/crop':
+                    target_shape = recipe_params['_tensor_a_shape']
+                    if result.shape != target_shape:
+                        slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(result.shape, target_shape))
+                        result = result[slices]
+
                 merged_state_dict[key] = result.to(save_torch_dtype).cpu().clone()
-            
+
             # Clean up tensor_a reference to allow GC
             del recipe_params['_tensor_a']
             pbar.update(1)
-        
+
         # Log summary
         if excluded_keys > 0:
             print(f"[Merger] Excluded {excluded_keys} keys from merge (kept Model A only)")
@@ -159,15 +193,22 @@ class MergerLogic:
             print(f"[Merger] Used Model A's values for {skipped_keys} keys due to mismatches")
         if error_keys:
             print(f"[Merger] Unexpected errors on {len(error_keys)} keys: {error_keys[:5]}{'...' if len(error_keys) > 5 else ''}")
-        
+
         for handler in handlers.values():
             handler.__exit__(None, None, None)
+
         output_folder = "loras" if calc_mode == "SVD LoRA Extraction" else model_type
-        output_dir = folder_paths.get_folder_paths(output_folder)[0]
+        # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
+        output_dir = folder_paths.get_folder_paths(output_folder)[-1]
         os.makedirs(output_dir, exist_ok=True)
         output_filename = recipe_params.get("output_filename")
         output_path = os.path.join(output_dir, f"{output_filename}.safetensors")
         save_file(merged_state_dict, output_path, metadata=metadata)
+
+        # Cleanup after heavy operation
+        del merged_state_dict
+        cleanup_after_operation()
+
         return f"{output_filename}.safetensors"
 
 
@@ -188,9 +229,11 @@ class CheckpointTwoMerger(io.ComfyNode):
                 io.Combo.Input("model_b", options=["None"] + folder_paths.get_filename_list("checkpoints")),
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_checkpoint"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -206,8 +249,8 @@ class CheckpointTwoMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
-                gamma: float, seed: int, output_filename: str, save_dtype: str,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
+                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
@@ -215,8 +258,8 @@ class CheckpointTwoMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "seed": seed,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
@@ -241,9 +284,11 @@ class ModelTwoMerger(io.ComfyNode):
                 io.Combo.Input("model_b", options=["None"] + folder_paths.get_filename_list("diffusion_models")),
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_model"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -259,8 +304,8 @@ class ModelTwoMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
-                gamma: float, seed: int, output_filename: str, save_dtype: str,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
+                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
@@ -268,8 +313,8 @@ class ModelTwoMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "seed": seed,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
@@ -294,9 +339,11 @@ class TextEncoderTwoMerger(io.ComfyNode):
                 io.Combo.Input("model_b", options=["None"] + folder_paths.get_filename_list("text_encoders")),
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_textencoder"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -312,8 +359,8 @@ class TextEncoderTwoMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
-                gamma: float, seed: int, output_filename: str, save_dtype: str,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
+                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
@@ -321,8 +368,8 @@ class TextEncoderTwoMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "seed": seed,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
@@ -347,9 +394,11 @@ class LoRATwoMerger(io.ComfyNode):
                 io.Combo.Input("model_b", options=["None"] + folder_paths.get_filename_list("loras")),
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_lora"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -365,8 +414,8 @@ class LoRATwoMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
-                gamma: float, seed: int, output_filename: str, save_dtype: str,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
+                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
@@ -374,8 +423,8 @@ class LoRATwoMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "seed": seed,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
@@ -400,9 +449,11 @@ class EmbeddingTwoMerger(io.ComfyNode):
                 io.Combo.Input("model_b", options=["None"] + folder_paths.get_filename_list("embeddings")),
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_embedding"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -418,8 +469,8 @@ class EmbeddingTwoMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
-                gamma: float, seed: int, output_filename: str, save_dtype: str,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
+                gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
@@ -427,8 +478,8 @@ class EmbeddingTwoMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
-            "alpha": alpha, "beta": beta, "gamma": gamma, "seed": seed,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
+            "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
@@ -456,6 +507,7 @@ class CheckpointThreeMerger(io.ComfyNode):
                 io.Combo.Input("model_c", options=["None"] + folder_paths.get_filename_list("checkpoints")),
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
@@ -475,7 +527,7 @@ class CheckpointThreeMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
@@ -484,7 +536,7 @@ class CheckpointThreeMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
             "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
@@ -511,6 +563,7 @@ class ModelThreeMerger(io.ComfyNode):
                 io.Combo.Input("model_c", options=["None"] + folder_paths.get_filename_list("diffusion_models")),
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
@@ -530,7 +583,7 @@ class ModelThreeMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
@@ -539,7 +592,7 @@ class ModelThreeMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
             "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
@@ -566,6 +619,7 @@ class TextEncoderThreeMerger(io.ComfyNode):
                 io.Combo.Input("model_c", options=["None"] + folder_paths.get_filename_list("text_encoders")),
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
@@ -585,7 +639,7 @@ class TextEncoderThreeMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
@@ -594,7 +648,7 @@ class TextEncoderThreeMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
             "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
@@ -621,6 +675,7 @@ class LoRAThreeMerger(io.ComfyNode):
                 io.Combo.Input("model_c", options=["None"] + folder_paths.get_filename_list("loras")),
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
@@ -640,7 +695,7 @@ class LoRAThreeMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
@@ -649,7 +704,7 @@ class LoRAThreeMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
             "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
@@ -676,6 +731,7 @@ class EmbeddingThreeMerger(io.ComfyNode):
                 io.Combo.Input("model_c", options=["None"] + folder_paths.get_filename_list("embeddings")),
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
+                io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
                 io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
                 io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
@@ -695,7 +751,7 @@ class EmbeddingThreeMerger(io.ComfyNode):
 
     @classmethod
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
-                calc_mode: str, mismatch_mode: str, alpha: float, beta: float,
+                calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
                 process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
@@ -704,7 +760,7 @@ class EmbeddingThreeMerger(io.ComfyNode):
 
         recipe_params = {
             "model_a": model_a, "model_b": model_b, "model_c": model_c, "calc_mode": calc_mode,
-            "mismatch_mode": mismatch_mode,
+            "mismatch_mode": mismatch_mode, "alignment_mode": alignment_mode,
             "alpha": alpha, "beta": beta, "gamma": gamma, "delta": delta, "seed": seed,
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,

nodes/merger_ops.py

@@ -15,6 +15,92 @@ class MissingTensorError(Exception):
     """Raised when a tensor is missing and behavior is ERROR."""
     pass
 
+# ################# #
+# UTILITIES
+# ################# #
+def resize_by_interpolation(tensor1, tensor2):
+    """Resizes tensor2 to match tensor1's shape using interpolation."""
+    if tensor1 is None or tensor2 is None:
+        return tensor1, tensor2
+    if tensor1.shape == tensor2.shape:
+        return tensor1, tensor2
+
+    dims1 = len(tensor1.shape)
+    dims2 = len(tensor2.shape)
+
+    if dims1 != dims2 or dims1 > 4 or dims1 == 0:
+        return tensor1, tensor2
+
+    orig_dtype = tensor2.dtype
+    t2_f = tensor2.float()
+
+    # F.interpolate expects [B, C, ...]
+    # We treat the tensor as [1, 1, ...dims]
+    view_shape = [1, 1] + list(tensor2.shape)
+    t2_f = t2_f.view(view_shape)
+
+    # Select mode based on dimensions
+    if dims1 == 4:
+        # [Out, In, H, W] -> [1, 1, Out, In, H*W] (5D)
+        view_shape = [1, 1, tensor2.shape[0], tensor2.shape[1], tensor2.shape[2] * tensor2.shape[3]]
+        target_size = [tensor1.shape[0], tensor1.shape[1], tensor1.shape[2] * tensor1.shape[3]]
+        t2_f = t2_f.view(view_shape)
+        t2_res = F.interpolate(t2_f, size=target_size, mode='trilinear', align_corners=False)
+        return tensor1, t2_res.view(tensor1.shape).to(orig_dtype)
+
+    if dims1 == 1:
+        mode = 'linear'
+    elif dims1 == 2:
+        mode = 'bilinear'
+    else:
+        mode = 'trilinear'
+
+    t2_res = F.interpolate(t2_f, size=list(tensor1.shape), mode=mode, align_corners=False)
+
+    return tensor1, t2_res.view(tensor1.shape).to(orig_dtype)
+
+
+def resize_tensors(tensor1, tensor2, mode='pad/crop'):
+    """Resizes tensors to match shapes using either padding or interpolation."""
+    if tensor1 is None or tensor2 is None:
+        return tensor1, tensor2
+    if tensor1.shape == tensor2.shape:
+        return tensor1, tensor2
+
+    if mode == 'interpolate':
+        return resize_by_interpolation(tensor1, tensor2)
+
+    dims1 = len(tensor1.shape)
+    dims2 = len(tensor2.shape)
+
+    if dims1 != dims2 or dims1 > 4 or dims1 == 0:
+        return tensor1, tensor2
+
+    padding1 = []
+    padding2 = []
+
+    # F.pad expects: (last_dim_left, last_dim_right, prev_dim_top, prev_dim_bottom, ...)
+    for i in range(dims1 - 1, -1, -1):
+        s1 = tensor1.shape[i]
+        s2 = tensor2.shape[i]
+        if s1 < s2:
+            padding1.extend([0, s2 - s1])
+            padding2.extend([0, 0])
+        elif s2 < s1:
+            padding1.extend([0, 0])
+            padding2.extend([0, s1 - s2])
+        else:
+            padding1.extend([0, 0])
+            padding2.extend([0, 0])
+
+    if any(p > 0 for p in padding1):
+        tensor1 = F.pad(tensor1, tuple(padding1))
+    if any(p > 0 for p in padding2):
+        tensor2 = F.pad(tensor2, tuple(padding2))
+
+    return tensor1, tensor2
+
+
 # ################# #
 # MERGE OPERATORS
 # ################# #
@@ -23,6 +109,7 @@ class Operation:
         self.key = key
         self.sources = tuple(sources)
         self.merge_func = self.recurse
+        self.alignment_mode = 'pad/crop'
         self.alpha = None
         self.beta = None
         self.gamma = None
@@ -61,7 +148,7 @@ class LoadTensor(Operation):
     def merge(self) -> torch.Tensor:
         if self.model_name not in self.handlers:
             raise ValueError(f"Model '{self.model_name}' is required for this mode but was not provided.")
-        
+
         handler = self.handlers[self.model_name]
         if self.key not in handler.keys():
             if self.on_missing == MissingTensorBehavior.ERROR:
@@ -77,7 +164,7 @@ class LoadTensor(Operation):
                     )
                 dtype = self.fallback_dtype if self.fallback_dtype else self.dtype
                 return torch.zeros(self.fallback_shape, device=self.device, dtype=dtype)
-        
+
         return handler.get_tensor(self.key).to(device=self.device, dtype=self.dtype)
 
 class Multiply(Operation):
@@ -96,6 +183,7 @@ class Add(Operation):
             return b
         if b is None:
             return a
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
         return a + b
 
 
@@ -103,15 +191,33 @@ class Sub(Operation):
     def oper(self, a, b) -> torch.Tensor:
         if a is None or b is None:
             return None  # Cannot compute difference with missing operand
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
         return a - b
 
 class SVD(Operation):
-    def __init__(self, key, alpha, beta, gamma, *sources):
+    def __init__(self, key, alpha, beta, gamma, *sources, target_shape=None):
         super().__init__(key, *sources)
         self.alpha, self.beta, self.gamma = int(alpha), int(beta), gamma
+        self.target_shape = target_shape
+
     def oper(self, a, b) -> torch.Tensor:
         if a is None or b is None:
             return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+
+        # For SVD extraction, we must ensure the diff matches the original Model A shape
+        # so the resulting LoRA weights are compatible with the target architecture.
+        if self.target_shape and a.shape != self.target_shape:
+            if self.alignment_mode == 'interpolate':
+                # Re-align both to match target_shape via interpolation
+                dummy = torch.empty(self.target_shape, device=a.device)
+                _, a = resize_by_interpolation(dummy, a)
+                _, b = resize_by_interpolation(dummy, b)
+            else:
+                slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(a.shape, self.target_shape))
+                a = a[slices]
+                b = b[slices]
+
         diff, weights, conv2d = a - b, {}, len(a.size()) == 4
         kernel_size = None if not conv2d else a.size()[2:4]
         conv2d_3x3 = conv2d and kernel_size != (1, 1)
@@ -141,6 +247,10 @@ class TrainDiff(Operation):
     def oper(self, a, b, c) -> torch.Tensor:
         if a is None or b is None or c is None:
             return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        a, c = resize_tensors(a, c, mode=self.alignment_mode)
+        b, c = resize_tensors(b, c, mode=self.alignment_mode)
+
         if torch.allclose(b.float(), c.float()): return torch.zeros_like(a)
         diff_bc = b.float() - c.float()
         dist_bc, dist_ab = torch.abs(diff_bc), torch.abs(b.float() - a.float())
@@ -153,11 +263,17 @@ class ExtractOp(Operation):
     def oper(self, base: torch.Tensor, a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
         if a is None or b is None:
             return None
+        if base is not None:
+            base, a = resize_tensors(base, a, mode=self.alignment_mode)
+            base, b = resize_tensors(base, b, mode=self.alignment_mode)
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+
+        dtype = base.dtype if base is not None else a.dtype
         base_f = base.float() if base is not None else 0
         a_f, b_f = a.float() - base_f, b.float() - base_f
-        c = torch.cosine_similarity(a_f.flatten(), b_f.flatten(), 0).clamp(-1, 1)
+        c = torch.cosine_similarity(a_f, b_f, -1).clamp(-1, 1).unsqueeze(-1)
         d = ((c + 1) / 2)**self.gamma
-        return (torch.lerp(a_f, b_f, self.alpha) * torch.lerp(d, 1 - d, self.beta)).to(a.dtype)
+        return (torch.lerp(a_f, b_f, self.alpha) * torch.lerp(d, 1 - d, self.beta)).to(dtype)
 
 
 class Similarities(ExtractOp):
@@ -172,10 +288,14 @@ class PowerUpOp(Operation):
     def oper(self, a, b):
         if a is None or b is None:
             return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
         delta = b - a
+        keep_prob = 1 - self.alpha
+        if keep_prob <= 0:
+            return torch.zeros_like(delta)
         rng = torch.Generator(device=delta.device).manual_seed(self.seed)
-        m = torch.bernoulli(torch.full_like(delta, 1 - self.alpha), generator=rng)
-        return (m * delta) / (1 - self.alpha) if (1 - self.alpha) > 0 else torch.zeros_like(delta)
+        m = (torch.empty_like(delta).uniform_(0, 1, generator=rng) < keep_prob).to(delta.dtype)
+        return (m * delta) / keep_prob
 
 
 class InterpolateDifference(Operation):
@@ -185,6 +305,7 @@ class InterpolateDifference(Operation):
     def oper(self, a, b):
         if a is None or b is None:
             return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
         alpha, delta = max(self.alpha, 0.001), torch.abs(a - b)
         max_delta = torch.max(delta)
         if max_delta == 0: return a
@@ -194,13 +315,72 @@ class InterpolateDifference(Operation):
         mask = torch.lerp(torch.bernoulli(torch.clamp(diff, 0, 1), generator=rng), diff, self.gamma)
         return a * (1 - mask) + b * mask
 
+class ManualEnhancedInterpolateDifferenceOp(Operation):
+    def __init__(self, key, alpha, beta, gamma, delta, seed, *sources):
+        super().__init__(key, *sources)
+        self.alpha, self.beta, self.gamma, self.delta, self.seed = alpha, beta, gamma, delta, seed
+
+    def oper(self, a, b):
+        if a is None or b is None: return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        delta = torch.abs(a - b)
+        diff = torch.nan_to_num((torch.max(delta) - delta) / torch.max(delta))
+        mean_diff = torch.mean(diff, 0, keepdim=True)
+        mask = torch.logical_and(self.beta < mean_diff, mean_diff < self.gamma)
+        powered_diff = torch.nan_to_num(diff ** (1 / max(self.alpha, 0.001) - 1))
+        masked_diff = powered_diff * mask.float()
+        rng = torch.Generator(device=a.device).manual_seed(self.seed)
+        random_mask = torch.bernoulli(torch.clamp(masked_diff, 0, 1), generator=rng)
+        interpolated_mask = torch.lerp(random_mask, masked_diff, self.delta)
+        return a * (1 - interpolated_mask) + b * interpolated_mask
+
+
+class AutoEnhancedInterpolateDifferenceOp(Operation):
+    def __init__(self, key, alpha, beta, gamma, seed, *sources):
+        super().__init__(key, *sources)
+        self.alpha, self.beta, self.gamma, self.seed = alpha, beta, gamma, seed
+
+    def oper(self, a, b):
+        if a is None or b is None: return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        delta = torch.abs(a - b)
+        max_delta = torch.max(delta)
+        diff = torch.nan_to_num((max_delta - delta) / max_delta)
+        mean_diff = torch.mean(diff)
+        lower_threshold = mean_diff * (1 - self.beta)
+        upper_threshold = mean_diff * (1 + self.beta)
+        mask = torch.logical_and(lower_threshold < diff, diff < upper_threshold)
+        powered_diff = torch.nan_to_num(diff ** (1 / max(self.alpha, 0.001) - 1))
+        masked_diff = powered_diff * mask.float()
+        rng = torch.Generator(device=a.device).manual_seed(self.seed)
+        random_mask = torch.bernoulli(torch.clamp(masked_diff, 0, 1), generator=rng)
+        interpolated_mask = torch.lerp(random_mask, masked_diff, self.gamma)
+        return a * (1 - interpolated_mask) + b * interpolated_mask
+
+
+class WeightSumCutoffOp(Operation):
+    def __init__(self, key, alpha, beta, gamma, *sources):
+        super().__init__(key, *sources)
+        self.alpha, self.beta, self.gamma = alpha, beta, gamma
+
+    def oper(self, a, b):
+        if a is None or b is None: return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        delta = torch.abs(a - b)
+        diff = torch.nan_to_num((torch.max(delta) - delta) / torch.max(delta))
+        mean = torch.mean(diff, 0, True)
+        mask = torch.logical_and(mean < self.beta, self.gamma < mean)
+        mul = self.alpha * mask
+        return a * (1 - mul) + b * mul
+
+
 # ################# #
 # CALCULATION MODES
 # ################# #
 class CalcMode:
     name = 'calcmode'; description = 'description'; models_used = []; param_docs = {}
     def create_recipe(self, key, **kwargs) -> Operation: raise NotImplementedError
-    
+
     @staticmethod
     def _get_secondary_loader_args(kwargs, mismatch_mode):
         """Helper to build loader args for secondary models (B, C) with fallback support."""
@@ -219,13 +399,13 @@ class CalcMode:
 
 class PreloadedTensor(Operation):
     """Returns a pre-loaded tensor instead of loading from file.
-    
+
     Used when Model A's tensor is pre-loaded in the merge loop.
     """
     def __init__(self, key, tensor):
         super().__init__(key)
         self.tensor = tensor
-    
+
     def merge(self) -> torch.Tensor:
         return self.tensor
 
@@ -236,22 +416,25 @@ class WeightSum(CalcMode):
     param_docs = {'alpha': 'Interpolation weight. 0.0 is 100% Model A, 1.0 is 100% Model B.'}
     def create_recipe(self, key, **kwargs):
         mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
-        
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
         # Use pre-loaded tensor_a if available
         if '_tensor_a' in kwargs:
             a = PreloadedTensor(key, kwargs['_tensor_a'])
         else:
             loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
             a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
-        
+
         loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
         b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
-        
+
         if kwargs['alpha'] >= 1: return b
         if kwargs['alpha'] <= 0: return a
         c = Multiply(key, 1 - kwargs['alpha'], a)
         d = Multiply(key, kwargs['alpha'], b)
-        return Add(key, c, d)
+        res = Add(key, c, d)
+        res.alignment_mode = alignment_mode
+        return res
 
 
 class AddDifference(CalcMode):
@@ -261,20 +444,24 @@ class AddDifference(CalcMode):
     param_docs = {'alpha': 'Multiplier for the (B - C) difference.', 'beta': 'Smoothing (0=Off, 1=On).'}
     def create_recipe(self, key, **kwargs):
         mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
-        
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
         if '_tensor_a' in kwargs:
             a = PreloadedTensor(key, kwargs['_tensor_a'])
         else:
             loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
             a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
-        
+
         loader_args_bc = self._get_secondary_loader_args(kwargs, mismatch_mode)
         b = LoadTensor(key, kwargs['model_b'], **loader_args_bc)
         c = LoadTensor(key, kwargs['model_c'], **loader_args_bc)
         diff = Sub(key, b, c)
+        diff.alignment_mode = alignment_mode
         if kwargs['beta'] == 1: diff = Smooth(key, diff)
         diffm = Multiply(key, kwargs['alpha'], diff)
-        return Add(key, a, diffm)
+        res = Add(key, a, diffm)
+        res.alignment_mode = alignment_mode
+        return res
 
 
 class TrainDifference(CalcMode):
@@ -284,19 +471,23 @@ class TrainDifference(CalcMode):
     param_docs = {'alpha': 'Multiplier for the calculated difference.'}
     def create_recipe(self, key, **kwargs):
         mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
-        
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
         if '_tensor_a' in kwargs:
             a = PreloadedTensor(key, kwargs['_tensor_a'])
         else:
             loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
             a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
-        
+
         loader_args_bc = self._get_secondary_loader_args(kwargs, mismatch_mode)
         b = LoadTensor(key, kwargs['model_b'], **loader_args_bc)
         c = LoadTensor(key, kwargs['model_c'], **loader_args_bc)
         diff = TrainDiff(key, a, b, c)
+        diff.alignment_mode = alignment_mode
         diffm = Multiply(key, kwargs['alpha'], diff)
-        return Add(key, a, diffm)
+        res = Add(key, a, diffm)
+        res.alignment_mode = alignment_mode
+        return res
 class InterpDifference(CalcMode):
     name = 'Comparative-Interpolation'
     description = 'Interpolates A and B based on relative tensor value differences.'
@@ -304,20 +495,23 @@ class InterpDifference(CalcMode):
     param_docs = {'alpha': 'Curve strength.', 'beta': 'Style (0=Similarity, 1=Difference).', 'gamma': 'Mix (0=Random, 1=Linear).'}
     def create_recipe(self, key, **kwargs):
         mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
-        
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
         if '_tensor_a' in kwargs:
             a = PreloadedTensor(key, kwargs['_tensor_a'])
         else:
             loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
             a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
-        
+
         # Skip certain keys entirely
         if key.startswith('cond_stage_model.transformer.text_model.embeddings'):
             return a
-        
+
         loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
         b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
-        return InterpolateDifference(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'], kwargs['seed'], a, b)
+        res = InterpolateDifference(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'], kwargs['seed'], a, b)
+        res.alignment_mode = alignment_mode
+        return res
 
 
 class Extract(CalcMode):
@@ -327,19 +521,23 @@ class Extract(CalcMode):
     param_docs = {'alpha': 'Weighting B vs C.', 'beta': 'Similarity vs dissimilarity.', 'gamma': 'Similarity bias.', 'delta': 'Multiplier for final features.'}
     def create_recipe(self, key, **kwargs):
         mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
-        
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
         if '_tensor_a' in kwargs:
             a = PreloadedTensor(key, kwargs['_tensor_a'])
         else:
             loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
             a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
-        
+
         loader_args_bc = self._get_secondary_loader_args(kwargs, mismatch_mode)
         b = LoadTensor(key, kwargs['model_b'], **loader_args_bc)
         c = LoadTensor(key, kwargs['model_c'], **loader_args_bc)
         extracted = ExtractOp(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'] * 15, a, b, c)
+        extracted.alignment_mode = alignment_mode
         multiplied = Multiply(key, kwargs['delta'], extracted)
-        return Add(key, a, multiplied)
+        res = Add(key, a, multiplied)
+        res.alignment_mode = alignment_mode
+        return res
 
 
 class AddDisimilarity(CalcMode):
@@ -349,19 +547,23 @@ class AddDisimilarity(CalcMode):
     param_docs = {'alpha': 'Weighting B vs C.', 'beta': 'Multiplier for final features.', 'gamma': 'Similarity bias.'}
     def create_recipe(self, key, **kwargs):
         mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
-        
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
         if '_tensor_a' in kwargs:
             a = PreloadedTensor(key, kwargs['_tensor_a'])
         else:
             loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
             a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
-        
+
         loader_args_bc = self._get_secondary_loader_args(kwargs, mismatch_mode)
         b = LoadTensor(key, kwargs['model_b'], **loader_args_bc)
         c = LoadTensor(key, kwargs['model_c'], **loader_args_bc)
         extracted = Similarities(key, kwargs['alpha'], 1, kwargs['gamma'] * 15, b, c)
+        extracted.alignment_mode = alignment_mode
         multiplied = Multiply(key, kwargs['beta'], extracted)
-        return Add(key, a, multiplied)
+        res = Add(key, a, multiplied)
+        res.alignment_mode = alignment_mode
+        return res
 
 
 class PowerUp(CalcMode):
@@ -371,18 +573,22 @@ class PowerUp(CalcMode):
     param_docs = {'alpha': 'Dropout rate (proportion of weights from B to *drop*).', 'beta': 'Multiplier for the final difference.'}
     def create_recipe(self, key, **kwargs):
         mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
-        
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
         if '_tensor_a' in kwargs:
             a = PreloadedTensor(key, kwargs['_tensor_a'])
         else:
             loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
             a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
-        
+
         loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
         b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
         deltahat = PowerUpOp(key, kwargs['alpha'], kwargs['seed'], a, b)
-        res = Multiply(key, kwargs['beta'], deltahat)
-        return Add(key, a, res)
+        deltahat.alignment_mode = alignment_mode # Important!
+        multiplied = Multiply(key, kwargs['beta'], deltahat)
+        res = Add(key, a, multiplied)
+        res.alignment_mode = alignment_mode
+        return res
 
 
 class SVDMode(CalcMode):
@@ -392,17 +598,92 @@ class SVDMode(CalcMode):
     param_docs = {'alpha': 'Rank for standard layers.', 'beta': 'Rank for 3x3 conv layers.', 'gamma': 'Clamp quantile for weights.'}
     def create_recipe(self, key, **kwargs):
         mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
-        
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
         if '_tensor_a' in kwargs:
             a = PreloadedTensor(key, kwargs['_tensor_a'])
         else:
             loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
             a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
-        
+
         loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
         b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
-        return SVD(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'], a, b)
+        res = SVD(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'], a, b, target_shape=kwargs.get('_tensor_a_shape'))
+        res.alignment_mode = alignment_mode
+        return res
 
 
-TWO_MODEL_MODES = [WeightSum(), InterpDifference(), PowerUp(), SVDMode()]
+class ManEnhInterpDifference(CalcMode):
+    name = 'Enhanced Man Interp'
+    description = 'Enhanced interpolation between each pair of values from A and B depending on their difference relative to other values.'
+    models_used = ['A', 'B']
+    param_docs = {'alpha': 'Interpolation strength.', 'beta': 'Lower mean threshold.', 'gamma': 'Upper mean threshold.', 'delta': 'Smoothness factor.'}
+    def create_recipe(self, key, **kwargs):
+        mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
+        if '_tensor_a' in kwargs:
+            a = PreloadedTensor(key, kwargs['_tensor_a'])
+        else:
+            loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
+            a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
+
+        if key.startswith('cond_stage_model.transformer.text_model.embeddings'):
+            return a
+
+        loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
+        b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
+        res = ManualEnhancedInterpolateDifferenceOp(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'], kwargs['delta'], kwargs['seed'], a, b)
+        res.alignment_mode = alignment_mode
+        return res
+
+
+class AutoEnhInterpDifference(CalcMode):
+    name = 'Enhanced Auto Interp'
+    description = 'Interpolates between each pair of values from A and B depending on their difference relative to other values.'
+    models_used = ['A', 'B']
+    param_docs = {'alpha': 'Interpolation strength.', 'beta': 'Threshold adjustment factor.', 'gamma': 'Smoothness factor.'}
+    def create_recipe(self, key, **kwargs):
+        mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
+        if '_tensor_a' in kwargs:
+            a = PreloadedTensor(key, kwargs['_tensor_a'])
+        else:
+            loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
+            a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
+
+        if key.startswith('cond_stage_model.transformer.text_model.embeddings'):
+            return a
+
+        loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
+        b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
+        res = AutoEnhancedInterpolateDifferenceOp(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'], kwargs['seed'], a, b)
+        res.alignment_mode = alignment_mode
+        return res
+
+
+class WeightSumCutoffMode(CalcMode):
+    name = 'Weight-Sum Cutoff'
+    description = 'Weight-sum with cutoff based on value differences.'
+    models_used = ['A', 'B']
+    param_docs = {'alpha': 'Interpolation weight.', 'beta': 'Upper threshold.', 'gamma': 'Lower threshold.'}
+    def create_recipe(self, key, **kwargs):
+        mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
+        if '_tensor_a' in kwargs:
+            a = PreloadedTensor(key, kwargs['_tensor_a'])
+        else:
+            loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
+            a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
+
+        loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
+        b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
+        res = WeightSumCutoffOp(key, kwargs['alpha'], kwargs['beta'], kwargs['gamma'], a, b)
+        res.alignment_mode = alignment_mode
+        return res
+
+
+TWO_MODEL_MODES = [WeightSum(), InterpDifference(), PowerUp(), SVDMode(), ManEnhInterpDifference(), AutoEnhInterpDifference(), WeightSumCutoffMode()]
 THREE_MODEL_MODES = [AddDifference(), TrainDifference(), Extract(), AddDisimilarity()]

nodes/merger_utils.py

@@ -1,36 +1,275 @@
+"""
+Memory-efficient safetensors utilities with parallel I/O and pinned memory support.
+
+Based on benchmarking, parallel reading with ThreadPoolExecutor provides 2-4x speedup
+for large models compared to sequential reads.
+
+Worker count is automatically optimized based on device capabilities when workers=None.
+Pinned memory enables faster CPU→GPU transfers (2-3x improvement).
+"""
+import os
+import mmap
 import json
 import struct
 import torch
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from typing import Dict, List, Optional, Any, Tuple
+
+# Import ComfyUI's memory pinning functions if available
+try:
+    from comfy.model_management import pin_memory, unpin_memory, is_device_cuda
+    PINNING_AVAILABLE = True
+except ImportError:
+    PINNING_AVAILABLE = False
+    def pin_memory(tensor): return False
+    def unpin_memory(tensor): return False
+    def is_device_cuda(device): return str(device).startswith('cuda')
+
+
+class PinnedTensor:
+    """Context manager for automatic pinned memory management.
+
+    Usage:
+        with PinnedTensor(cpu_tensor) as pinned:
+            gpu_tensor = pinned.to('cuda')
+        # Memory automatically unpinned after transfer
+    """
+
+    def __init__(self, tensor: torch.Tensor, auto_pin: bool = True):
+        self.tensor = tensor
+        self.is_pinned = False
+        self.auto_pin = auto_pin and PINNING_AVAILABLE
+
+    def __enter__(self) -> torch.Tensor:
+        if self.auto_pin and self.tensor.device.type == 'cpu':
+            self.is_pinned = pin_memory(self.tensor)
+        return self.tensor
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.is_pinned:
+            unpin_memory(self.tensor)
+            self.is_pinned = False
+
+
+def transfer_to_gpu_pinned(
+    tensor: torch.Tensor,
+    device: str = 'cuda',
+    dtype: Optional[torch.dtype] = None
+) -> torch.Tensor:
+    """Transfer tensor to GPU using pinned memory for faster transfer.
+
+    Args:
+        tensor: CPU tensor to transfer
+        device: Target GPU device
+        dtype: Optional dtype conversion
+
+    Returns:
+        Tensor on GPU device
+    """
+    if not PINNING_AVAILABLE or not is_device_cuda(torch.device(device)):
+        # Fall back to regular transfer
+        if dtype:
+            return tensor.to(device=device, dtype=dtype)
+        return tensor.to(device=device)
+
+    # Pin memory for faster transfer
+    was_pinned = pin_memory(tensor)
+    try:
+        if dtype:
+            result = tensor.to(device=device, dtype=dtype, non_blocking=True)
+        else:
+            result = tensor.to(device=device, non_blocking=True)
+        # Sync to ensure transfer is complete before unpinning
+        if was_pinned:
+            torch.cuda.current_stream().synchronize()
+        return result
+    finally:
+        if was_pinned:
+            unpin_memory(tensor)
+
 
 class MemoryEfficientSafeOpen:
-    def __init__(self, filename, device='cpu'):
+    """Memory-efficient safetensors file reader with mmap, parallel I/O, and pinned memory support.
+
+    Features:
+    - mmap mode: Zero-copy tensor access via memory-mapped file
+    - low_memory mode: Direct file reads to minimize OS page cache usage
+    - Parallel loading: Multi-threaded tensor reads for 2-4x speedup
+    - Sorted batch reads: Keys sorted by file offset for sequential I/O
+    - Auto-optimized workers: Adjusts parallelism based on device capabilities
+    - Pinned memory: Faster CPU→GPU transfers when transferring to CUDA
+
+    Args:
+        filename: Path to safetensors file
+        device: Target device (default 'cpu')
+        mmap_mode: Use memory-mapped file for zero-copy (default True)
+        low_memory: Use direct file reads to minimize memory (overrides mmap_mode)
+    """
+
+    def __init__(self, filename: str, device: str = 'cpu', mmap_mode: bool = True, low_memory: bool = False):
         self.filename = filename
         self.device = device
+        # low_memory mode forces mmap off to avoid OS page caching
+        self.low_memory = low_memory
+        self.mmap_mode = mmap_mode and not low_memory
         self.header, self.header_size = self._read_header()
         self.file = open(filename, "rb")
+        self.mmap_obj = None
+
+        if self.mmap_mode:
+            self.mmap_obj = mmap.mmap(self.file.fileno(), 0, access=mmap.ACCESS_READ)
 
     def __enter__(self):
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.mmap_obj:
+            self.mmap_obj.close()
         self.file.close()
 
-    def keys(self):
+
+    def keys(self) -> List[str]:
+        """Return all tensor keys (excluding metadata)."""
         return [k for k in self.header.keys() if k != "__metadata__"]
 
-    def get_tensor(self, key):
+    def metadata(self) -> Dict[str, str]:
+        """Return file metadata."""
+        return self.header.get("__metadata__", {})
+
+    def keys_sorted_by_offset(self) -> List[str]:
+        """Return keys sorted by file offset for optimal sequential I/O."""
+        keys_with_offsets = []
+        for key in self.keys():
+            offset = self.header[key]["data_offsets"][0]
+            keys_with_offsets.append((key, offset))
+        keys_with_offsets.sort(key=lambda x: x[1])
+        return [k for k, _ in keys_with_offsets]
+
+    def get_tensor(self, key: str) -> torch.Tensor:
+        """Load a single tensor by key."""
         if key not in self.header:
             raise KeyError(f"Tensor '{key}' not found in the file")
 
         metadata = self.header[key]
         offset_start, offset_end = metadata["data_offsets"]
 
-        tensor_bytes = None
-        if offset_start != offset_end:
+        if self.mmap_mode and self.mmap_obj:
+            if offset_start != offset_end:
+                start = self.header_size + 8 + offset_start
+                end = self.header_size + 8 + offset_end
+                tensor_bytes = memoryview(self.mmap_obj)[start:end]
+            else:
+                tensor_bytes = None
+        else:
+            # Non-mmap mode: use pre-allocated bytearray with readinto for minimal copies
+            tensor_bytes = None
+            if offset_start != offset_end:
+                self.file.seek(self.header_size + 8 + offset_start)
+                # Pre-allocate writable buffer and read directly into it
+                tensor_bytes = bytearray(offset_end - offset_start)
+                self.file.readinto(tensor_bytes)
+
+        return self._deserialize_tensor(tensor_bytes, metadata)
+
+
+    def get_tensor_to_gpu(
+        self,
+        key: str,
+        device: str = 'cuda',
+        dtype: Optional[torch.dtype] = None
+    ) -> torch.Tensor:
+        """Load a tensor and transfer to GPU using pinned memory for faster transfer.
+
+        Args:
+            key: Tensor key to load
+            device: Target GPU device (default 'cuda')
+            dtype: Optional dtype conversion
+
+        Returns:
+            Tensor on GPU device
+        """
+        cpu_tensor = self.get_tensor(key)
+        return transfer_to_gpu_pinned(cpu_tensor, device, dtype)
+
+    def get_tensors_parallel(
+        self,
+        keys: Optional[List[str]] = None,
+        workers: Optional[int] = None,
+        sort_by_offset: bool = True
+    ) -> Dict[str, torch.Tensor]:
+        """Load multiple tensors in parallel using ThreadPoolExecutor.
+
+        Args:
+            keys: List of keys to load (default: all keys)
+            workers: Number of worker threads (default: auto-calculated based on device)
+            sort_by_offset: Sort keys by file offset before loading (default: True)
+
+        Returns:
+            Dict mapping key -> tensor
+        """
+        if keys is None:
+            keys = self.keys()
+
+        # Auto-calculate workers based on device capabilities
+        if workers is None:
+            try:
+                from .device_utils import get_optimal_workers, estimate_model_size
+                model_size = estimate_model_size(self.filename)
+                workers = get_optimal_workers(model_size)
+            except ImportError:
+                workers = min(4, os.cpu_count() or 4)
+
+        if sort_by_offset:
+            # Sort by offset for better I/O patterns
+            keys_with_offsets = [(k, self.header[k]["data_offsets"][0]) for k in keys]
+            keys_with_offsets.sort(key=lambda x: x[1])
+            keys = [k for k, _ in keys_with_offsets]
+
+        result = {}
+
+        with ThreadPoolExecutor(max_workers=workers) as executor:
+            futures = {
+                executor.submit(
+                    _read_tensor_from_file,
+                    self.filename,
+                    key,
+                    self.header[key],
+                    self.header_size
+                ): key
+                for key in keys
+            }
+
+            for future in as_completed(futures):
+                key = futures[future]
+                try:
+                    tensor = future.result()
+                    if tensor is not None:
+                        result[key] = tensor
+                except Exception as e:
+                    print(f"[MemoryEfficientSafeOpen] Failed to load {key}: {e}")
+
+        return result
+
+    def get_tensor_as_dict(self, key: str) -> Dict[str, Any]:
+        """Load a tensor and decode as JSON dict (for config tensors)."""
+        if key not in self.header:
+            raise KeyError(f"Tensor '{key}' not found")
+
+        metadata = self.header[key]
+        offset_start, offset_end = metadata["data_offsets"]
+
+        if offset_start == offset_end:
+            return {}
+
+        if self.mmap_mode and self.mmap_obj:
+            start = self.header_size + 8 + offset_start
+            end = self.header_size + 8 + offset_end
+            tensor_bytes = bytes(self.mmap_obj[start:end])
+        else:
             self.file.seek(self.header_size + 8 + offset_start)
             tensor_bytes = self.file.read(offset_end - offset_start)
 
-        return self._deserialize_tensor(tensor_bytes, metadata)
+        return json.loads(tensor_bytes.decode("utf-8"))
 
     def _read_header(self):
         with open(self.filename, "rb") as f:
@@ -46,13 +285,19 @@ class MemoryEfficientSafeOpen:
         if tensor_bytes is None:
             byte_tensor = torch.empty(0, dtype=torch.uint8)
         else:
-            byte_tensor = torch.frombuffer(bytearray(tensor_bytes), dtype=torch.uint8)
+            # Avoid extra copy if already a bytearray (low_memory mode)
+            if isinstance(tensor_bytes, bytearray):
+                byte_tensor = torch.frombuffer(tensor_bytes, dtype=torch.uint8)
+            else:
+                # mmap memoryview needs copy to create writable tensor
+                byte_tensor = torch.frombuffer(bytearray(tensor_bytes), dtype=torch.uint8)
 
         if dtype_str in ["F8_E5M2", "F8_E4M3"]:
             return self._convert_float8(byte_tensor, dtype_str, shape)
 
         return byte_tensor.view(dtype).reshape(shape)
 
+
     @staticmethod
     def _get_torch_dtype(dtype_str):
         dtype_map = {
@@ -80,11 +325,52 @@ class MemoryEfficientSafeOpen:
             raise ValueError(f"Unsupported float8 type: {dtype_str}. Your PyTorch version may be too old.")
 
 
-def get_union_keys(*handlers):
-    """Returns the union of all tensor keys across multiple model handlers.
-    
-    Useful when you want to iterate over all unique keys present in any model.
+def _read_tensor_from_file(
+    filepath: str,
+    key: str,
+    metadata: Dict,
+    header_size: int
+) -> Optional[torch.Tensor]:
+    """Helper function to read a single tensor from file (for parallel execution).
+
+    Each worker opens its own file handle for thread safety.
     """
+    offset_start, offset_end = metadata["data_offsets"]
+    dtype_str = metadata["dtype"]
+    shape = metadata["shape"]
+
+    if offset_start == offset_end:
+        return None
+
+    dtype_map = {
+        "F64": torch.float64, "F32": torch.float32, "F16": torch.float16, "BF16": torch.bfloat16,
+        "I64": torch.int64, "I32": torch.int32, "I16": torch.int16, "I8": torch.int8,
+        "U8": torch.uint8, "BOOL": torch.bool,
+    }
+    if hasattr(torch, "float8_e5m2"):
+        dtype_map["F8_E5M2"] = torch.float8_e5m2
+    if hasattr(torch, "float8_e4m3fn"):
+        dtype_map["F8_E4M3"] = torch.float8_e4m3fn
+
+    dtype = dtype_map.get(dtype_str, torch.float32)
+
+    with open(filepath, "rb") as f:
+        f.seek(header_size + 8 + offset_start)
+        tensor_bytes = f.read(offset_end - offset_start)
+
+    byte_tensor = torch.frombuffer(bytearray(tensor_bytes), dtype=torch.uint8)
+
+    if dtype_str in ["F8_E5M2", "F8_E4M3"]:
+        if dtype_str == "F8_E5M2" and hasattr(torch, "float8_e5m2"):
+            return byte_tensor.view(torch.float8_e5m2).reshape(shape)
+        elif dtype_str == "F8_E4M3" and hasattr(torch, "float8_e4m3fn"):
+            return byte_tensor.view(torch.float8_e4m3fn).reshape(shape)
+
+    return byte_tensor.view(dtype).reshape(shape)
+
+
+def get_union_keys(*handlers) -> List[str]:
+    """Returns the union of all tensor keys across multiple model handlers."""
     all_keys = set()
     for handler in handlers:
         if handler is not None:
@@ -92,11 +378,8 @@ def get_union_keys(*handlers):
     return sorted(all_keys)
 
 
-def get_intersection_keys(*handlers):
-    """Returns only keys present in ALL model handlers.
-    
-    Useful when you want to iterate only over keys that exist in every model.
-    """
+def get_intersection_keys(*handlers) -> List[str]:
+    """Returns only keys present in ALL model handlers."""
     handlers = [h for h in handlers if h is not None]
     if not handlers:
         return []

nodes/metakeys.py

@@ -1,15 +1,17 @@
 import os
+import json
+import struct
 import folder_paths
+import comfy.utils
+from tqdm import tqdm
 from comfy_api.latest import io
-from .utils import convert_pt_to_safetensors, load_metadata_from_safetensors
+from safetensors import safe_open
+from .utils import convert_pt_to_safetensors
 
 
 def _get_metakeys(model_name: str, model_type: str) -> tuple[str, str]:
     """Shared logic for all MetaKeys nodes."""
     model_path = folder_paths.get_full_path_or_raise(model_type, model_name)
-    model_weights = {}
-    metadata = {}
-    layer_shapes = ""
 
     if model_path.endswith(('.pt', '.pth', '.bin', '.ckpt')):
         temp_safe_path = model_path + ".safetensors"
@@ -19,13 +21,26 @@ def _get_metakeys(model_name: str, model_type: str) -> tuple[str, str]:
                 return f"Conversion failed: {error_message}", ""
         model_path = temp_safe_path
 
-    model_weights, metadata = load_metadata_from_safetensors(model_path)
+    # Read header only - no tensor loading needed
+    with open(model_path, "rb") as f:
+        header_size = struct.unpack("<Q", f.read(8))[0]
+        header_json = f.read(header_size).decode("utf-8")
+    header = json.loads(header_json)
+
+    # Extract metadata and layer shapes from header
+    metadata = header.pop("__metadata__", {})
     metadata_str = str(metadata)
-    for layer_name, tensor in model_weights.items():
-        shape = str(tensor.shape)
-        shape = shape.replace("torch.Size(", "").replace(")", "")
+
+    layer_shapes = ""
+    keys = [k for k in header.keys()]
+    pbar = comfy.utils.ProgressBar(len(keys))
+    for layer_name in tqdm(keys, desc="Reading layer shapes", unit="layers"):
+        info = header[layer_name]
+        shape = str(info.get("shape", []))
         name_shape = f"{layer_name}, {shape}\n"
         layer_shapes += name_shape
+        pbar.update(1)
+
     return metadata_str, layer_shapes
 
 

nodes/prunekeys.py

@@ -1,12 +1,17 @@
 import os
 import re
 import folder_paths
+import comfy.utils
+from tqdm import tqdm
 from comfy_api.latest import io
 from safetensors.torch import save_file
-from .utils import convert_pt_to_safetensors, load_metadata_from_safetensors
+from safetensors import safe_open
+from .utils import convert_pt_to_safetensors
+from .merger_utils import MemoryEfficientSafeOpen
+from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 
 
-def _prune_keys(model_name: str, model_type: str, keys_to_prune_str: str, 
+def _prune_keys(model_name: str, model_type: str, keys_to_prune_str: str,
                 use_regex: bool, output_filename: str) -> str:
     """Shared logic for all PruneKeys nodes."""
     model_path = folder_paths.get_full_path_or_raise(model_type, model_name)
@@ -21,31 +26,47 @@ def _prune_keys(model_name: str, model_type: str, keys_to_prune_str: str,
     else:
         model_path_to_load = model_path
 
-    model_weights, metadata = load_metadata_from_safetensors(model_path_to_load)
+    # 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 {}
 
     patterns = [p.strip() for p in keys_to_prune_str.strip().split('\n') if p.strip()]
 
     if not patterns:
         raise ValueError("No keys/patterns provided to prune.")
 
+    # Stream tensors and filter on the fly
     pruned_tensors = {}
-    for key, tensor in model_weights.items():
-        is_match = False
-        if use_regex:
-            if any(re.search(pattern, key) for pattern in patterns):
-                is_match = True
-        else:
-            if any(pattern in key for pattern in patterns):
-                is_match = True
+    with MemoryEfficientSafeOpen(model_path_to_load) as handler:
+        all_keys = handler.keys()
+        pbar = comfy.utils.ProgressBar(len(all_keys))
+        for key in tqdm(all_keys, desc="Pruning keys", unit="keys"):
+            is_match = False
+            if use_regex:
+                if any(re.search(pattern, key) for pattern in patterns):
+                    is_match = True
+            else:
+                if any(pattern in key for pattern in patterns):
+                    is_match = True
 
-        if not is_match:
-            pruned_tensors[key] = tensor
+            if not is_match:
+                pruned_tensors[key] = handler.get_tensor(key)
+            pbar.update(1)
 
-    model_dir = folder_paths.get_folder_paths(model_type)[0]
+    # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
+    model_dir = folder_paths.get_folder_paths(model_type)[-1]
     output_path = os.path.join(model_dir, f"{output_filename.strip()}.safetensors")
 
     save_file(pruned_tensors, output_path, metadata)
 
+    # Cleanup after operation
+    del pruned_tensors
+    cleanup_after_operation()
+
     return output_path
 
 
@@ -72,9 +93,9 @@ class ModelPruneKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, diffusionmodel_name: str, keys_to_prune: str, 
+    def execute(cls, diffusionmodel_name: str, keys_to_prune: str,
                 use_regex: bool, output_filename: str) -> io.NodeOutput:
-        path = _prune_keys(diffusionmodel_name, "diffusion_models", 
+        path = _prune_keys(diffusionmodel_name, "diffusion_models",
                           keys_to_prune, use_regex, output_filename)
         return io.NodeOutput(path)
 
@@ -102,9 +123,9 @@ class TextEncoderPruneKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, textencoder_name: str, keys_to_prune: str, 
+    def execute(cls, textencoder_name: str, keys_to_prune: str,
                 use_regex: bool, output_filename: str) -> io.NodeOutput:
-        path = _prune_keys(textencoder_name, "text_encoders", 
+        path = _prune_keys(textencoder_name, "text_encoders",
                           keys_to_prune, use_regex, output_filename)
         return io.NodeOutput(path)
 
@@ -132,9 +153,9 @@ class LoRAPruneKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, lora_name: str, keys_to_prune: str, 
+    def execute(cls, lora_name: str, keys_to_prune: str,
                 use_regex: bool, output_filename: str) -> io.NodeOutput:
-        path = _prune_keys(lora_name, "loras", 
+        path = _prune_keys(lora_name, "loras",
                           keys_to_prune, use_regex, output_filename)
         return io.NodeOutput(path)
 
@@ -162,9 +183,9 @@ class CheckpointPruneKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, ckpt_name: str, keys_to_prune: str, 
+    def execute(cls, ckpt_name: str, keys_to_prune: str,
                 use_regex: bool, output_filename: str) -> io.NodeOutput:
-        path = _prune_keys(ckpt_name, "checkpoints", 
+        path = _prune_keys(ckpt_name, "checkpoints",
                           keys_to_prune, use_regex, output_filename)
         return io.NodeOutput(path)
 
@@ -192,8 +213,8 @@ class EmbeddingPruneKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, embedding: str, keys_to_prune: str, 
+    def execute(cls, embedding: str, keys_to_prune: str,
                 use_regex: bool, output_filename: str) -> io.NodeOutput:
-        path = _prune_keys(embedding, "embeddings", 
+        path = _prune_keys(embedding, "embeddings",
                           keys_to_prune, use_regex, output_filename)
         return io.NodeOutput(path)

nodes/renamekeys.py

@@ -1,11 +1,16 @@
 import os
 import folder_paths
+import comfy.utils
+from tqdm import tqdm
 from comfy_api.latest import io
 from safetensors.torch import save_file
-from .utils import convert_pt_to_safetensors, load_metadata_from_safetensors
+from safetensors import safe_open
+from .utils import convert_pt_to_safetensors
+from .merger_utils import MemoryEfficientSafeOpen
+from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 
 
-def _rename_keys(model_name: str, model_type: str, old_keys_str: str, 
+def _rename_keys(model_name: str, model_type: str, old_keys_str: str,
                  new_keys_str: str, output_filename: str) -> str:
     """Shared logic for all RenameKeys nodes."""
     model_path = folder_paths.get_full_path_or_raise(model_type, model_name)
@@ -20,7 +25,13 @@ def _rename_keys(model_name: str, model_type: str, old_keys_str: str,
     else:
         model_path_to_load = model_path
 
-    model_weights, metadata = load_metadata_from_safetensors(model_path_to_load)
+    # 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 {}
 
     old_keys = [key.strip() for key in old_keys_str.strip().split('\n') if key.strip()]
     new_keys = [key.strip() for key in new_keys_str.strip().split('\n') if key.strip()]
@@ -30,17 +41,26 @@ def _rename_keys(model_name: str, model_type: str, old_keys_str: str,
 
     key_map = dict(zip(old_keys, new_keys))
 
-    original_keys = list(model_weights.keys())
+    # Stream tensors and rename on the fly
     renamed_tensors = {}
-    for key in original_keys:
-        new_key = key_map.get(key, key)
-        renamed_tensors[new_key] = model_weights[key]
+    with MemoryEfficientSafeOpen(model_path_to_load) as handler:
+        original_keys = handler.keys()
+        pbar = comfy.utils.ProgressBar(len(original_keys))
+        for key in tqdm(original_keys, desc="Renaming keys", unit="keys"):
+            new_key = key_map.get(key, key)
+            renamed_tensors[new_key] = handler.get_tensor(key)
+            pbar.update(1)
 
-    model_dir = folder_paths.get_folder_paths(model_type)[0]
+    # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
+    model_dir = folder_paths.get_folder_paths(model_type)[-1]
     output_path = os.path.join(model_dir, f"{output_filename.strip()}.safetensors")
 
     save_file(renamed_tensors, output_path, metadata)
 
+    # Cleanup after operation
+    del renamed_tensors
+    cleanup_after_operation()
+
     return output_path
 
 
@@ -67,9 +87,9 @@ class ModelRenameKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, diffusionmodel_name: str, old_keys: str, 
+    def execute(cls, diffusionmodel_name: str, old_keys: str,
                 new_keys: str, output_filename: str) -> io.NodeOutput:
-        path = _rename_keys(diffusionmodel_name, "diffusion_models", 
+        path = _rename_keys(diffusionmodel_name, "diffusion_models",
                            old_keys, new_keys, output_filename)
         return io.NodeOutput(path)
 
@@ -97,9 +117,9 @@ class TextEncoderRenameKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, textencoder_name: str, old_keys: str, 
+    def execute(cls, textencoder_name: str, old_keys: str,
                 new_keys: str, output_filename: str) -> io.NodeOutput:
-        path = _rename_keys(textencoder_name, "text_encoders", 
+        path = _rename_keys(textencoder_name, "text_encoders",
                            old_keys, new_keys, output_filename)
         return io.NodeOutput(path)
 
@@ -127,9 +147,9 @@ class LoRARenameKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, lora_name: str, old_keys: str, 
+    def execute(cls, lora_name: str, old_keys: str,
                 new_keys: str, output_filename: str) -> io.NodeOutput:
-        path = _rename_keys(lora_name, "loras", 
+        path = _rename_keys(lora_name, "loras",
                            old_keys, new_keys, output_filename)
         return io.NodeOutput(path)
 
@@ -157,9 +177,9 @@ class CheckpointRenameKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, ckpt_name: str, old_keys: str, 
+    def execute(cls, ckpt_name: str, old_keys: str,
                 new_keys: str, output_filename: str) -> io.NodeOutput:
-        path = _rename_keys(ckpt_name, "checkpoints", 
+        path = _rename_keys(ckpt_name, "checkpoints",
                            old_keys, new_keys, output_filename)
         return io.NodeOutput(path)
 
@@ -187,8 +207,8 @@ class EmbeddingRenameKeys(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, embedding: str, old_keys: str, 
+    def execute(cls, embedding: str, old_keys: str,
                 new_keys: str, output_filename: str) -> io.NodeOutput:
-        path = _rename_keys(embedding, "embeddings", 
+        path = _rename_keys(embedding, "embeddings",
                            old_keys, new_keys, output_filename)
         return io.NodeOutput(path)

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "ComfyUI-ModelUtils"
 description = "[WIP]Custom nodes for handling, inspecting, modifying and creating various model files."
-version = "0.1.7"
+version = "0.2.7"
 license = { file = "LICENSE" }
 
 [project.urls]

requirements.txt

@@ -0,0 +1 @@
+# No external dependencies required