Add node to extract text encoder LoRA

Feature/preview meta wf loader

__init__.py

@@ -22,7 +22,7 @@ from .nodes.merger import (
 )
 from .nodes.lora_extract_svd import (
     LoRAExtractFixed, LoRAExtractRatio, LoRAExtractQuantile,
-    LoRAExtractKnee, LoRAExtractFrobenius
+    LoRAExtractKnee, LoRAExtractFrobenius, LoRATextEncoderExtractFixed
 )
 from .nodes.lora_resize import (
     LoRAResizeFixed, LoRAResizeRatio,
@@ -30,8 +30,12 @@ from .nodes.lora_resize import (
     LoRAMergeToModel
 )
 from .nodes.downloader_nodes import (
-    CheckpointDownloader, DiffusionModelDownloader, LoRADownloader, EmbeddingDownloader,
-    VAEDownloader, ControlNetDownloader, ManualPathDownloader
+    CheckpointInfoMetaDownloader, DiffusionModelInfoMetaDownloader, LoRAInfoMetaDownloader, EmbeddingInfoMetaDownloader,
+    VAEInfoMetaDownloader, ControlNetInfoMetaDownloader, ManualPathInfoMetaDownloader
+)
+from .nodes.model_info_nodes import (
+    CheckpointInfoLoader, LoRAInfoLoader, EmbeddingInfoLoader,
+    VAEInfoLoader, ControlNetInfoLoader, DiffusionModelInfoLoader
 )
 
 
@@ -56,7 +60,7 @@ class ModelUtilsExtension(ComfyExtension):
             CheckpointThreeMerger, EmbeddingThreeMerger,
             # LoRA Extraction
             LoRAExtractFixed, LoRAExtractRatio, LoRAExtractQuantile,
-            LoRAExtractKnee, LoRAExtractFrobenius,
+            LoRAExtractKnee, LoRAExtractFrobenius, LoRATextEncoderExtractFixed,
             # LoRA Resize
             LoRAResizeFixed, LoRAResizeRatio,
             LoRAResizeFrobenius, LoRAResizeCumulative,
@@ -65,8 +69,11 @@ class ModelUtilsExtension(ComfyExtension):
             # LoRA Merge To Model
             LoRAMergeToModel,
             # Downloaders
-            CheckpointDownloader, DiffusionModelDownloader, LoRADownloader, EmbeddingDownloader,
-            VAEDownloader, ControlNetDownloader, ManualPathDownloader,
+            CheckpointInfoMetaDownloader, DiffusionModelInfoMetaDownloader, LoRAInfoMetaDownloader, EmbeddingInfoMetaDownloader,
+            VAEInfoMetaDownloader, ControlNetInfoMetaDownloader, ManualPathInfoMetaDownloader,
+            # Info Loaders
+            CheckpointInfoLoader, LoRAInfoLoader, EmbeddingInfoLoader,
+            VAEInfoLoader, ControlNetInfoLoader, DiffusionModelInfoLoader,
         ]
 
 

docs/merger_2_model_modes.md

@@ -33,6 +33,30 @@
 
 ---
 
+## Power-Up (DARE+TIES)
+> Combines DARE (Drop and Rescale) with TIES (Trim, Elect Sign) — the current best-practice algorithm for merging fine-tuned models, particularly LoRA-tuned ones. DARE sparsifies the task vector via random dropout and rescaling; TIES then removes residual low-magnitude noise and enforces sign consistency. The result is a cleaner, more targeted capability transfer from B to A with less interference.
+
+**Models Used:** A, B
+**Parameters:**
+- **Alpha:** DARE drop rate. Fraction of delta parameters randomly zeroed. Higher values produce a sparser, more targeted delta. Typical range: `0.5`–`0.9`.
+- **Beta:** TIES trim quantile. Fraction of the smallest-magnitude delta parameters zeroed after DARE, removing residual noise. `0.0` disables trimming. Typical range: `0.1`–`0.3`.
+- **Gamma:** Lambda scale. Final multiplier applied to the merged delta before adding to Model A. Equivalent to the task arithmetic scaling coefficient. Typical range: `0.5`–`1.5`.
+- **Seed:** Random seed for the DARE dropout mask. Controls reproducibility.
+
+**Algorithm:**
+1. Compute task vector `δ = B − A`
+2. **DARE:** apply random binary mask with keep probability `1 − alpha`; rescale survivors by `1/(1−alpha)`
+3. **TIES trim:** zero out parameters below the `beta`-quantile magnitude threshold
+4. **TIES elect:** determine dominant sign per position; zero out disagreeing parameters
+5. Return `A + gamma × δ_filtered`
+
+**Guidance:**
+- Start with `alpha=0.9, beta=0.2, gamma=1.0` for LoRA-derived models
+- Lower `alpha` (e.g. `0.5`) for full fine-tunes where more delta mass should be retained
+- Increase `gamma` beyond `1.0` to amplify B's influence; decrease below `1.0` to soften it
+
+---
+
 ## Enhanced Man Interp
 > Sophisticated interpolation between values from A and B depending on their difference relative to other values, with manual threshold control.
 

nodes/downloader_nodes.py

@@ -26,18 +26,18 @@ def get_subdirectories(folder_name: str) -> list[str]:
 
     return sorted(list(subdirs))
 
-class BaseDownloaderNode(io.ComfyNode):
+class BaseInfoMetaDownloaderNode(io.ComfyNode):
     CATEGORY = "ModelUtils/Downloader"
     OUTPUT_NODE = True
 
     @classmethod
     def get_common_inputs(cls):
         return {
-            "recursive": io.Boolean.Input("recursive", default=True),
-            "nsfw_level": io.Combo.Input("nsfw_level", options=["None", "Soft", "Mature", "X", "XXX", "All"], default="All"),
-            "max_examples": io.Int.Input("max_examples", default=0, min=0, max=100),
-            "threads": io.Int.Input("threads", default=4, min=1, max=32),
-            "api_key": io.String.Input("api_key", default="")
+            "recursive": io.Boolean.Input("recursive", default=True, tooltip="Whether to scan subdirectories recursively."),
+            "nsfw_level": io.Combo.Input("nsfw_level", options=["None", "Soft", "Mature", "X", "XXX", "All"], default="All", tooltip="Filter models based on NSFW level in metadata."),
+            "max_examples": io.Int.Input("max_examples", default=0, min=0, max=100, tooltip="Maximum number of examples to process."),
+            "threads": io.Int.Input("threads", default=4, min=1, max=32, tooltip="Number of threads to use for processing."),
+            "api_key": io.String.Input("api_key", default="", tooltip="API key for accessing the model repository.")
         }
 
     @classmethod
@@ -49,6 +49,7 @@ class BaseDownloaderNode(io.ComfyNode):
             inputs=list(cls.get_common_inputs().values()),
             outputs=[io.String.Output(display_name="status")],
             is_output_node=True,
+            description=("Scans specified directories for models, extracts metadata, generates previews, and optionally compresses files.")
         )
 
     @classmethod
@@ -84,7 +85,7 @@ class BaseDownloaderNode(io.ComfyNode):
     def get_scan_dirs(cls, **kwargs) -> list[str]:
         return []
 
-class CheckpointDownloader(BaseDownloaderNode):
+class CheckpointInfoMetaDownloader(BaseInfoMetaDownloaderNode):
     @classmethod
     def define_schema(cls):
         inputs = list(cls.get_common_inputs().values())
@@ -121,7 +122,7 @@ class CheckpointDownloader(BaseDownloaderNode):
                 target_paths.append(target_path)
         return target_paths
 
-class DiffusionModelDownloader(BaseDownloaderNode):
+class DiffusionModelInfoMetaDownloader(BaseInfoMetaDownloaderNode):
     @classmethod
     def define_schema(cls):
         inputs = list(cls.get_common_inputs().values())
@@ -158,7 +159,7 @@ class DiffusionModelDownloader(BaseDownloaderNode):
                 target_paths.append(target_path)
         return target_paths
 
-class LoRADownloader(BaseDownloaderNode):
+class LoRAInfoMetaDownloader(BaseInfoMetaDownloaderNode):
     @classmethod
     def define_schema(cls):
         inputs = list(cls.get_common_inputs().values())
@@ -195,7 +196,7 @@ class LoRADownloader(BaseDownloaderNode):
                 target_paths.append(target_path)
         return target_paths
 
-class EmbeddingDownloader(BaseDownloaderNode):
+class EmbeddingInfoMetaDownloader(BaseInfoMetaDownloaderNode):
     @classmethod
     def define_schema(cls):
         inputs = list(cls.get_common_inputs().values())
@@ -232,7 +233,7 @@ class EmbeddingDownloader(BaseDownloaderNode):
                 target_paths.append(target_path)
         return target_paths
 
-class VAEDownloader(BaseDownloaderNode):
+class VAEInfoMetaDownloader(BaseInfoMetaDownloaderNode):
     @classmethod
     def define_schema(cls):
         inputs = list(cls.get_common_inputs().values())
@@ -269,7 +270,7 @@ class VAEDownloader(BaseDownloaderNode):
                 target_paths.append(target_path)
         return target_paths
 
-class ControlNetDownloader(BaseDownloaderNode):
+class ControlNetInfoMetaDownloader(BaseInfoMetaDownloaderNode):
     @classmethod
     def define_schema(cls):
         inputs = list(cls.get_common_inputs().values())
@@ -306,7 +307,7 @@ class ControlNetDownloader(BaseDownloaderNode):
                 target_paths.append(target_path)
         return target_paths
 
-class ManualPathDownloader(BaseDownloaderNode):
+class ManualPathInfoMetaDownloader(BaseInfoMetaDownloaderNode):
     @classmethod
     def define_schema(cls):
         inputs = list(cls.get_common_inputs().values())

nodes/downloader_utils.py

@@ -36,12 +36,7 @@ except ImportError:
 # Assuming downloader_utils.py is inside nodes/, the root is one level up
 root_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 sys.path.append(os.path.join(root_dir, 'unifiedefficientloader-reference'))
-try:
-    from unifiedefficientloader import UnifiedSafetensorsLoader
-    HAS_UNIFIED_LOADER = True
-except ImportError:
-    HAS_UNIFIED_LOADER = False
-    print("Warning: unifiedefficientloader not found. Safetensors metadata extraction disabled.")
+from unifiedefficientloader import UnifiedSafetensorsLoader
 
 # --- Database Path Logic ---
 try:
@@ -416,8 +411,6 @@ def filter_images_by_nsfw(images, threshold_level):
     return valid_images
 
 def extract_safetensors_metadata(filepath):
-    if not HAS_UNIFIED_LOADER:
-        return None
     try:
         # Use low_memory=True to only read the header and metadata without loading tensors
         loader = UnifiedSafetensorsLoader(filepath, low_memory=True)
@@ -576,6 +569,60 @@ def process_file(filepath, cache, api, nsfw_threshold, max_examples):
                         for k in stats: stats[k] += res.get(k, 0)
     return stats
 
+import numpy as np
+from PIL import Image, ImageOps
+import torch
+
+def load_image_tensor(path):
+    if not os.path.exists(path):
+        return None
+    try:
+        img = Image.open(path)
+        img = ImageOps.exif_transpose(img)
+        image = img.convert("RGB")
+        image = np.array(image).astype(np.float32) / 255.0
+        image = torch.from_numpy(image)[None,]
+        return image
+    except Exception as e:
+        print(f"Error loading image {path}: {e}")
+        return None
+
+def get_model_workflows(base_path):
+    workflows = []
+    # Main workflow
+    main_json = f"{base_path}.json"
+    if os.path.exists(main_json):
+        workflows.append(main_json)
+
+    # Example workflows
+    # They are named like base_path.example.N.json
+    # or base_path.preview.json (if extracted from preview)
+    preview_json = f"{base_path}.preview.json"
+    if os.path.exists(preview_json) and preview_json not in workflows:
+        workflows.append(preview_json)
+
+    example_jsons = glob.glob(f"{base_path}.example.*.json")
+    # Sort them by index
+    example_jsons.sort()
+    for ej in example_jsons:
+        if ej not in workflows:
+            workflows.append(ej)
+
+    return workflows
+
+def get_model_metadata_file(base_path):
+    # Prefer civitai.info if available
+    civitai_info = f"{base_path}.civitai.info"
+    if os.path.exists(civitai_info):
+        return civitai_info
+
+    # Fallback to metadata.json (safetensors metadata)
+    metadata_json = f"{base_path}.metadata.json"
+    if os.path.exists(metadata_json):
+        return metadata_json
+
+    return None
+
 def scan_and_process(scan_dirs, recursive=True, db_path=None, nsfw_level="All", max_examples=0, api_key=None, threads=4):
     """
     Scans directories for models and fetches their metadata/previews.

nodes/lora_extract_svd.py

@@ -13,12 +13,15 @@ import comfy.utils
 from tqdm import tqdm
 from comfy_api.latest import io
 from safetensors.torch import save_file
-from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
 from .device_utils import (
     estimate_model_size, prepare_for_large_operation,
     cleanup_after_operation, get_device_capabilities
 )
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
+
+
+
 
 # =============================================================================
 # Rank Index Functions - Determine LoRA rank from singular values
@@ -438,6 +441,7 @@ def extract_lora_from_files(
     svd_niter: int = 2,
     lazy_load: bool = True,
     force_clear_cache: bool = True,
+    is_text_encoder: bool = False,
 ) -> dict[str, torch.Tensor]:
     """
     Extract LoRA from difference between two models.
@@ -457,6 +461,7 @@ def extract_lora_from_files(
         skip_patterns_str: Patterns for layers to skip
         mismatch_mode: How to handle mismatches
         chunk_large_layers: Enable chunked extraction
+        is_text_encoder: Whether extracting from text encoder models
     """
     output_sd = {}
 
@@ -484,7 +489,7 @@ def extract_lora_from_files(
         stats = {"extracted": 0, "full": 0, "skipped": 0, "chunked": 0}
 
         def _process_layer(key):
-            lora_name = _format_lora_key(key)
+            lora_name = _format_lora_key(key, is_text_encoder=is_text_encoder)
 
             if _matches_any_pattern(key, skip_patterns):
                 return "skipped", None
@@ -644,6 +649,15 @@ def _get_model_inputs():
     ]
 
 
+def _get_clip_model_inputs():
+    return [
+        io.Combo.Input("model_a", options=folder_paths.get_filename_list("text_encoders"),
+                      tooltip="Finetuned CLIP/Text Encoder (A - B = LoRA)"),
+        io.Combo.Input("model_b", options=folder_paths.get_filename_list("text_encoders"),
+                      tooltip="Base CLIP/Text Encoder (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
@@ -680,18 +694,26 @@ def _reconstruct_dots(key: str) -> str:
     return key
 
 
-def _format_lora_key(key: str) -> str:
+def _format_lora_key(key: str, is_text_encoder: bool = False) -> str:
     """
     Format the key for LoRA saving.
     Standardizes to 'diffusion_model.' or 'transformer.' prefix.
+    If is_text_encoder is True, use 'text_encoders.transformer.'
     """
     if key.endswith(".weight"):
         key = key[:-7]
 
+    if is_text_encoder:
+        return f"text_encoders.transformer.{key}"
+
     # Handle ComfyUI Checkpoint wrapper
-    if key.startswith("model.diffusion_model."):
+    if key.startswith("model.diffusion_model.") or key.startswith("net."):
         # Check if it's a Diffusers-format transformer inside a checkpoint
-        inner_key = key[22:]  # len("model.diffusion_model.")
+        if key.startswith("model.diffusion_model."):
+            inner_key = key[22:]  # len("model.diffusion_model.")
+        else:
+            inner_key = key[4:]   # len("net.")
+
         if inner_key.startswith("transformer_blocks") or inner_key.startswith("single_transformer_blocks"):
             return f"transformer.{inner_key}"
         # Standard checkpoint or Original Flux
@@ -955,3 +977,43 @@ class LoRAExtractFrobenius(io.ComfyNode):
         )
 
         return io.NodeOutput(_save_lora(output_sd, output_filename))
+
+
+class LoRATextEncoderExtractFixed(io.ComfyNode):
+    """Extract LoRA from Text Encoder with fixed rank."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LoRATextEncoderExtractFixed",
+            display_name="LoRA Extract (Text Encoder, Fixed Rank)",
+            category="ModelUtils/LoRA",
+            description="Extract LoRA from Text Encoder/CLIP models with specified fixed rank.",
+            inputs=[
+                *_get_clip_model_inputs(),
+                io.Int.Input("linear_dim", default=32, min=1, max=16384,
+                            tooltip="Rank for linear layers"),
+                io.Int.Input("svd_niter", default=2, min=0, max=10,
+                            tooltip="SVD power iterations"),
+                *_get_common_inputs(),
+            ],
+            outputs=[io.String.Output(display_name="output_path")],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, model_a, model_b, linear_dim, svd_niter, chunk_large_layers,
+                clamp_quantile, min_diff, mismatch_mode, output_filename,
+                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
+
+        model_a_path = folder_paths.get_full_path_or_raise("text_encoders", model_a)
+        model_b_path = folder_paths.get_full_path_or_raise("text_encoders", model_b)
+
+        output_sd = extract_lora_from_files(
+            model_a_path, model_b_path, "fixed", linear_dim, 0,
+            device, save_dtype, linear_dim, 0,
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers, svd_niter,
+            lazy_load, force_clear_cache, is_text_encoder=True
+        )
+
+        return io.NodeOutput(_save_lora(output_sd, output_filename))

nodes/lora_merger.py

@@ -10,7 +10,6 @@ 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,
@@ -21,6 +20,10 @@ from .lora_resize import (
     _format_lora_key
 )
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
+
+
+
 def merge_multi_loras(
     lora_paths: List[str],
     lora_weights: List[float],

nodes/lora_resize.py

@@ -14,9 +14,12 @@ from tqdm import tqdm
 from comfy_api.latest import io
 from safetensors.torch import save_file
 from typing import Optional, Dict, Tuple, List
-from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
 from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
+
+
+
 # Reuse rank functions from extraction module
 from .lora_extract_svd import (
     _index_sv_ratio,
@@ -189,7 +192,7 @@ def detect_lora_rank(handler: MemoryEfficientSafeOpen, pairs: Dict) -> Tuple[int
         # Get dim from down weight shape
         if network_dim is None:
             down_key = block_keys["down"]
-            shape = handler.header[down_key]["shape"]
+            shape = handler.get_shape(down_key)
             # Linear: [rank, in_features] or Conv: [rank, in_ch, k, k]
             network_dim = shape[0]
 

nodes/merger.py

@@ -7,11 +7,12 @@ from tqdm import tqdm
 from comfy_api.latest import io
 from safetensors.torch import save_file
 from .merger_ops import TWO_MODEL_MODES, THREE_MODEL_MODES, MissingTensorBehavior, MissingTensorError
-from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
 from .device_utils import (
     estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 )
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned, IncrementalSafetensorsWriter
+
 
 def load_documentation_from_file(filename):
     """Loads documentation from a markdown file in the ../docs/ directory."""
@@ -76,6 +77,9 @@ class MergerLogic:
             process_device = recipe_params.get('device', 'cpu')
             print(f"[Merger] Preparing memory for {total_size_gb:.2f}GB merge operation...")
             prepare_for_large_operation(total_size_gb * 1.2, torch.device(process_device))
+            flush_threshold_bytes = int(total_size_gb * 1024**3 / 16)
+        else:
+            flush_threshold_bytes = 256 * 1024 * 1024  # 256MB fallback (min 8GB RAM assumed)
 
         handlers = {}
         for name in model_names.values():
@@ -124,12 +128,42 @@ class MergerLogic:
         discarded_keys = 0
         error_keys = []
 
+        output_folder = "loras" if calc_mode == "SVD LoRA Extraction" else model_type
+        # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
+        output_dir = folder_paths.get_folder_paths(output_folder)[-1]
+        os.makedirs(output_dir, exist_ok=True)
+        output_filename = recipe_params.get("output_filename")
+        output_path = os.path.join(output_dir, f"{output_filename}.safetensors")
+
+        writer = None
+
+        writer = IncrementalSafetensorsWriter(output_path, metadata=metadata)
+        writer.__enter__()
+
+        batch_buffer = {}
+        batch_bytes = 0
+        batch_key_count = 0
+
+        def _flush_batch():
+            nonlocal batch_buffer, batch_bytes, batch_key_count
+            if not batch_buffer:
+                return
+            writer.write_dict(batch_buffer)
+            pbar.update(batch_key_count)
+            batch_buffer = {}
+            batch_bytes = 0
+            batch_key_count = 0
+            if recipe_params.get('force_clear_cache', True):
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
         with torch.no_grad():
             for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
                 # Check discard patterns first - skip entirely
                 if _matches_any_pattern(key, discard_patterns):
                     discarded_keys += 1
-                    pbar.update(1)
                     continue
 
                 # Pre-load Model A's tensor with pinned memory for CUDA
@@ -142,9 +176,15 @@ class MergerLogic:
 
                 # Check exclude patterns - use Model A only, no merge
                 if _matches_any_pattern(key, exclude_patterns):
-                    merged_state_dict[key] = tensor_a.detach().to(save_torch_dtype).cpu().clone()
+                    t = tensor_a.detach().to(save_torch_dtype).cpu()
+            
+                    batch_buffer[key] = t
+                    batch_bytes += t.numel() * t.element_size()
+                    batch_key_count += 1
                     excluded_keys += 1
-                    pbar.update(1)
+                    del tensor_a
+                    if (batch_bytes >= flush_threshold_bytes or batch_key_count >= 32):
+                        _flush_batch()
                     continue
 
                 # Pass tensor_a metadata to recipes for zeros mode and fallback
@@ -157,6 +197,9 @@ class MergerLogic:
                     result = recipe.merge()
                 except MissingTensorError as e:
                     if mismatch_mode == MissingTensorBehavior.ERROR:
+                
+                        _flush_batch()
+                        writer.__exit__(None, None, None)
                         raise ValueError(f"Layer mismatch error (mismatch_mode='error'): {e}")
                     result = None
                     error_keys.append(key)
@@ -168,7 +211,10 @@ class MergerLogic:
 
                 if isinstance(result, dict):
                     for r_key, r_tensor in result.items():
-                        merged_state_dict[r_key] = r_tensor.detach().to(save_torch_dtype).cpu().clone()
+                        t = r_tensor.detach().to(save_torch_dtype).cpu()
+                
+                        batch_buffer[r_key] = t
+                        batch_bytes += t.numel() * t.element_size()
                 else:
                     # Ensure compatibility with Model A's architecture.
                     # If alignment_mode is 'pad/crop', we crop results that were padded.
@@ -179,7 +225,13 @@ class MergerLogic:
                             slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(result.shape, target_shape))
                             result = result[slices]
 
-                    merged_state_dict[key] = result.detach().to(save_torch_dtype).cpu().clone()
+                    t = result.detach().to(save_torch_dtype).cpu()
+            
+                    batch_buffer[key] = t
+                    batch_bytes += t.numel() * t.element_size()
+
+        
+                batch_key_count += 1
 
                 # Clean up references to allow GC immediately.
                 # Local loop variables must be explicitly deleted to prevent PyTorch from keeping tensors in VRAM.
@@ -189,13 +241,12 @@ class MergerLogic:
                 del recipe
                 del result
 
-                if recipe_params.get('force_clear_cache', True):
-                    import gc
-                    gc.collect()
-                    if torch.cuda.is_available():
-                        torch.cuda.empty_cache()
+                if (batch_bytes >= flush_threshold_bytes or batch_key_count >= 32):
+                    _flush_batch()
 
-                pbar.update(1)
+
+        _flush_batch()
+        writer.__exit__(None, None, None)
 
         # Log summary
         if excluded_keys > 0:
@@ -216,10 +267,9 @@ class MergerLogic:
         os.makedirs(output_dir, exist_ok=True)
         output_filename = recipe_params.get("output_filename")
         output_path = os.path.join(output_dir, f"{output_filename}.safetensors")
-        save_file(merged_state_dict, output_path, metadata=metadata)
-
+        if os.path.exists(output_path):
+            print(f"[Merger] Output saved to {output_path}")
         # Cleanup after heavy operation
-        del merged_state_dict
         cleanup_after_operation()
 
         return f"{output_filename}.safetensors"
@@ -243,10 +293,10 @@ class CheckpointTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_checkpoint"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -302,10 +352,10 @@ class ModelTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_model"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -361,10 +411,10 @@ class TextEncoderTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_textencoder"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -420,10 +470,10 @@ class LoRATwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_lora"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -479,10 +529,10 @@ class EmbeddingTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_2_embedding"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -541,10 +591,10 @@ class CheckpointThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_checkpoint"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -601,10 +651,10 @@ class ModelThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_model"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -661,10 +711,10 @@ class TextEncoderThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_textencoder"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -721,10 +771,10 @@ class LoRAThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_lora"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),
@@ -781,10 +831,10 @@ class EmbeddingThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
+                io.Float.Input("delta", default=0.5, min=-10.0, max=10.0, step=0.01),
                 io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff),
                 io.String.Input("output_filename", default="merged_3_embedding"),
                 io.Combo.Input("save_dtype", options=["fp32", "fp16", "bf16"]),

nodes/merger_ops.py

@@ -4,7 +4,7 @@ import torch.nn.functional as F
 from enum import Enum
 
 
-from .merger_utils import transfer_to_gpu_pinned
+from unifiedefficientloader import transfer_to_gpu_pinned
 
 class MissingTensorBehavior(Enum):
     """Controls behavior when a tensor key is missing from a model."""
@@ -319,6 +319,53 @@ class PowerUpOp(Operation):
         return (m * delta) / keep_prob
 
 
+class DARETIESOp(Operation):
+    """DARE + TIES combined operation on a single task vector (delta = b - a).
+
+    Step 1 — DARE: randomly drop `drop_rate` fraction of delta parameters,
+                   rescale survivors by 1/(1-drop_rate).
+    Step 2 — TIES trim: zero out parameters below the `trim_quantile` magnitude threshold.
+    Step 3 — TIES elect: determine dominant sign per parameter position.
+    Step 4 — TIES merge: zero out parameters that disagree with the dominant sign.
+
+    For a two-model merge the TIES elect/merge step degenerates to a no-op (single vector,
+    sign is always self-consistent) but the trim step still reduces interference from
+    near-zero noisy deltas. The dominant sign filter is included for correctness and
+    forward compatibility with n-model extensions.
+    """
+    def __init__(self, key, alpha, beta, gamma, seed, *sources):
+        super().__init__(key, *sources)
+        # alpha = drop_rate (DARE), beta = trim_quantile (TIES), gamma = lambda scale
+        self.alpha, self.beta, self.gamma, self.seed = alpha, beta, gamma, seed
+
+    def oper(self, a, b):
+        if a is None or b is None:
+            return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        delta = (b - a).float()
+
+        # --- DARE: drop and rescale ---
+        keep_prob = 1.0 - self.alpha
+        if keep_prob <= 0:
+            return torch.zeros_like(a)
+        rng = torch.Generator(device=delta.device).manual_seed(self.seed)
+        dare_mask = (torch.empty_like(delta).uniform_(0, 1, generator=rng) < keep_prob).float()
+        delta = (delta * dare_mask) / keep_prob
+
+        # --- TIES trim: zero parameters below magnitude threshold ---
+        if self.beta > 0:
+            flat_abs = delta.abs().flatten()
+            k = max(1, int(flat_abs.numel() * self.beta))
+            threshold = torch.kthvalue(flat_abs, k).values
+            delta = torch.where(delta.abs() < threshold, torch.zeros_like(delta), delta)
+
+        # --- TIES elect: dominant sign (trivially self-consistent for single vector) ---
+        dominant_sign = torch.sign(delta)
+        delta = torch.where(torch.sign(delta) == dominant_sign, delta, torch.zeros_like(delta))
+
+        return (a + self.gamma * delta).to(a.dtype)
+
+
 class InterpolateDifference(Operation):
     def __init__(self, key, alpha, beta, gamma, seed, *sources):
         super().__init__(key, *sources)
@@ -710,5 +757,41 @@ class WeightSumCutoffMode(CalcMode):
         return res
 
 
-TWO_MODEL_MODES = [WeightSum(), InterpDifference(), PowerUp(), SVDMode(), ManEnhInterpDifference(), AutoEnhInterpDifference(), WeightSumCutoffMode()]
+class DARETIESMode(CalcMode):
+    name = 'Power-Up (DARE+TIES)'
+    description = 'Adds capabilities of B to A using DARE (Drop and Rescale) followed by TIES (Trim, Elect Sign). Best practice for LoRA-tuned model merging.'
+    models_used = ['A', 'B']
+    param_docs = {
+        'alpha': 'DARE drop rate — fraction of delta parameters randomly zeroed. Higher = more aggressive sparsification.',
+        'beta':  'TIES trim quantile — fraction of smallest-magnitude delta parameters zeroed after DARE. Removes residual noise.',
+        'gamma': 'Lambda — scale multiplier applied to the final merged delta before adding to Model A.',
+        'seed':  'Random seed for the DARE dropout mask.',
+    }
+
+    def create_recipe(self, key, **kwargs):
+        mismatch_mode = kwargs.get('mismatch_mode', MissingTensorBehavior.SKIP)
+        alignment_mode = kwargs.get('alignment_mode', 'pad/crop')
+
+        if '_tensor_a' in kwargs:
+            a = PreloadedTensor(key, kwargs['_tensor_a'])
+        else:
+            loader_args_a = {"handlers": kwargs['handlers'], "device": kwargs['device'], "dtype": kwargs['dtype']}
+            a = LoadTensor(key, kwargs['model_a'], **loader_args_a)
+
+        loader_args_b = self._get_secondary_loader_args(kwargs, mismatch_mode)
+        b = LoadTensor(key, kwargs['model_b'], **loader_args_b)
+
+        op = DARETIESOp(
+            key,
+            kwargs['alpha'],   # drop_rate
+            kwargs['beta'],    # trim_quantile
+            kwargs['gamma'],   # lambda scale
+            kwargs['seed'],
+            a, b,
+        )
+        op.alignment_mode = alignment_mode
+        return op
+
+
+TWO_MODEL_MODES = [WeightSum(), InterpDifference(), PowerUp(), DARETIESMode(), SVDMode(), ManEnhInterpDifference(), AutoEnhInterpDifference(), WeightSumCutoffMode()]
 THREE_MODEL_MODES = [AddDifference(), TrainDifference(), Extract(), AddDisimilarity()]

nodes/merger_utils.py

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

nodes/model_info_nodes.py

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

nodes/prunekeys.py

@@ -7,9 +7,10 @@ from comfy_api.latest import io
 from safetensors.torch import save_file
 from safetensors import safe_open
 from .utils import convert_pt_to_safetensors
-from .merger_utils import MemoryEfficientSafeOpen
 from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 
+from unifiedefficientloader import MemoryEfficientSafeOpen
+
 
 def _prune_keys(model_name: str, model_type: str, keys_to_prune_str: str,
                 use_regex: bool, output_filename: str) -> str:

nodes/renamekeys.py

@@ -6,9 +6,11 @@ from comfy_api.latest import io
 from safetensors.torch import save_file
 from safetensors import safe_open
 from .utils import convert_pt_to_safetensors
-from .merger_utils import MemoryEfficientSafeOpen
 from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 
+from unifiedefficientloader import MemoryEfficientSafeOpen
+
+
 
 def _rename_keys(model_name: str, model_type: str, old_keys_str: str,
                  new_keys_str: str, output_filename: str) -> str:

pyproject.toml

@@ -1,9 +1,24 @@
 [project]
 name = "ComfyUI-ModelUtils"
 description = "[WIP]Custom nodes for handling, inspecting, modifying and creating various model files."
-version = "0.2.9"
+version = "0.3.5"
 license = { file = "LICENSE" }
 
+classifiers = [
+    "Development Status :: 5 - Production/Stable",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+    "unifiedefficientloader==0.4.4"
+]
+
 [project.urls]
 Repository = "https://github.com/silveroxides/ComfyUI-ModelUtils"
 #  Used by Comfy Registry https://comfyregistry.org

requirements.txt

@@ -1,5 +1,5 @@
 requests
 Pillow
 mutagen
-unifiedefficientloader
+unifiedefficientloader==0.4.4
 av