comfyui_nunchaku_lora_patch

Official Fix (2025-07-28)

The official fix is being tested: https://github.com/nunchaku-tech/nunchaku/pull/557

--

✅ Patched LoRAs

⚡ Usage

• These patched LoRAs are compatible with ComfyUI-nunchaku.
• Use the Nunchaku FLUX LoRA Loader node to load LoRA modules for SVDQuant FLUX models.

🛠️ Patch References

Some original FLUX LoRA files were missing required final_layer.adaLN weights needed by ComfyUI-nunchaku’s FLUX LoRA Loader.
This patch script automatically adds dummy adaLN tensors to make the LoRA compatible with SVDQuant FLUX models.

Script: patch_comfyui_nunchaku_lora.py

Based on:

• Nunchaku Issue: ComfyUI-nunchaku #340

  Node Type: NunchakuFluxLoraLoader
  Exception Type: KeyError
  Exception Message: 'lora_unet_final_layer_adaLN_modulation_1.lora_down.weight'

• Example Gist: akedia/e0a132b5...

---

🔄 Example Patch Log

Running patch_comfyui_nunchaku_lora.py

🔄 Universal final_layer.adaLN LoRA patcher (.safetensors)

Enter path to input LoRA .safetensors file: Flux_kontext_deblur.safetensors
Enter path to save patched LoRA .safetensors file: flux_kontext_deblur_nunchaku.safetensors

✅ Loaded 610 tensors from: Flux_kontext_deblur.safetensors

🔑 Found final_layer-related keys:
  - lora_unet_final_layer_linear.lora_down.weight
  - lora_unet_final_layer_linear.lora_up.weight

🔍 Checking for final_layer keys with prefix 'lora_unet_final_layer'
  Linear down: lora_unet_final_layer_linear.lora_down.weight
  Linear up:   lora_unet_final_layer_linear.lora_up.weight
  ✅ Has final_layer.linear: True
  ✅ Has final_layer.adaLN_modulation_1: False

✅ Added dummy adaLN weights:
  - lora_unet_final_layer_adaLN_modulation_1.lora_down.weight (shape: torch.Size([16, 3072]))
  - lora_unet_final_layer_adaLN_modulation_1.lora_up.weight   (shape: torch.Size([64, 16]))

✅ Patched file saved to: flux_kontext_deblur_nunchaku.safetensors
   Total tensors now: 612

🔍 Verifying patched keys:
  - lora_unet_final_layer_adaLN_modulation_1.lora_down.weight
  - lora_unet_final_layer_adaLN_modulation_1.lora_up.weight
  - lora_unet_final_layer_linear.lora_down.weight
  - lora_unet_final_layer_linear.lora_up.weight

✅ Contains adaLN after patch: True