patch_comfyui_nunchaku_lora

Deprecated — This issue has been fixed in the latest Nunchaku release. See nunchaku-tech/ComfyUI-nunchaku#406.

A single-file Python utility that patches .safetensors LoRA models missing adaLN (Adaptive Layer Normalization) weights in their final layer — restoring compatibility with the Nunchaku LoRA loader in ComfyUI.

Based on: lym00/comfyui_nunchaku_lora_patch

The Problem

Some LoRA models are trained without final_layer.adaLN_modulation_1 weights. The Nunchaku loader strictly requires these keys and fails with a key mismatch error when they are absent.

How It Works

1. Scans lora/ directory for all .safetensors files (skips _patched files)
2. For each file: detects final_layer.linear weights and checks for missing adaLN_modulation_1 counterparts
3. Generates zero-filled dummy tensors matching the shape of the linear weights
4. Saves a new _patched.safetensors file — original is never modified

The patched model produces identical visual output to the original. The dummy weights are zero-filled placeholders purely for loader compatibility — they have no functional effect.

Tech Stack

Dependency	Purpose
Python 3.8+	Runtime
torch	Tensor creation and shape introspection
safetensors	Read / write .safetensors model files

Usage

pip install torch safetensors

1. Place LoRA models to patch in the lora/ directory
2. Run the script:

   python patch_comfyui_nunchaku_lora.py

3. Patched files are saved as lora/<original_name>_patched.safetensors

Example output:

🔄 Universal final_layer.adaLN LoRA patcher (.safetensors)
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).

License

MIT