deprecated - as it has been fixed in latest nunchaku version (see nunchaku-tech/ComfyUI-nunchaku#406 (comment) )

---

based on: https://huggingface.co/lym00/comfyui_nunchaku_lora_patch/blob/main/patch_comfyui_nunchaku_lora.py

Universal LoRA Final Layer adaLN Patcher

This script patches .safetensors LoRA models that are missing adaLN (Adaptive Layer Normalization) weights in their final layer. This is a common issue with some LoRA models that can cause compatibility problems with certain loaders, such as the Nunchaku LoRA loader in ComfyUI, which expect these weights to be present.

The Problem

When a LoRA model is trained without adaLN weights for the final layer, loaders that strictly require them will fail to load the model, often with a key mismatch error. This script provides a simple solution by adding dummy (zero-filled) adaLN weights, allowing the model to be loaded without errors.

How It Works

The script automatically scans the lora/ directory for any .safetensors files. For each file found, it inspects it for final_layer.linear weights. If these weights exist but the corresponding final_layer.adaLN_modulation_1 weights are missing, the script performs the following actions:

1. Reads the existing final_layer.linear tensors to determine their shape.
2. Generates new zero-filled "dummy" tensors for adaLN_modulation_1 that match the shape of the linear layer tensors.
3. Adds these new dummy tensors to the model's state dictionary under the correct keys.
4. Saves the modified state dictionary to a new .safetensors file in the same directory, with a _patched suffix (e.g., my_lora_patched.safetensors).

The script is designed to be robust and checks for common key prefixes (e.g., lora_unet_final_layer, final_layer) to maximize compatibility. It will skip any files that already have the _patched suffix.

Usage

Prerequisites

Make sure you have Python installed, along with the torch and safetensors libraries. If you don't have them, you can install them using pip:

pip install torch safetensors

Steps

1. Place all the LoRA models you want to patch into the lora/ directory.
2. Run the script from your terminal in the project directory:

   python patch_comfyui_nunchaku_lora.py

3. The script will automatically find, patch, and save the new files in the lora/ directory.

Example Output

$ python patch_comfyui_nunchaku_lora.py

🔄 Universal final_layer.adaLN LoRA patcher (.safetensors)
Looking for .safetensors files in the 'lora' directory...

Found 1 file(s) to process.

-----------------------------------------------------
Processing: lora\fal-Realism-Detailer-Kontext.safetensors
✅ Loaded 130 tensors.
✅ Patch applied using prefix 'lora_unet_final_layer'.
✅ Patched file saved to: lora\fal-Realism-Detailer-Kontext_patched.safetensors
✅ Verification successful: `adaLN` keys are present.
-----------------------------------------------------
🎉 Done. Patched 1 file(s).

Note: This script does not create functional adaLN layers; it only adds zero-filled placeholder weights to ensure compatibility. The patched model should produce the same visual output as the original.