scripts/MergeSafetensors3.py

"""
Convert Diffusers-format FLUX model to ComfyUI-compatible checkpoint.
This handles the proper folder structure and key naming.
"""

from safetensors.torch import save_file, load_file
import os
import json
from pathlib import Path

def convert_diffusers_to_comfyui(
    diffusers_folder,
    output_path,
    fp16=False
):
    """
    Convert a Diffusers FLUX model folder to a single ComfyUI checkpoint.
    
    Args:
        diffusers_folder: Path to folder containing model_index.json
        output_path: Output path for the merged .safetensors file
        fp16: If True, convert to float16 to save space
    """
    
    diffusers_folder = Path(diffusers_folder)
    
    # Verify it's a Diffusers model
    model_index = diffusers_folder / "model_index.json"
    if not model_index.exists():
        raise ValueError(f"Not a Diffusers model folder. Missing: {model_index}")
    
    with open(model_index) as f:
        config = json.load(f)
    
    print("=" * 80)
    print("DIFFUSERS TO COMFYUI CONVERTER")
    print("=" * 80)
    print(f"\nModel: {config.get('_name_or_path', 'Unknown')}")
    print(f"Format: {config.get('_class_name', 'Unknown')}")
    
    merged_state = {}
    
    # ========================================================================
    # 1. Load Transformer (main FLUX model)
    # ========================================================================
    print("\n" + "=" * 80)
    print("Loading Transformer...")
    print("=" * 80)
    
    transformer_path = diffusers_folder / "transformer"
    transformer_file = None
    
    # Find the safetensors file
    for file in transformer_path.glob("*.safetensors"):
        transformer_file = file
        break
    
    if not transformer_file:
        raise ValueError(f"No safetensors file found in {transformer_path}")
    
    print(f"Found: {transformer_file.name}")
    transformer_state = load_file(str(transformer_file))
    print(f"Loaded {len(transformer_state)} transformer parameters")
    
    # Add transformer weights (keep original keys or minimal prefix)
    for key, value in transformer_state.items():
        if fp16 and value.dtype.is_floating_point:
            value = value.half()
        merged_state[key] = value
    
    # ========================================================================
    # 2. Load VAE
    # ========================================================================
    print("\n" + "=" * 80)
    print("Loading VAE...")
    print("=" * 80)
    
    vae_path = diffusers_folder / "vae"
    vae_file = None
    
    for file in vae_path.glob("*.safetensors"):
        vae_file = file
        break
    
    if not vae_file:
        print("⚠️  No VAE file found, skipping...")
    else:
        print(f"Found: {vae_file.name}")
        vae_state = load_file(str(vae_file))
        print(f"Loaded {len(vae_state)} VAE parameters")
        
        # Add VAE weights with proper prefix
        for key, value in vae_state.items():
            if fp16 and value.dtype.is_floating_point:
                value = value.half()
            # Keep original Diffusers VAE key structure
            merged_state[key] = value
    
    # ========================================================================
    # 3. Load Text Encoders (CLIP + T5)
    # ========================================================================
    print("\n" + "=" * 80)
    print("Loading Text Encoders...")
    print("=" * 80)
    
    # CLIP (text_encoder)
    clip_path = diffusers_folder / "text_encoder"
    if clip_path.exists():
        clip_file = None
        for file in clip_path.glob("*.safetensors"):
            clip_file = file
            break
        
        if clip_file:
            print(f"Found CLIP: {clip_file.name}")
            clip_state = load_file(str(clip_file))
            print(f"Loaded {len(clip_state)} CLIP parameters")
            
            for key, value in clip_state.items():
                if fp16 and value.dtype.is_floating_point:
                    value = value.half()
                # Keep original structure
                merged_state[key] = value
        else:
            print("⚠️  No CLIP file found")
    
    # T5 (text_encoder_2) - often the largest component
    t5_path = diffusers_folder / "text_encoder_2"
    if t5_path.exists():
        t5_file = None
        for file in t5_path.glob("*.safetensors"):
            t5_file = file
            break
        
        if t5_file:
            print(f"Found T5: {t5_file.name}")
            print("⚠️  Loading T5 (this may take a while, it's large)...")
            t5_state = load_file(str(t5_file))
            print(f"Loaded {len(t5_state)} T5 parameters")
            
            for key, value in t5_state.items():
                if fp16 and value.dtype.is_floating_point:
                    value = value.half()
                merged_state[key] = value
        else:
            print("⚠️  No T5 file found")
    
    # ========================================================================
    # Save merged checkpoint
    # ========================================================================
    print("\n" + "=" * 80)
    print("Saving merged checkpoint...")
    print("=" * 80)
    
    print(f"Total parameters: {len(merged_state):,}")
    print(f"Output: {output_path}")
    
    save_file(merged_state, output_path)
    
    size_gb = os.path.getsize(output_path) / (1024**3)
    print(f"\n✅ Conversion complete!")
    print(f"File size: {size_gb:.2f} GB")
    
    # Show key structure
    print("\n" + "=" * 80)
    print("Key Structure in Merged File")
    print("=" * 80)
    
    sample_keys = list(merged_state.keys())[:10]
    print("\nFirst 10 keys:")
    for key in sample_keys:
        print(f"  {key}")
    
    return output_path


def convert_with_working_template(
    diffusers_folder,
    working_checkpoint,
    output_path,
    replace_transformer_only=True
):
    """
    Use a working checkpoint as template, replacing components from Diffusers model.
    This ensures key naming matches what ComfyUI expects.
    
    Args:
        diffusers_folder: Path to Diffusers model folder
        working_checkpoint: Path to a working ComfyUI checkpoint
        output_path: Output path for merged checkpoint
        replace_transformer_only: If True, only replace transformer, keep VAE/encoders from template
    """
    
    print("=" * 80)
    print("TEMPLATE-BASED CONVERSION")
    print("=" * 80)
    
    # Load working checkpoint as template
    print("\nLoading template checkpoint...")
    template_state = load_file(working_checkpoint)
    print(f"Template has {len(template_state)} parameters")
    
    # Get key prefixes from template
    template_keys = set(template_state.keys())
    transformer_keys = {k for k in template_keys if 'transformer' in k or 'double_blocks' in k or 'single_blocks' in k}
    vae_keys = {k for k in template_keys if 'vae' in k.lower() or 'first_stage' in k}
    text_encoder_keys = {k for k in template_keys if 'text_encoder' in k or 'clip' in k.lower()}
    
    print(f"\nTemplate structure:")
    print(f"  Transformer keys: {len(transformer_keys)}")
    print(f"  VAE keys: {len(vae_keys)}")
    print(f"  Text encoder keys: {len(text_encoder_keys)}")
    
    # Load transformer from Diffusers
    diffusers_folder = Path(diffusers_folder)
    transformer_path = diffusers_folder / "transformer"
    
    transformer_file = next(transformer_path.glob("*.safetensors"))
    print(f"\nLoading new transformer from: {transformer_file.name}")
    new_transformer = load_file(str(transformer_file))
    
    # Replace transformer weights
    print("\nReplacing transformer weights...")
    merged_state = dict(template_state)  # Copy template
    
    # Replace matching keys
    replaced = 0
    for key in transformer_keys:
        if key in new_transformer:
            merged_state[key] = new_transformer[key]
            replaced += 1
    
    print(f"Replaced {replaced} transformer parameters")
    
    if not replace_transformer_only:
        print("\n⚠️  Also replacing VAE and text encoders...")
        # Load and replace VAE
        vae_file = next((diffusers_folder / "vae").glob("*.safetensors"), None)
        if vae_file:
            vae_state = load_file(str(vae_file))
            for key in vae_keys:
                if key in vae_state:
                    merged_state[key] = vae_state[key]
        
        # Similar for text encoders...
    
    # Save
    print(f"\nSaving to {output_path}...")
    save_file(merged_state, output_path)
    
    size_gb = os.path.getsize(output_path) / (1024**3)
    print(f"✅ Done! File size: {size_gb:.2f} GB")


# Example usage
if __name__ == "__main__":
    # Method 1: Direct conversion
    # convert_diffusers_to_comfyui(
    #     diffusers_folder="../",
    #     output_path="flux1-depth-dev_ComfyMerged.safetensors",
    #     fp16=True  # Set False to keep original precision
    # )
    
    #Method 2: Use working checkpoint as template (RECOMMENDED)
    convert_with_working_template(
        diffusers_folder="../",
        working_checkpoint="../quantized/svdq-fp4_r32-flux.1-depth-dev.safetensors",
        output_path="svdq-fp4_r32-flux.1-depth-dev_ComfyMerged.safetensors",
        replace_transformer_only=True
    )