Add ability to adjust paramaters on a ler layer basis using regex/substring for layer colon separated with parameter=float for specific parameters

Feature/preview meta wf loader

__init__.py

@@ -29,6 +29,14 @@ from .nodes.lora_resize import (
     LoRAResizeFrobenius, LoRAResizeCumulative,
     LoRAMergeToModel
 )
+from .nodes.downloader_nodes import (
+    CheckpointInfoMetaDownloader, DiffusionModelInfoMetaDownloader, LoRAInfoMetaDownloader, EmbeddingInfoMetaDownloader,
+    VAEInfoMetaDownloader, ControlNetInfoMetaDownloader, ManualPathInfoMetaDownloader
+)
+from .nodes.model_info_nodes import (
+    CheckpointInfoLoader, LoRAInfoLoader, EmbeddingInfoLoader,
+    VAEInfoLoader, ControlNetInfoLoader, DiffusionModelInfoLoader
+)
 
 
 class ModelUtilsExtension(ComfyExtension):
@@ -60,6 +68,12 @@ class ModelUtilsExtension(ComfyExtension):
             LoRAMultiMerge, LoRAMultiMergeDARE,
             # LoRA Merge To Model
             LoRAMergeToModel,
+            # Downloaders
+            CheckpointInfoMetaDownloader, DiffusionModelInfoMetaDownloader, LoRAInfoMetaDownloader, EmbeddingInfoMetaDownloader,
+            VAEInfoMetaDownloader, ControlNetInfoMetaDownloader, ManualPathInfoMetaDownloader,
+            # Info Loaders
+            CheckpointInfoLoader, LoRAInfoLoader, EmbeddingInfoLoader,
+            VAEInfoLoader, ControlNetInfoLoader, DiffusionModelInfoLoader,
         ]
 
 

docs/merger_2_model_modes.md

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

nodes/downloader_nodes.py

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

nodes/downloader_utils.py

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

nodes/lora_extract_svd.py

@@ -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
@@ -46,14 +49,14 @@ def _index_sv_ratio(S: torch.Tensor, ratio: float) -> int:
 
 def _index_sv_cumulative(S: torch.Tensor, target: float, max_rank: int = None) -> int:
     """Cumulative mode - keep enough SVs to reach target % of total.
-    
+
     Calculates relative to max_rank if provided, otherwise relative to full.
     """
     if max_rank is not None and max_rank < len(S):
         total = torch.sum(S[:max_rank])
     else:
         total = torch.sum(S)
-    
+
     if total < 1e-8:
         return 1
     cumsum = torch.cumsum(S, dim=0) / total
@@ -63,7 +66,7 @@ def _index_sv_cumulative(S: torch.Tensor, target: float, max_rank: int = None) -
 
 def _index_sv_fro(S: torch.Tensor, target: float, max_rank: int = None) -> int:
     """Frobenius norm mode - preserve target fraction of Frobenius norm.
-    
+
     Calculates relative to max_rank if provided, otherwise relative to full.
     This means "retain target% of what's achievable within max_rank".
     """
@@ -75,10 +78,10 @@ def _index_sv_fro(S: torch.Tensor, target: float, max_rank: int = None) -> int:
     else:
         S_sq = S.pow(2)
         total_sq = torch.sum(S_sq)
-    
+
     if total_sq < 1e-8:
         return 1
-    
+
     # Cumsum of all S (not capped) to find where we reach target
     cumsum = torch.cumsum(S.pow(2), dim=0) / total_sq
     rank = int(torch.searchsorted(cumsum, target ** 2).item()) + 1
@@ -172,27 +175,27 @@ def _svd_extract_linear_lowrank(
 ) -> tuple:
     """
     Low-rank SVD decomposition using svd_lowrank for fixed rank extraction.
-    
+
     Much faster than full SVD when rank << min(m, n) because it only computes
     the top-k singular values using randomized algorithms.
-    
+
     Args:
         niter: Power iterations for SVD accuracy (higher = more accurate but slower)
-    
+
     Returns:
         (lora_down, lora_up, diff), "low rank" or (weight, "full")
     """
     weight = weight.to(device=device, dtype=torch.float32)
     out_dim, in_dim = weight.shape
-    
+
     # Clamp rank to valid range
     max_possible_rank = min(out_dim, in_dim)
-    rank = max(1, min(rank, max_possible_rank - 1))
+    rank = max(1, min(rank, max_possible_rank))
-    
+
     # Check if decomposition is worthwhile
-    if rank >= max_possible_rank // 2:
+    if rank >= max_possible_rank:
         return weight, "full"
-    
+
     # Use svd_lowrank for faster low-rank approximation
     try:
         U, S, V = torch.svd_lowrank(weight, q=rank, niter=niter)
@@ -200,7 +203,7 @@ def _svd_extract_linear_lowrank(
         Vh = V.T  # [rank, in_dim]
     except Exception as e:
         raise RuntimeError(f"svd_lowrank failed: {e}")
-    
+
     # Clamp outliers
     if clamp_quantile < 1.0 and len(S) > 0:
         try:
@@ -208,14 +211,14 @@ def _svd_extract_linear_lowrank(
             S = S.clamp(max=max_val)
         except RuntimeError:
             pass
-    
+
     # Construct LoRA matrices
     lora_up = U @ torch.diag(S)  # [out_dim, rank]
     lora_down = Vh               # [rank, in_dim]
-    
+
     # Compute reconstruction diff
     diff = weight - (lora_up @ lora_down)
-    
+
     return (lora_down, lora_up, diff), "low rank"
 
 
@@ -239,7 +242,7 @@ def _svd_extract_linear(
     """
     weight = weight.to(device=device, dtype=torch.float32)
     out_dim, in_dim = weight.shape
-    
+
     # For fixed rank, use the faster svd_lowrank
     if mode == "fixed" and mode_param > 0:
         target_rank = int(mode_param)
@@ -257,7 +260,7 @@ def _svd_extract_linear(
     rank = _compute_rank(S, mode, mode_param, max_rank)
 
     # Check if decomposition is worthwhile
-    if rank >= min(out_dim, in_dim) // 2:
+    if rank >= min(out_dim, in_dim):
         return weight, "full"
 
     # Truncate to rank
@@ -303,7 +306,7 @@ def _svd_extract_conv(
 
     # Flatten for SVD
     if is_1x1:
-        mat = weight.squeeze()  # [out_ch, in_ch]
+        mat = weight.view(out_ch, in_ch)  # [out_ch, in_ch]
     else:
         mat = weight.reshape(out_ch, -1)  # [out_ch, in_ch*k*k]
 
@@ -382,8 +385,11 @@ def _extract_chunked_layer(
             return None, None, 0
 
     # Combine chunks
+    # Note: Using the first chunk's down matrix as the basis for all chunks
+    # is a better approximation than a simple mean if the chunks share the same input space (like QKV).
+    # Ideally, we would use a more sophisticated joint SVD, but this is a reasonable fallback.
     combined_lora_up = torch.cat(all_lora_up, dim=0)
-    combined_lora_down = torch.stack(all_lora_down, dim=0).mean(dim=0)
+    combined_lora_down = all_lora_down[0]
     rank = combined_lora_down.shape[0]
 
     return combined_lora_up, combined_lora_down, rank
@@ -433,6 +439,8 @@ def extract_lora_from_files(
     mismatch_mode: str = "skip",
     chunk_large_layers: bool = True,
     svd_niter: int = 2,
+    lazy_load: bool = True,
+    force_clear_cache: bool = True,
 ) -> dict[str, torch.Tensor]:
     """
     Extract LoRA from difference between two models.
@@ -468,45 +476,46 @@ def extract_lora_from_files(
     print(f"[LoRA Extract] Preparing memory for {total_size_gb:.2f}GB operation...")
     prepare_for_large_operation(total_size_gb * 1.5, torch.device(device))  # 1.5x for SVD headroom
 
-    handler_a = MemoryEfficientSafeOpen(model_a_path)
+    handler_a = MemoryEfficientSafeOpen(model_a_path, low_memory=lazy_load)
-    handler_b = MemoryEfficientSafeOpen(model_b_path)
+    handler_b = MemoryEfficientSafeOpen(model_b_path, low_memory=lazy_load)
 
     try:
         keys_a = set(handler_a.keys())
         keys_b = set(handler_b.keys())
-
         weight_keys = [k for k in keys_a if k.endswith(".weight")]
-
         pbar = comfy.utils.ProgressBar(len(weight_keys))
-
         stats = {"extracted": 0, "full": 0, "skipped": 0, "chunked": 0}
 
-        for key in tqdm(weight_keys, desc="Extracting LoRA", unit="layers"):
+        def _process_layer(key):
             lora_name = _format_lora_key(key)
 
             if _matches_any_pattern(key, skip_patterns):
-                stats["skipped"] += 1
+                return "skipped", None
-                pbar.update(1)
-                continue
 
             # Load tensors with pinned memory for CUDA
             use_pinned = device == 'cuda'
-            
+
             if key not in keys_b:
                 if mismatch_mode == "skip":
-                    stats["skipped"] += 1
+                    return "skipped", None
-                    pbar.update(1)
+                if mismatch_mode == "error":
-                    continue
+                    raise ValueError(f"Key {key} not found in model B")
+
                 cpu_a = handler_a.get_tensor(key)
                 if use_pinned:
                     weight_diff = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
                 else:
                     weight_diff = cpu_a.to(device=device, dtype=torch.float32)
                 del cpu_a
+
+                if mismatch_mode == "zeros":
+                    # For zeros mode, we treat the missing base as a zeroed tensor of same shape
+                    # which is already captured by weight_diff = tensor_a
+                    pass
             else:
                 cpu_a = handler_a.get_tensor(key)
                 cpu_b = handler_b.get_tensor(key)
-                
+
                 if use_pinned:
                     tensor_a = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
                     tensor_b = transfer_to_gpu_pinned(cpu_b, device, torch.float32)
@@ -517,10 +526,12 @@ def extract_lora_from_files(
 
                 if tensor_a.shape != tensor_b.shape:
                     if mismatch_mode == "skip":
-                        stats["skipped"] += 1
-                        pbar.update(1)
                         del tensor_a, tensor_b
-                        continue
+                        return "skipped", None
+                    if mismatch_mode == "error":
+                        raise ValueError(f"Shape mismatch for {key}: {tensor_a.shape} vs {tensor_b.shape}")
+
+                    # For zeros/fallback, use tensor_a as the difference
                     weight_diff = tensor_a
                     del tensor_b
                 else:
@@ -529,19 +540,16 @@ def extract_lora_from_files(
 
             # Skip small differences
             if min_diff > 0 and weight_diff.abs().max() < min_diff:
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
+                return "skipped", None
-                continue
 
             # Skip 1D tensors
             if weight_diff.ndim < 2:
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
+                return "skipped", None
-                continue
 
             is_conv = weight_diff.ndim == 4
+            layer_results = {}
 
             try:
                 if is_conv:
@@ -562,32 +570,42 @@ def extract_lora_from_files(
                             weight_diff, num_chunks, mode, linear_param, device, linear_max_rank
                         )
                         if lora_up is not None:
-                            output_sd[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.lora_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()
+                            layer_results[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)
+                            layer_results[f"{lora_name}.alpha"] = torch.tensor(rank).to(save_torch_dtype)
-                            stats["chunked"] += 1
                             del weight_diff
-                            pbar.update(1)
+                            return "chunked", layer_results
-                            continue
 
                 print(f"[LoRA Extract] Failed: {key}: {e}")
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
+                return "skipped", None
-                continue
 
             # Store result
             if mode_str == "full":
-                output_sd[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
-                stats["full"] += 1
+                status = "full"
             else:
                 lora_down, lora_up, _ = result
-                output_sd[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.lora_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()
+                layer_results[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)
+                layer_results[f"{lora_name}.alpha"] = torch.tensor(lora_down.shape[0]).to(save_torch_dtype)
-                stats["extracted"] += 1
+                status = "extracted"
 
             del weight_diff
+            return status, layer_results
+
+        for key in tqdm(weight_keys, desc="Extracting LoRA", unit="layers"):
+            status, layer_sd = _process_layer(key)
+            stats[status] += 1
+            if layer_sd:
+                output_sd.update(layer_sd)
+
+            if force_clear_cache:
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
             pbar.update(1)
 
         print(f"[LoRA Extract] Done: {stats['extracted']} extracted, {stats['chunked']} chunked, "
@@ -674,9 +692,13 @@ def _format_lora_key(key: str) -> str:
         key = key[:-7]
 
     # 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
@@ -698,7 +720,7 @@ def _format_lora_key(key: str) -> str:
     # 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}"
@@ -709,6 +731,8 @@ def _format_lora_key(key: str) -> str:
 
 def _get_common_inputs():
     return [
+        io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+        io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
         io.Boolean.Input("chunk_large_layers", default=False,
                         tooltip="Split large fused layers (QKV, MLP) into chunks"),
         io.Float.Input("clamp_quantile", default=0.99, min=0.5, max=1.0, step=0.01,
@@ -736,9 +760,9 @@ class LoRAExtractFixed(io.ComfyNode):
             description="Extract LoRA with specified fixed rank for each layer type.",
             inputs=[
                 *_get_model_inputs(),
-                io.Int.Input("linear_dim", default=64, min=1, max=4096,
+                io.Int.Input("linear_dim", default=64, min=1, max=16384,
                             tooltip="Rank for linear/attention layers"),
-                io.Int.Input("conv_dim", default=32, min=1, max=4096,
+                io.Int.Input("conv_dim", default=32, min=1, max=16384,
                             tooltip="Rank for conv layers"),
                 io.Int.Input("svd_niter", default=2, min=0, max=10,
                             tooltip="SVD power iterations (higher = more accurate but slower)"),
@@ -751,7 +775,7 @@ class LoRAExtractFixed(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, linear_dim, conv_dim, svd_niter, chunk_large_layers,
                 clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
@@ -759,7 +783,8 @@ class LoRAExtractFixed(io.ComfyNode):
         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
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers, svd_niter,
+            lazy_load, force_clear_cache
         )
 
         return io.NodeOutput(_save_lora(output_sd, output_filename))
@@ -781,9 +806,9 @@ class LoRAExtractRatio(io.ComfyNode):
                               tooltip="Ratio threshold for linear layers (higher = more SVs kept)"),
                 io.Float.Input("conv_ratio", default=2.0, min=1.0, max=100.0, step=0.1,
                               tooltip="Ratio threshold for conv layers (higher = more SVs kept)"),
-                io.Int.Input("linear_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for linear layers"),
-                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -794,7 +819,7 @@ class LoRAExtractRatio(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, linear_ratio, conv_ratio, linear_max_rank, conv_max_rank,
                 chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
@@ -802,7 +827,8 @@ class LoRAExtractRatio(io.ComfyNode):
         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
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache
         )
 
         return io.NodeOutput(_save_lora(output_sd, output_filename))
@@ -824,9 +850,9 @@ class LoRAExtractQuantile(io.ComfyNode):
                               tooltip="Target cumulative % for linear layers"),
                 io.Float.Input("conv_quantile", default=0.9, min=0.0, max=1.0, step=0.01,
                               tooltip="Target cumulative % for conv layers"),
-                io.Int.Input("linear_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for linear layers"),
-                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -837,7 +863,7 @@ class LoRAExtractQuantile(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, linear_quantile, conv_quantile, linear_max_rank, conv_max_rank,
                 chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
@@ -845,7 +871,8 @@ class LoRAExtractQuantile(io.ComfyNode):
         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
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache
         )
 
         return io.NodeOutput(_save_lora(output_sd, output_filename))
@@ -865,9 +892,9 @@ class LoRAExtractKnee(io.ComfyNode):
                 *_get_model_inputs(),
                 io.Combo.Input("knee_method", options=["sv_knee", "sv_cumulative_knee"],
                               default="sv_knee", tooltip="Knee detection method"),
-                io.Int.Input("linear_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for linear layers"),
-                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -878,7 +905,7 @@ class LoRAExtractKnee(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, knee_method, linear_max_rank, conv_max_rank,
                 chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
@@ -886,7 +913,8 @@ class LoRAExtractKnee(io.ComfyNode):
         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
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache
         )
 
         return io.NodeOutput(_save_lora(output_sd, output_filename))
@@ -908,9 +936,9 @@ class LoRAExtractFrobenius(io.ComfyNode):
                               tooltip="Target Frobenius norm fraction for linear"),
                 io.Float.Input("conv_target", default=0.9, min=0.0, max=1.0, step=0.01,
                               tooltip="Target Frobenius norm fraction for conv"),
-                io.Int.Input("linear_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for linear layers"),
-                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -921,7 +949,7 @@ class LoRAExtractFrobenius(io.ComfyNode):
     @classmethod
     def execute(cls, model_a, model_b, linear_target, conv_target, linear_max_rank, conv_max_rank,
                 chunk_large_layers, clamp_quantile, min_diff, mismatch_mode, output_filename,
-                save_dtype, device, skip_patterns) -> io.NodeOutput:
+                save_dtype, device, skip_patterns, lazy_load, force_clear_cache) -> io.NodeOutput:
 
         model_a_path = folder_paths.get_full_path_or_raise("diffusion_models", model_a)
         model_b_path = folder_paths.get_full_path_or_raise("diffusion_models", model_b)
@@ -929,7 +957,8 @@ class LoRAExtractFrobenius(io.ComfyNode):
         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
+            clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
+            lazy_load=lazy_load, force_clear_cache=force_clear_cache
         )
 
         return io.NodeOutput(_save_lora(output_sd, output_filename))

nodes/lora_merger.py

@@ -10,17 +10,21 @@ 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 .merger_utils import parse_custom_weights, get_custom_parameters
 from .lora_resize import (
-    detect_lora_format, 
+    detect_lora_format,
-    extract_lora_pairs, 
+    extract_lora_pairs,
     detect_lora_rank,
     _compile_patterns,
     _matches_any_pattern,
     _format_lora_key
 )
 
+from unifiedefficientloader import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
+
+
+
 def merge_multi_loras(
     lora_paths: List[str],
     lora_weights: List[float],
@@ -28,11 +32,12 @@ def merge_multi_loras(
     device: str,
     save_dtype: torch.dtype,
     output_filename: str,
+    custom_weights: 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.
@@ -45,22 +50,24 @@ def merge_multi_loras(
     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]
+    parsed_custom_weights = parse_custom_weights(custom_weights)
-    
+
+    handlers = [MemoryEfficientSafeOpen(p, low_memory=True) 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,
@@ -71,10 +78,10 @@ def merge_multi_loras(
                 "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:
@@ -90,41 +97,51 @@ def merge_multi_loras(
         # 2. Merge each core layer
         with torch.no_grad():
             for core, indices in tqdm(layer_map.items(), desc="Merging layers", unit="layers"):
+                # Apply per-layer overrides (regex/substring matching)
+                layer_overrides = get_custom_parameters(core, parsed_custom_weights, {})
+
                 downs = []
                 ups = []
                 is_conv = False
                 max_rank = 0
-                
+
                 # First pass: load and determine max rank
                 for idx in indices:
-                    info = lora_infos[idx]
+                    info = lora_infos[idx].copy()
+                    # Apply overrides: w1, w2, w3... or 'weight'
+                    lora_key = f'w{idx+1}'
+                    if lora_key in layer_overrides:
+                        info["weight"] = layer_overrides[lora_key]
+                    elif "weight" in layer_overrides:
+                        info["weight"] = layer_overrides["weight"]
+
                     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))
 
@@ -148,10 +165,10 @@ def merge_multi_loras(
                     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"]
@@ -159,7 +176,7 @@ def merge_multi_loras(
                         # 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)
@@ -168,15 +185,15 @@ def merge_multi_loras(
                             # 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)
@@ -185,12 +202,18 @@ def merge_multi_loras(
                 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
-                
+                downs.clear()
+                ups.clear()
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
                 pbar.update(1)
 
         # 3. Final Summary
@@ -199,7 +222,7 @@ def merge_multi_loras(
             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")
@@ -210,16 +233,16 @@ def merge_multi_loras(
             "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:
@@ -265,10 +288,11 @@ class LoRAMultiMerge(io.ComfyNode):
                 # 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.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights for each LoRA. Syntax: 'regex:w1,w2,w3...' or 'regex:w1=0.5,w2=1.0'"),
                 io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
                 io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
             ],
@@ -280,35 +304,36 @@ class LoRAMultiMerge(io.ComfyNode):
     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:
+                merge_mode, output_filename, custom_weights, 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
+            output_filename=output_filename,
+            custom_weights=custom_weights
         )
-        
+
         return io.NodeOutput(path)
 
 
@@ -347,6 +372,7 @@ class LoRAMultiMergeDARE(io.ComfyNode):
                 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.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights for each LoRA. Syntax: 'regex:w1,w2,w3...' or 'regex:w1=0.5,w2=1.0'"),
                 io.Combo.Input("save_dtype", options=["fp16", "bf16", "fp32"], default="fp16"),
                 io.Combo.Input("device", options=["cuda", "cpu"], default="cuda"),
             ],
@@ -358,26 +384,26 @@ class LoRAMultiMergeDARE(io.ComfyNode):
     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:
+                drop_rate, trim_quantile, seed, output_filename, custom_weights, 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,
@@ -386,9 +412,10 @@ class LoRAMultiMergeDARE(io.ComfyNode):
             seed=seed,
             device=device,
             save_dtype=dtype,
-            output_filename=output_filename
+            output_filename=output_filename,
+            custom_weights=custom_weights
         )
-        
+
         return io.NodeOutput(path)
 
 def merge_multi_loras_dare(
@@ -400,6 +427,7 @@ def merge_multi_loras_dare(
     device: str,
     save_dtype: torch.dtype,
     output_filename: str,
+    custom_weights: str = "",
     verbose: bool = True,
 ) -> str:
     """
@@ -412,8 +440,11 @@ def merge_multi_loras_dare(
         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]
+
-    
+    parsed_custom_weights = parse_custom_weights(custom_weights)
+
+    handlers = [MemoryEfficientSafeOpen(p, low_memory=True) for p in lora_paths]
+
     rng = torch.Generator(device=device).manual_seed(seed)
 
     try:
@@ -433,7 +464,7 @@ def merge_multi_loras_dare(
                 "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}")
 
@@ -451,19 +482,29 @@ def merge_multi_loras_dare(
         # 2. Merge
         with torch.no_grad():
             for core, indices in tqdm(layer_map.items(), desc="Merging layers (DARE)", unit="layers"):
+                # Apply per-layer overrides (regex/substring matching)
+                layer_overrides = get_custom_parameters(core, parsed_custom_weights, {})
+
                 downs = []
                 ups = []
                 max_rank = 0
-                
+
                 for idx in indices:
-                    info = lora_infos[idx]
+                    info = lora_infos[idx].copy()
+                    # Apply overrides: w1, w2, w3... or 'weight'
+                    lora_key = f'w{idx+1}'
+                    if lora_key in layer_overrides:
+                        info["weight"] = layer_overrides[lora_key]
+                    elif "weight" in layer_overrides:
+                        info["weight"] = layer_overrides["weight"]
+
                     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"]))
@@ -484,33 +525,37 @@ def merge_multi_loras_dare(
                             padding[rev_dim*2 + 1] = max_rank - t.shape[dim_to_pad]
                             t = torch.nn.functional.pad(t, tuple(padding))
                         padded.append(t * w)
-                    
+
+                    # Apply per-layer overrides for DARE/TIES parameters
+                    current_drop_rate = layer_overrides.get('alpha', layer_overrides.get('drop_rate', drop_rate))
+                    current_trim_quantile = layer_overrides.get('beta', layer_overrides.get('trim_quantile', trim_quantile))
+
                     # DARE
-                    if drop_rate > 0:
+                    if current_drop_rate > 0:
                         for i in range(len(padded)):
-                            mask = (torch.rand(padded[i].shape, generator=rng, device=device) > drop_rate).float()
+                            mask = (torch.rand(padded[i].shape, generator=rng, device=device) > current_drop_rate).float()
-                            padded[i] = (padded[i] * mask) / (1 - drop_rate)
+                            padded[i] = (padded[i] * mask) / (1 - current_drop_rate)
-                    
+
                     # TIES
                     # 1. Trim
-                    if trim_quantile > 0:
+                    if current_trim_quantile > 0:
                         for i in range(len(padded)):
                             flat = padded[i].abs().flatten()
-                            k = int(len(flat) * trim_quantile)
+                            k = int(len(flat) * current_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)
@@ -522,7 +567,17 @@ def merge_multi_loras_dare(
                 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)
-                
+
+                del merged_down, merged_up
+                for d, _ in downs: del d
+                for u, _ in ups: del u
+                downs.clear()
+                ups.clear()
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
                 pbar.update(1)
 
         # Final Summary
@@ -531,7 +586,7 @@ def merge_multi_loras_dare(
             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")

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]
 
@@ -461,6 +464,12 @@ def resize_lora_file(
                 if alpha_suffix:
                     output_sd[f"{new_block_name}{alpha_suffix}"] = torch.tensor(result["new_alpha"]).to(save_dtype)
 
+                del result
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
                 pbar.update(1)
 
         if verbose and fro_list:
@@ -854,6 +863,10 @@ def merge_loras_to_model(
                     del cpu_base
                     stats["copied"] += 1
 
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
 
                 pbar.update(1)
 
@@ -886,11 +899,11 @@ class LoRAMergeToModel(io.ComfyNode):
             node_id="LoRAMergeToModel",
             display_name="LoRA Merge To Model",
             category="ModelUtils/LoRA/Merge",
-            description="Merge 1-4 LoRAs into a base model and save the result. Saves to base model directory.",
+            description="Merge 1-8 LoRAs into a base model and save the result. Saves to base model directory.",
             inputs=[
                 io.Combo.Input("base_model", options=folder_paths.get_filename_list("diffusion_models"),
                               tooltip="Base model the LoRAs were trained on"),
-                io.Combo.Input("lora_count", options=["1", "2", "3", "4"], default="2",
+                io.Combo.Input("lora_count", options=["1", "2", "3", "4", "5", "6", "7", "8"], default="2",
                               tooltip="Number of LoRAs to merge"),
                 # LoRA 1
                 io.Combo.Input("lora_1", options=folder_paths.get_filename_list("loras"),
@@ -912,6 +925,26 @@ class LoRAMergeToModel(io.ComfyNode):
                               default="None", tooltip="Fourth LoRA"),
                 io.Float.Input("weight_4", default=1.0, min=-10.0, max=10.0, step=0.01,
                               tooltip="Weight strength for LoRA 4"),
+                # LoRA 5
+                io.Combo.Input("lora_5", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Fifth LoRA"),
+                io.Float.Input("weight_5", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 5"),
+                # LoRA 6
+                io.Combo.Input("lora_6", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Sixth LoRA"),
+                io.Float.Input("weight_6", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 6"),
+                # LoRA 7
+                io.Combo.Input("lora_7", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Seventh LoRA"),
+                io.Float.Input("weight_7", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 7"),
+                # LoRA 8
+                io.Combo.Input("lora_8", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Eighth LoRA"),
+                io.Float.Input("weight_8", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 8"),
                 # Settings
                 io.String.Input("skip_patterns", default="", multiline=True,
                                tooltip="Regex patterns for layers to skip"),
@@ -926,12 +959,13 @@ class LoRAMergeToModel(io.ComfyNode):
     @classmethod
     def execute(cls, base_model, lora_count,
                 lora_1, weight_1, lora_2, weight_2, lora_3, weight_3, lora_4, weight_4,
+                lora_5, weight_5, lora_6, weight_6, lora_7, weight_7, lora_8, weight_8,
                 skip_patterns, output_filename, save_dtype, device) -> io.NodeOutput:
 
         # Build LoRA list based on count
         count = int(lora_count)
-        lora_names = [lora_1, lora_2, lora_3, lora_4][:count]
+        lora_names = [lora_1, lora_2, lora_3, lora_4, lora_5, lora_6, lora_7, lora_8][:count]
-        lora_weights = [weight_1, weight_2, weight_3, weight_4][:count]
+        lora_weights = [weight_1, weight_2, weight_3, weight_4, weight_5, weight_6, weight_7, weight_8][:count]
 
         # Filter out "None" entries
         valid_loras = [(name, weight) for name, weight in zip(lora_names, lora_weights) if name != "None"]

nodes/merger.py

@@ -7,11 +7,13 @@ 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 .merger_utils import parse_custom_weights, get_custom_parameters
 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 +78,9 @@ class MergerLogic:
             process_device = recipe_params.get('device', 'cpu')
             print(f"[Merger] Preparing memory for {total_size_gb:.2f}GB merge operation...")
             prepare_for_large_operation(total_size_gb * 1.2, torch.device(process_device))
+            flush_threshold_bytes = int(total_size_gb * 1024**3 / 16)
+        else:
+            flush_threshold_bytes = 256 * 1024 * 1024  # 256MB fallback (min 8GB RAM assumed)
 
         handlers = {}
         for name in model_names.values():
@@ -83,11 +88,11 @@ class MergerLogic:
                 path = folder_paths.get_full_path(model_type, name)
                 if not path:
                     raise FileNotFoundError(f"Model '{name}' not found.")
-                handlers[name] = MemoryEfficientSafeOpen(path)
+                handlers[name] = MemoryEfficientSafeOpen(path, low_memory=True)
 
         primary_handler = handlers[primary_model_name]
         all_keys = primary_handler.keys()
-        metadata = primary_handler.header.get("__metadata__", {})
+        metadata = primary_handler.metadata()
 
         # Convert mismatch_mode string to enum
         mismatch_mode_str = recipe_params.get('mismatch_mode', 'skip')
@@ -99,6 +104,9 @@ class MergerLogic:
         exclude_patterns = _compile_patterns(recipe_params.get('exclude_patterns', ''))
         discard_patterns = _compile_patterns(recipe_params.get('discard_patterns', ''))
 
+        # Parse custom per-layer weights
+        parsed_custom_weights = parse_custom_weights(recipe_params.get('custom_weights', ''))
+
         # Pre-compute key differences for logging
         primary_keys = set(all_keys)
         for name, handler in handlers.items():
@@ -124,65 +132,127 @@ class MergerLogic:
         discarded_keys = 0
         error_keys = []
 
-        for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
+        output_folder = "loras" if calc_mode == "SVD LoRA Extraction" else model_type
-            # Check discard patterns first - skip entirely
+        # Use [-1] for diffusion_models to get the actual diffusion_models folder, not legacy unet
-            if _matches_any_pattern(key, discard_patterns):
+        output_dir = folder_paths.get_folder_paths(output_folder)[-1]
-                discarded_keys += 1
+        os.makedirs(output_dir, exist_ok=True)
-                pbar.update(1)
+        output_filename = recipe_params.get("output_filename")
-                continue
+        output_path = os.path.join(output_dir, f"{output_filename}.safetensors")
 
-            # Pre-load Model A's tensor with pinned memory for CUDA
+        writer = None
-            cpu_tensor = primary_handler.get_tensor(key)
-            if process_device == 'cuda':
-                tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
-            else:
-                tensor_a = cpu_tensor.to(device=process_device, dtype=process_dtype)
-            del cpu_tensor
 
-            # Check exclude patterns - use Model A only, no merge
+        writer = IncrementalSafetensorsWriter(output_path, metadata=metadata)
-            if _matches_any_pattern(key, exclude_patterns):
+        writer.__enter__()
-                merged_state_dict[key] = tensor_a.to(save_torch_dtype).cpu().clone()
-                excluded_keys += 1
-                pbar.update(1)
-                continue
 
-            # Pass tensor_a metadata to recipes for zeros mode and fallback
+        batch_buffer = {}
-            recipe_params['_tensor_a'] = tensor_a
+        batch_bytes = 0
-            recipe_params['_tensor_a_shape'] = tensor_a.shape
+        batch_key_count = 0
-            recipe_params['_tensor_a_dtype'] = tensor_a.dtype
 
-            try:
+        def _flush_batch():
-                recipe = calc_mode_class.create_recipe(key=key, **recipe_params)
+            nonlocal batch_buffer, batch_bytes, batch_key_count
-                result = recipe.merge()
+            if not batch_buffer:
-            except MissingTensorError as e:
+                return
-                if mismatch_mode == MissingTensorBehavior.ERROR:
+            writer.write_dict(batch_buffer)
-                    raise ValueError(f"Layer mismatch error (mismatch_mode='error'): {e}")
+            pbar.update(batch_key_count)
-                result = None
+            batch_buffer = {}
-                error_keys.append(key)
+            batch_bytes = 0
+            batch_key_count = 0
+            if recipe_params.get('force_clear_cache', True):
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
 
-            # Handle None result (mismatch occurred with skip mode)
+        with torch.no_grad():
-            if result is None:
+            for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
-                result = tensor_a
+                # Check discard patterns first - skip entirely
-                skipped_keys += 1
+                if _matches_any_pattern(key, discard_patterns):
+                    discarded_keys += 1
+                    continue
 
-            if isinstance(result, dict):
+                # Pre-load Model A's tensor with pinned memory for CUDA
-                for r_key, r_tensor in result.items():
+                cpu_tensor = primary_handler.get_tensor(key)
-                    merged_state_dict[r_key] = r_tensor.to(save_torch_dtype).cpu().clone()
+                if process_device == 'cuda':
-            else:
+                    tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
-                # Ensure compatibility with Model A's architecture.
+                else:
-                # If alignment_mode is 'pad/crop', we crop results that were padded.
+                    tensor_a = cpu_tensor.to(device=process_device, dtype=process_dtype)
-                # If alignment_mode is 'interpolate', resizing happened during operators.
+                del cpu_tensor
-                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()
+                # Check exclude patterns - use Model A only, no merge
+                if _matches_any_pattern(key, exclude_patterns):
+                    t = tensor_a.detach().to(save_torch_dtype).cpu()
 
-            # Clean up tensor_a reference to allow GC
+                    batch_buffer[key] = t
-            del recipe_params['_tensor_a']
+                    batch_bytes += t.numel() * t.element_size()
-            pbar.update(1)
+                    batch_key_count += 1
+                    excluded_keys += 1
+                    del tensor_a
+                    if (batch_bytes >= flush_threshold_bytes or batch_key_count >= 32):
+                        _flush_batch()
+                    continue
+
+                # Pass tensor_a metadata to recipes for zeros mode and fallback
+                recipe_params['_tensor_a'] = tensor_a
+                recipe_params['_tensor_a_shape'] = tensor_a.shape
+                recipe_params['_tensor_a_dtype'] = tensor_a.dtype
+
+                try:
+                    # Apply per-layer parameter overrides (regex/substring matching)
+                    layer_params = get_custom_parameters(key, parsed_custom_weights, recipe_params)
+                    recipe = calc_mode_class.create_recipe(key=key, **layer_params)
+                    result = recipe.merge()
+                except MissingTensorError as e:
+                    if mismatch_mode == MissingTensorBehavior.ERROR:
+
+                        _flush_batch()
+                        writer.__exit__(None, None, None)
+                        raise ValueError(f"Layer mismatch error (mismatch_mode='error'): {e}")
+                    result = None
+                    error_keys.append(key)
+
+                # Handle None result (mismatch occurred with skip mode)
+                if result is None:
+                    result = tensor_a
+                    skipped_keys += 1
+
+                if isinstance(result, dict):
+                    for r_key, r_tensor in result.items():
+                        t = r_tensor.detach().to(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.
+                    # If alignment_mode is 'interpolate', resizing happened during operators.
+                    if alignment_mode == 'pad/crop':
+                        target_shape = recipe_params['_tensor_a_shape']
+                        if result.shape != target_shape:
+                            slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(result.shape, target_shape))
+                            result = result[slices]
+
+                    t = result.detach().to(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.
+                recipe.clean()
+                del recipe_params['_tensor_a']
+                del tensor_a
+                del recipe
+                del result
+
+                if (batch_bytes >= flush_threshold_bytes or batch_key_count >= 32):
+                    _flush_batch()
+
+
+        _flush_batch()
+        writer.__exit__(None, None, None)
 
         # Log summary
         if excluded_keys > 0:
@@ -203,10 +273,9 @@ class MergerLogic:
         os.makedirs(output_dir, exist_ok=True)
         output_filename = recipe_params.get("output_filename")
         output_path = os.path.join(output_dir, f"{output_filename}.safetensors")
-        save_file(merged_state_dict, output_path, metadata=metadata)
+        if os.path.exists(output_path):
-
+            print(f"[Merger] Output saved to {output_path}")
         # Cleanup after heavy operation
-        del merged_state_dict
         cleanup_after_operation()
 
         return f"{output_filename}.safetensors"
@@ -230,16 +299,19 @@ class CheckpointTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -251,7 +323,9 @@ class CheckpointTwoMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -263,6 +337,8 @@ class CheckpointTwoMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -285,16 +361,19 @@ class ModelTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -306,7 +385,9 @@ class ModelTwoMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -318,6 +399,8 @@ class ModelTwoMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -340,16 +423,19 @@ class TextEncoderTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -361,7 +447,9 @@ class TextEncoderTwoMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -373,6 +461,8 @@ class TextEncoderTwoMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -395,16 +485,19 @@ class LoRATwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -416,7 +509,9 @@ class LoRATwoMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -428,6 +523,8 @@ class LoRATwoMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -450,16 +547,19 @@ class EmbeddingTwoMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in TWO_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.99, min=0.0, max=1.0, step=0.001),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -471,7 +571,9 @@ class EmbeddingTwoMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_2_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -483,6 +585,8 @@ class EmbeddingTwoMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b}
         filename = MergerLogic.execute_merge(model_names, calc_mode, TWO_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -508,16 +612,19 @@ class CheckpointThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -529,7 +636,9 @@ class CheckpointThreeMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -541,6 +650,8 @@ class CheckpointThreeMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
         filename = MergerLogic.execute_merge(model_names, calc_mode, THREE_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -564,16 +675,19 @@ class ModelThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -585,7 +699,9 @@ class ModelThreeMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -597,6 +713,8 @@ class ModelThreeMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
         filename = MergerLogic.execute_merge(model_names, calc_mode, THREE_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -620,16 +738,19 @@ class TextEncoderThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.99, min=-10.0, max=10.0, step=0.001),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -641,7 +762,9 @@ class TextEncoderThreeMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -653,6 +776,8 @@ class TextEncoderThreeMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
         filename = MergerLogic.execute_merge(model_names, calc_mode, THREE_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -676,16 +801,19 @@ class LoRAThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -697,7 +825,9 @@ class LoRAThreeMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -709,6 +839,8 @@ class LoRAThreeMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
         filename = MergerLogic.execute_merge(model_names, calc_mode, THREE_MODEL_MODES, recipe_params, cls.MODEL_TYPE)
@@ -732,16 +864,19 @@ class EmbeddingThreeMerger(io.ComfyNode):
                 io.Combo.Input("calc_mode", options=[m.name for m in THREE_MODEL_MODES]),
                 io.Combo.Input("mismatch_mode", options=["skip", "zeros", "error"], default="skip"),
                 io.Combo.Input("alignment_mode", options=["pad/crop", "interpolate"], default="pad/crop"),
-                io.Float.Input("alpha", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("alpha", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("beta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("beta", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("gamma", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("gamma", default=0.5, min=-10.0, max=10.0, step=0.01),
-                io.Float.Input("delta", default=0.5, min=-2.0, max=3.0, step=0.01),
+                io.Float.Input("delta", default=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"]),
                 io.Combo.Input("process_device", options=["cuda", "cpu"]),
                 io.String.Input("exclude_patterns", default="", multiline=True),
                 io.String.Input("discard_patterns", default="", multiline=True),
+                io.String.Input("custom_weights", default="", multiline=True, tooltip="Per-layer weights/params. Syntax: 'regex:weight' or 'regex:param=val,param2=val'"),
+                io.Boolean.Input("lazy_load", default=True, tooltip="Low memory mode: load tensors from disk on demand"),
+                io.Boolean.Input("force_clear_cache", default=True, tooltip="Clear CUDA cache after each layer"),
             ],
             outputs=[
                 io.String.Output(display_name="output_filename"),
@@ -753,7 +888,9 @@ class EmbeddingThreeMerger(io.ComfyNode):
     def execute(cls, execution_mode: str, model_a: str, model_b: str, model_c: str,
                 calc_mode: str, mismatch_mode: str, alignment_mode: str, alpha: float, beta: float,
                 gamma: float, delta: float, seed: int, output_filename: str, save_dtype: str,
-                process_device: str, exclude_patterns: str, discard_patterns: str) -> io.NodeOutput:
+                process_device: str, exclude_patterns: str, discard_patterns: str,
+                custom_weights: str,
+                lazy_load: bool, force_clear_cache: bool) -> io.NodeOutput:
         doc = load_documentation_from_file('merger_3_model_modes.md')
         if execution_mode == "DOCUMENTATION ONLY":
             return io.NodeOutput("Documentation mode active. No merge performed.", doc)
@@ -765,6 +902,8 @@ class EmbeddingThreeMerger(io.ComfyNode):
             "output_filename": output_filename, "save_dtype": save_dtype,
             "device": process_device, "dtype": torch.float32,
             "exclude_patterns": exclude_patterns, "discard_patterns": discard_patterns,
+            "custom_weights": custom_weights,
+            "lazy_load": lazy_load, "force_clear_cache": force_clear_cache,
         }
         model_names = {"model_a": model_a, "model_b": model_b, "model_c": model_c}
         filename = MergerLogic.execute_merge(model_names, calc_mode, THREE_MODEL_MODES, recipe_params, cls.MODEL_TYPE)

nodes/merger_ops.py

@@ -4,6 +4,8 @@ import torch.nn.functional as F
 from enum import Enum
 
 
+from unifiedefficientloader import transfer_to_gpu_pinned
+
 class MissingTensorBehavior(Enum):
     """Controls behavior when a tensor key is missing from a model."""
     ERROR = "error"      # Raise exception (strict mode)
@@ -127,12 +129,21 @@ class Operation:
         raise NotImplementedError
 
     def recurse(self, operation):
-        source_tensors = [source_oper.merge() for source_oper in operation.sources]
+        self._source_tensors = [source_oper.merge() for source_oper in operation.sources]
-        return operation.oper(*source_tensors)
+        return operation.oper(*self._source_tensors)
 
     def merge(self):
         return self.merge_func(self)
 
+    def clean(self):
+        if hasattr(self, '_source_tensors'):
+            for t in self._source_tensors:
+                del t
+            del self._source_tensors
+        for source in self.sources:
+            if hasattr(source, 'clean'):
+                source.clean()
+
 class LoadTensor(Operation):
     def __init__(self, key, model_name, handlers, device, dtype,
                  on_missing=MissingTensorBehavior.ERROR, fallback_shape=None, fallback_dtype=None):
@@ -165,7 +176,17 @@ class LoadTensor(Operation):
                 dtype = self.fallback_dtype if self.fallback_dtype else self.dtype
                 return torch.zeros(self.fallback_shape, device=self.device, dtype=dtype)
 
-        return handler.get_tensor(self.key).to(device=self.device, dtype=self.dtype)
+        cpu_tensor = handler.get_tensor(self.key)
+        if self.device == 'cuda':
+            self._tensor = transfer_to_gpu_pinned(cpu_tensor, self.device, self.dtype)
+        else:
+            self._tensor = cpu_tensor.to(device=self.device, dtype=self.dtype)
+        del cpu_tensor
+        return self._tensor
+
+    def clean(self):
+        if hasattr(self, '_tensor'):
+            del self._tensor
 
 class Multiply(Operation):
     def __init__(self, key, alpha, *sources):
@@ -298,6 +319,53 @@ class PowerUpOp(Operation):
         return (m * delta) / keep_prob
 
 
+class DARETIESOp(Operation):
+    """DARE + TIES combined operation on a single task vector (delta = b - a).
+
+    Step 1 — DARE: randomly drop `drop_rate` fraction of delta parameters,
+                   rescale survivors by 1/(1-drop_rate).
+    Step 2 — TIES trim: zero out parameters below the `trim_quantile` magnitude threshold.
+    Step 3 — TIES elect: determine dominant sign per parameter position.
+    Step 4 — TIES merge: zero out parameters that disagree with the dominant sign.
+
+    For a two-model merge the TIES elect/merge step degenerates to a no-op (single vector,
+    sign is always self-consistent) but the trim step still reduces interference from
+    near-zero noisy deltas. The dominant sign filter is included for correctness and
+    forward compatibility with n-model extensions.
+    """
+    def __init__(self, key, alpha, beta, gamma, seed, *sources):
+        super().__init__(key, *sources)
+        # alpha = drop_rate (DARE), beta = trim_quantile (TIES), gamma = lambda scale
+        self.alpha, self.beta, self.gamma, self.seed = alpha, beta, gamma, seed
+
+    def oper(self, a, b):
+        if a is None or b is None:
+            return None
+        a, b = resize_tensors(a, b, mode=self.alignment_mode)
+        delta = (b - a).float()
+
+        # --- DARE: drop and rescale ---
+        keep_prob = 1.0 - self.alpha
+        if keep_prob <= 0:
+            return torch.zeros_like(a)
+        rng = torch.Generator(device=delta.device).manual_seed(self.seed)
+        dare_mask = (torch.empty_like(delta).uniform_(0, 1, generator=rng) < keep_prob).float()
+        delta = (delta * dare_mask) / keep_prob
+
+        # --- TIES trim: zero parameters below magnitude threshold ---
+        if self.beta > 0:
+            flat_abs = delta.abs().flatten()
+            k = max(1, int(flat_abs.numel() * self.beta))
+            threshold = torch.kthvalue(flat_abs, k).values
+            delta = torch.where(delta.abs() < threshold, torch.zeros_like(delta), delta)
+
+        # --- TIES elect: dominant sign (trivially self-consistent for single vector) ---
+        dominant_sign = torch.sign(delta)
+        delta = torch.where(torch.sign(delta) == dominant_sign, delta, torch.zeros_like(delta))
+
+        return (a + self.gamma * delta).to(a.dtype)
+
+
 class InterpolateDifference(Operation):
     def __init__(self, key, alpha, beta, gamma, seed, *sources):
         super().__init__(key, *sources)
@@ -409,6 +477,10 @@ class PreloadedTensor(Operation):
     def merge(self) -> torch.Tensor:
         return self.tensor
 
+    def clean(self):
+        if hasattr(self, 'tensor'):
+            del self.tensor
+
 class WeightSum(CalcMode):
     name = 'Weight-Sum'
     description = 'A * (1 - α) + B * α. Simple linear interpolation.'
@@ -685,5 +757,41 @@ 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

@@ -1,373 +1,8 @@
 """
-Memory-efficient safetensors utilities with parallel I/O and pinned memory support.
+Utilities for model merging operations.
-
-Based on benchmarking, parallel reading with ThreadPoolExecutor provides 2-4x speedup
-for large models compared to sequential reads.
-
-Worker count is automatically optimized based on device capabilities when workers=None.
-Pinned memory enables faster CPU→GPU transfers (2-3x improvement).
 """
-import os
+from typing import Dict, List, Tuple, Any
-import mmap
+import re
-import json
-import struct
-import torch
-from concurrent.futures import ThreadPoolExecutor, as_completed
-from typing import Dict, List, Optional, Any, Tuple
-
-# Import ComfyUI's memory pinning functions if available
-try:
-    from comfy.model_management import pin_memory, unpin_memory, is_device_cuda
-    PINNING_AVAILABLE = True
-except ImportError:
-    PINNING_AVAILABLE = False
-    def pin_memory(tensor): return False
-    def unpin_memory(tensor): return False
-    def is_device_cuda(device): return str(device).startswith('cuda')
-
-
-class PinnedTensor:
-    """Context manager for automatic pinned memory management.
-
-    Usage:
-        with PinnedTensor(cpu_tensor) as pinned:
-            gpu_tensor = pinned.to('cuda')
-        # Memory automatically unpinned after transfer
-    """
-
-    def __init__(self, tensor: torch.Tensor, auto_pin: bool = True):
-        self.tensor = tensor
-        self.is_pinned = False
-        self.auto_pin = auto_pin and PINNING_AVAILABLE
-
-    def __enter__(self) -> torch.Tensor:
-        if self.auto_pin and self.tensor.device.type == 'cpu':
-            self.is_pinned = pin_memory(self.tensor)
-        return self.tensor
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        if self.is_pinned:
-            unpin_memory(self.tensor)
-            self.is_pinned = False
-
-
-def transfer_to_gpu_pinned(
-    tensor: torch.Tensor,
-    device: str = 'cuda',
-    dtype: Optional[torch.dtype] = None
-) -> torch.Tensor:
-    """Transfer tensor to GPU using pinned memory for faster transfer.
-
-    Args:
-        tensor: CPU tensor to transfer
-        device: Target GPU device
-        dtype: Optional dtype conversion
-
-    Returns:
-        Tensor on GPU device
-    """
-    if not PINNING_AVAILABLE or not is_device_cuda(torch.device(device)):
-        # Fall back to regular transfer
-        if dtype:
-            return tensor.to(device=device, dtype=dtype)
-        return tensor.to(device=device)
-
-    # Pin memory for faster transfer
-    was_pinned = pin_memory(tensor)
-    try:
-        if dtype:
-            result = tensor.to(device=device, dtype=dtype, non_blocking=True)
-        else:
-            result = tensor.to(device=device, non_blocking=True)
-        # Sync to ensure transfer is complete before unpinning
-        if was_pinned:
-            torch.cuda.current_stream().synchronize()
-        return result
-    finally:
-        if was_pinned:
-            unpin_memory(tensor)
-
-
-class MemoryEfficientSafeOpen:
-    """Memory-efficient safetensors file reader with mmap, parallel I/O, and pinned memory support.
-
-    Features:
-    - mmap mode: Zero-copy tensor access via memory-mapped file
-    - low_memory mode: Direct file reads to minimize OS page cache usage
-    - Parallel loading: Multi-threaded tensor reads for 2-4x speedup
-    - Sorted batch reads: Keys sorted by file offset for sequential I/O
-    - Auto-optimized workers: Adjusts parallelism based on device capabilities
-    - Pinned memory: Faster CPU→GPU transfers when transferring to CUDA
-
-    Args:
-        filename: Path to safetensors file
-        device: Target device (default 'cpu')
-        mmap_mode: Use memory-mapped file for zero-copy (default True)
-        low_memory: Use direct file reads to minimize memory (overrides mmap_mode)
-    """
-
-    def __init__(self, filename: str, device: str = 'cpu', mmap_mode: bool = True, low_memory: bool = False):
-        self.filename = filename
-        self.device = device
-        # low_memory mode forces mmap off to avoid OS page caching
-        self.low_memory = low_memory
-        self.mmap_mode = mmap_mode and not low_memory
-        self.header, self.header_size = self._read_header()
-        self.file = open(filename, "rb")
-        self.mmap_obj = None
-
-        if self.mmap_mode:
-            self.mmap_obj = mmap.mmap(self.file.fileno(), 0, access=mmap.ACCESS_READ)
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        if self.mmap_obj:
-            self.mmap_obj.close()
-        self.file.close()
-
-
-    def keys(self) -> List[str]:
-        """Return all tensor keys (excluding metadata)."""
-        return [k for k in self.header.keys() if k != "__metadata__"]
-
-    def metadata(self) -> Dict[str, str]:
-        """Return file metadata."""
-        return self.header.get("__metadata__", {})
-
-    def keys_sorted_by_offset(self) -> List[str]:
-        """Return keys sorted by file offset for optimal sequential I/O."""
-        keys_with_offsets = []
-        for key in self.keys():
-            offset = self.header[key]["data_offsets"][0]
-            keys_with_offsets.append((key, offset))
-        keys_with_offsets.sort(key=lambda x: x[1])
-        return [k for k, _ in keys_with_offsets]
-
-    def get_tensor(self, key: str) -> torch.Tensor:
-        """Load a single tensor by key."""
-        if key not in self.header:
-            raise KeyError(f"Tensor '{key}' not found in the file")
-
-        metadata = self.header[key]
-        offset_start, offset_end = metadata["data_offsets"]
-
-        if self.mmap_mode and self.mmap_obj:
-            if offset_start != offset_end:
-                start = self.header_size + 8 + offset_start
-                end = self.header_size + 8 + offset_end
-                tensor_bytes = memoryview(self.mmap_obj)[start:end]
-            else:
-                tensor_bytes = None
-        else:
-            # Non-mmap mode: use pre-allocated bytearray with readinto for minimal copies
-            tensor_bytes = None
-            if offset_start != offset_end:
-                self.file.seek(self.header_size + 8 + offset_start)
-                # Pre-allocate writable buffer and read directly into it
-                tensor_bytes = bytearray(offset_end - offset_start)
-                self.file.readinto(tensor_bytes)
-
-        return self._deserialize_tensor(tensor_bytes, metadata)
-
-
-    def get_tensor_to_gpu(
-        self,
-        key: str,
-        device: str = 'cuda',
-        dtype: Optional[torch.dtype] = None
-    ) -> torch.Tensor:
-        """Load a tensor and transfer to GPU using pinned memory for faster transfer.
-
-        Args:
-            key: Tensor key to load
-            device: Target GPU device (default 'cuda')
-            dtype: Optional dtype conversion
-
-        Returns:
-            Tensor on GPU device
-        """
-        cpu_tensor = self.get_tensor(key)
-        return transfer_to_gpu_pinned(cpu_tensor, device, dtype)
-
-    def get_tensors_parallel(
-        self,
-        keys: Optional[List[str]] = None,
-        workers: Optional[int] = None,
-        sort_by_offset: bool = True
-    ) -> Dict[str, torch.Tensor]:
-        """Load multiple tensors in parallel using ThreadPoolExecutor.
-
-        Args:
-            keys: List of keys to load (default: all keys)
-            workers: Number of worker threads (default: auto-calculated based on device)
-            sort_by_offset: Sort keys by file offset before loading (default: True)
-
-        Returns:
-            Dict mapping key -> tensor
-        """
-        if keys is None:
-            keys = self.keys()
-
-        # Auto-calculate workers based on device capabilities
-        if workers is None:
-            try:
-                from .device_utils import get_optimal_workers, estimate_model_size
-                model_size = estimate_model_size(self.filename)
-                workers = get_optimal_workers(model_size)
-            except ImportError:
-                workers = min(4, os.cpu_count() or 4)
-
-        if sort_by_offset:
-            # Sort by offset for better I/O patterns
-            keys_with_offsets = [(k, self.header[k]["data_offsets"][0]) for k in keys]
-            keys_with_offsets.sort(key=lambda x: x[1])
-            keys = [k for k, _ in keys_with_offsets]
-
-        result = {}
-
-        with ThreadPoolExecutor(max_workers=workers) as executor:
-            futures = {
-                executor.submit(
-                    _read_tensor_from_file,
-                    self.filename,
-                    key,
-                    self.header[key],
-                    self.header_size
-                ): key
-                for key in keys
-            }
-
-            for future in as_completed(futures):
-                key = futures[future]
-                try:
-                    tensor = future.result()
-                    if tensor is not None:
-                        result[key] = tensor
-                except Exception as e:
-                    print(f"[MemoryEfficientSafeOpen] Failed to load {key}: {e}")
-
-        return result
-
-    def get_tensor_as_dict(self, key: str) -> Dict[str, Any]:
-        """Load a tensor and decode as JSON dict (for config tensors)."""
-        if key not in self.header:
-            raise KeyError(f"Tensor '{key}' not found")
-
-        metadata = self.header[key]
-        offset_start, offset_end = metadata["data_offsets"]
-
-        if offset_start == offset_end:
-            return {}
-
-        if self.mmap_mode and self.mmap_obj:
-            start = self.header_size + 8 + offset_start
-            end = self.header_size + 8 + offset_end
-            tensor_bytes = bytes(self.mmap_obj[start:end])
-        else:
-            self.file.seek(self.header_size + 8 + offset_start)
-            tensor_bytes = self.file.read(offset_end - offset_start)
-
-        return json.loads(tensor_bytes.decode("utf-8"))
-
-    def _read_header(self):
-        with open(self.filename, "rb") as f:
-            header_size = struct.unpack("<Q", f.read(8))[0]
-            header_json = f.read(header_size).decode("utf-8")
-        return json.loads(header_json), header_size
-
-    def _deserialize_tensor(self, tensor_bytes, metadata):
-        dtype_str = metadata["dtype"]
-        shape = metadata["shape"]
-        dtype = self._get_torch_dtype(dtype_str)
-
-        if tensor_bytes is None:
-            byte_tensor = torch.empty(0, dtype=torch.uint8)
-        else:
-            # Avoid extra copy if already a bytearray (low_memory mode)
-            if isinstance(tensor_bytes, bytearray):
-                byte_tensor = torch.frombuffer(tensor_bytes, dtype=torch.uint8)
-            else:
-                # mmap memoryview needs copy to create writable tensor
-                byte_tensor = torch.frombuffer(bytearray(tensor_bytes), dtype=torch.uint8)
-
-        if dtype_str in ["F8_E5M2", "F8_E4M3"]:
-            return self._convert_float8(byte_tensor, dtype_str, shape)
-
-        return byte_tensor.view(dtype).reshape(shape)
-
-
-    @staticmethod
-    def _get_torch_dtype(dtype_str):
-        dtype_map = {
-            "F64": torch.float64, "F32": torch.float32, "F16": torch.float16, "BF16": torch.bfloat16,
-            "I64": torch.int64, "I32": torch.int32, "I16": torch.int16, "I8": torch.int8,
-            "U8": torch.uint8, "BOOL": torch.bool,
-        }
-        if hasattr(torch, "float8_e5m2"):
-            dtype_map["F8_E5M2"] = torch.float8_e5m2
-        if hasattr(torch, "float8_e4m3fn"):
-            dtype_map["F8_E4M3"] = torch.float8_e4m3fn
-
-        dtype = dtype_map.get(dtype_str)
-        if dtype is None:
-            raise ValueError(f"Unsupported dtype: {dtype_str}")
-        return dtype
-
-    @staticmethod
-    def _convert_float8(byte_tensor, dtype_str, shape):
-        if dtype_str == "F8_E5M2" and hasattr(torch, "float8_e5m2"):
-            return byte_tensor.view(torch.float8_e5m2).reshape(shape)
-        elif dtype_str == "F8_E4M3" and hasattr(torch, "float8_e4m3fn"):
-            return byte_tensor.view(torch.float8_e4m3fn).reshape(shape)
-        else:
-            raise ValueError(f"Unsupported float8 type: {dtype_str}. Your PyTorch version may be too old.")
-
-
-def _read_tensor_from_file(
-    filepath: str,
-    key: str,
-    metadata: Dict,
-    header_size: int
-) -> Optional[torch.Tensor]:
-    """Helper function to read a single tensor from file (for parallel execution).
-
-    Each worker opens its own file handle for thread safety.
-    """
-    offset_start, offset_end = metadata["data_offsets"]
-    dtype_str = metadata["dtype"]
-    shape = metadata["shape"]
-
-    if offset_start == offset_end:
-        return None
-
-    dtype_map = {
-        "F64": torch.float64, "F32": torch.float32, "F16": torch.float16, "BF16": torch.bfloat16,
-        "I64": torch.int64, "I32": torch.int32, "I16": torch.int16, "I8": torch.int8,
-        "U8": torch.uint8, "BOOL": torch.bool,
-    }
-    if hasattr(torch, "float8_e5m2"):
-        dtype_map["F8_E5M2"] = torch.float8_e5m2
-    if hasattr(torch, "float8_e4m3fn"):
-        dtype_map["F8_E4M3"] = torch.float8_e4m3fn
-
-    dtype = dtype_map.get(dtype_str, torch.float32)
-
-    with open(filepath, "rb") as f:
-        f.seek(header_size + 8 + offset_start)
-        tensor_bytes = f.read(offset_end - offset_start)
-
-    byte_tensor = torch.frombuffer(bytearray(tensor_bytes), dtype=torch.uint8)
-
-    if dtype_str in ["F8_E5M2", "F8_E4M3"]:
-        if dtype_str == "F8_E5M2" and hasattr(torch, "float8_e5m2"):
-            return byte_tensor.view(torch.float8_e5m2).reshape(shape)
-        elif dtype_str == "F8_E4M3" and hasattr(torch, "float8_e4m3fn"):
-            return byte_tensor.view(torch.float8_e4m3fn).reshape(shape)
-
-    return byte_tensor.view(dtype).reshape(shape)
-
 
 def get_union_keys(*handlers) -> List[str]:
     """Returns the union of all tensor keys across multiple model handlers."""
@@ -386,4 +21,85 @@ 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)
+
+
+def parse_custom_weights(custom_weights_str: str) -> List[Tuple[re.Pattern, Dict[str, float]]]:
+    """Parses multiline string of custom weights.
+    Format: pattern:weight OR pattern:param=value,param2=value
+    Returns list of (compiled_regex, parameter_dict).
+    """
+    if not custom_weights_str or not custom_weights_str.strip():
+        return []
+
+    results = []
+    lines = custom_weights_str.strip().split('\n')
+    for line in lines:
+        line = line.strip()
+        if not line or line.startswith('#'):
+            continue
+
+        if ':' not in line:
+            continue
+
+        pattern_str, weight_str = line.rsplit(':', 1)
+        pattern_str = pattern_str.strip()
+        weight_str = weight_str.strip()
+
+        try:
+            pattern = re.compile(pattern_str)
+        except re.error:
+            # Fallback to literal substring if regex invalid
+            pattern = re.compile(re.escape(pattern_str))
+
+        params = {}
+        if '=' in weight_str:
+            # param=value,param2=value
+            parts = weight_str.split(',')
+            for part in parts:
+                if '=' in part:
+                    k, v = part.split('=', 1)
+                    k = k.strip()
+                    # Map common aliases
+                    if k in ['a', 'alpha']: k = 'alpha'
+                    elif k in ['b', 'beta']: k = 'beta'
+                    elif k in ['g', 'gamma']: k = 'gamma'
+                    elif k in ['d', 'delta']: k = 'delta'
+                    elif k in ['w', 'weight']: k = 'weight'
+                    elif re.match(r'^w[1-8]$', k): pass # Keep w1-w8 as is
+
+                    try:
+                        params[k] = float(v.strip())
+                    except ValueError:
+                        pass
+        else:
+            # single float or comma-separated list of floats (for multi-LoRA)
+            if ',' in weight_str:
+                parts = weight_str.split(',')
+                for i, p in enumerate(parts):
+                    try:
+                        params[f'w{i+1}'] = float(p.strip())
+                    except ValueError:
+                        pass
+            else:
+                try:
+                    val = float(weight_str)
+                    params['alpha'] = val # default for merger
+                    params['weight'] = val # default for lora_merger
+                except ValueError:
+                    pass
+
+        if params:
+            results.append((pattern, params))
+
+    return results
+
+
+def get_custom_parameters(key: str, parsed_weights: List[Tuple[re.Pattern, Dict[str, float]]], base_params: Dict[str, Any]) -> Dict[str, Any]:
+    """Returns updated parameters if key matches any pattern."""
+    new_params = base_params.copy()
+    for pattern, custom_dict in parsed_weights:
+        if pattern.search(key):
+            new_params.update(custom_dict)
+            break # First match wins
+    return new_params

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 = "0.2.7"
+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 +1,5 @@
-# No external dependencies required
+requests
+Pillow
+mutagen
+unifiedefficientloader==0.4.4
+av