'utf-8' codec can't decode byte 0xc1 in position 0: invalid start byte

[judian17]

There are a few issues when using LoRA: First, if I load a regular LoRA and select the ComfyUI (or XLab) type, the following error occurs:
model_type FLUX
!!! Exception during processing !!!
Traceback (most recent call last):
File "E:\comfyui\ComfyUI-aki-v1.3\execution.py", line 327, in execute
output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\comfyui\ComfyUI-aki-v1.3\execution.py", line 202, in get_output_data
return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\comfyui\ComfyUI-aki-v1.3\execution.py", line 174, in _map_node_over_list
process_inputs(input_dict, i)
File "E:\comfyui\ComfyUI-aki-v1.3\execution.py", line 163, in process_inputs
results.append(getattr(obj, func)(**inputs))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\comfyui\ComfyUI-aki-v1.3\custom_nodes\svdquant\nodes\lora\flux.py", line 94, in load_lora
input_lora = comfyui2diffusers(lora_path)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\comfyui\ComfyUI-aki-v1.3\python312\Lib\site-packages\nunchaku\lora\flux\comfyui_converter.py", line 53, in comfyui2diffusers
assert "lora_unet_single_blocks" in k
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError

If I switch to the Diffusers type:

[2025-02-25 01:38:22.066] [info] Loading weights from E:\comfyui\ComfyUI-aki-v1.3\models\diffusion_models\svdqfluxdev\transformer_blocks.safetensors
[2025-02-25 01:38:23.854] [info] Done.
model_type FLUX
Converting 57 transformer blocks...
Converting LoRA branch for block single_transformer_blocks.0...

• Found single_transformer_blocks.0 LoRA of qkv_proj (rank: 32)
  • Using original LoRA
• Found single_transformer_blocks.0 LoRA of out_proj (rank: 32)
  • Using original LoRA

......

• Found transformer_blocks.18 LoRA of mlp_fc2 (rank: 32)
  • Using original LoRA
• Found transformer_blocks.18 LoRA of mlp_context_fc1 (rank: 32)
  • Using original LoRA
• Found transformer_blocks.18 LoRA of mlp_context_fc2 (rank: 32)
  • Using original LoRA
    [2025-02-25 01:38:26.979] [info] Loading partial weights from C:\Users\Admin\AppData\Local\Temp\tmp5hm0i2ys.safetensors
    !!! Exception during processing !!! 'utf-8' codec can't decode byte 0xc1 in position 0: invalid start byte
    Traceback (most recent call last):
    File "E:\comfyui\ComfyUI-aki-v1.3\execution.py", line 327, in execute
    output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "E:\comfyui\ComfyUI-aki-v1.3\execution.py", line 202, in get_output_data
    return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "E:\comfyui\ComfyUI-aki-v1.3\execution.py", line 174, in _map_node_over_list
    process_inputs(input_dict, i)
    File "E:\comfyui\ComfyUI-aki-v1.3\execution.py", line 163, in process_inputs
    results.append(getattr(obj, func)(**inputs))
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "E:\comfyui\ComfyUI-aki-v1.3\custom_nodes\svdquant\nodes\lora\flux.py", line 110, in load_lora
    model.model.diffusion_model.model.update_lora_params(tmp_file.name)
    File "E:\comfyui\ComfyUI-aki-v1.3\python312\Lib\site-packages\nunchaku\models\transformer_flux.py", line 166, in update_lora_params
    block.m.load(path, True)
    UnicodeDecodeError: 'utf-8' codec can't decode byte 0xc1 in position 0: invalid start byte

[judian17]

in addition, if I use the command python -m nunchaku.lora.flux.convert --lora-format diffusers --lora-path "E:\comfyui\ComfyUI-aki-v1.3\models\loras\flux\Flux-Dev-Real-Anime-LoRA.safetensors" --lora-name testrealdiff to convert the model, the model can be generated (however, whether I use a 60MB or 600MB LoRA, the final model size is always the same at 233MB). But when I load this LoRA using the svdquant type, it generates random noise images without any pattern.

Additionally, there is a bug related to a ComfyUI node: After the model is loaded, its state seems unable to change. If I first load a problematic LoRA and generate a noise image, then disable the LoRA node, the same noise image is still generated again. Only after reloading the model can a normal image be generated (for other models, disabling the LoRA loading node usually means making it inactive without needing to reload the model). I used these loras:https://civitai.com/models/101242/crystal-liu-or-liu-yifei-cn-actress-sd15-and-pony-and-flux https://huggingface.co/prithivMLmods/Flux-Dev-Real-Anime-LoRA

[lmxyy]

in addition, if I use the command python -m nunchaku.lora.flux.convert --lora-format diffusers --lora-path "E:\comfyui\ComfyUI-aki-v1.3\models\loras\flux\Flux-Dev-Real-Anime-LoRA.safetensors" --lora-name testrealdiff to convert the model, the model can be generated (however, whether I use a 60MB or 600MB LoRA, the final model size is always the same at 233MB). But when I load this LoRA using the svdquant type, it generates random noise images without any pattern.

Additionally, there is a bug related to a ComfyUI node: After the model is loaded, its state seems unable to change. If I first load a problematic LoRA and generate a noise image, then disable the LoRA node, the same noise image is still generated again. Only after reloading the model can a normal image be generated (for other models, disabling the LoRA loading node usually means making it inactive without needing to reload the model). I used these loras:https://civitai.com/models/101242/crystal-liu-or-liu-yifei-cn-actress-sd15-and-pony-and-flux https://huggingface.co/prithivMLmods/Flux-Dev-Real-Anime-LoRA

This issue seems to be fixed in #100 Which nunchaku version are you using?

[judian17]

@lmxyy Sorry, regarding the issue of the generated image being a noise image, I successfully downloaded nunchaku yesterday, so the version I'm using should be the latest one (right?). It seems the problem was that I wasn't using the latest svdq flux dev model. After switching to the model that automatically downloads from Hugging Face within the node, I was able to successfully convert it into an SVDQuant type LoRA. However, there are still a few issues.

First, I still can't use the ComfyUI node to convert the LoRA, and the same error occurs: 'utf-8' codec can't decode byte 0xc1 in position 0: invalid start byte. I successfully converted the LoRA and loaded it with the svdquant type using the following command:

python -m nunchaku.lora.flux.convert --lora-format comfyui --lora-path "E:\comfyui\ComfyUI-aki-v1.3\models\loras\flux\AsianGirl.safetensors" --lora-name asiancomfy

However, the effect of the LoRA isn't good; the images have a severe smudging effect (I verified that the LoRA itself is fine by testing it with the regular Flux FP8).

[lmxyy]

Ok. Let me take a look at your lora.

[lmxyy]

I cannot open this url: https://civitai.com/models/101242/crystal-liu-or-liu-yifei-cn-actress-sd15-and-pony-and-flux

[lmxyy]

in addition, if I use the command python -m nunchaku.lora.flux.convert --lora-format diffusers --lora-path "E:\comfyui\ComfyUI-aki-v1.3\models\loras\flux\Flux-Dev-Real-Anime-LoRA.safetensors" --lora-name testrealdiff to convert the model, the model can be generated (however, whether I use a 60MB or 600MB LoRA, the final model size is always the same at 233MB). But when I load this LoRA using the svdquant type, it generates random noise images without any pattern.

Additionally, there is a bug related to a ComfyUI node: After the model is loaded, its state seems unable to change. If I first load a problematic LoRA and generate a noise image, then disable the LoRA node, the same noise image is still generated again. Only after reloading the model can a normal image be generated (for other models, disabling the LoRA loading node usually means making it inactive without needing to reload the model). I used these loras:https://civitai.com/models/101242/crystal-liu-or-liu-yifei-cn-actress-sd15-and-pony-and-flux https://huggingface.co/prithivMLmods/Flux-Dev-Real-Anime-LoRA

The Real-Anime LoRA is in ComfyUI format. Therefore, the conversion command should be

python -m nunchaku.lora.flux.convert --lora-format comfyui --lora-path loras/comfyui-real_anime.safetensors --lora-name svdq-int4-flux.1-dev-real_anime.safetensors --lora-format comfyui

In our current code, the LoRA weights are amplified, so you need to set LoRA strength to 0.5 for now.

I will update the codebase tomorrow morning to fix this LoRA scale issue and also support automatic LoRA format detection. If you want to use the correct converted LoRA now, you can find it here.

[lmxyy]

Since I cannot open that Yifei LoRA, I do not what caused the issue. You can upload it somewhere else.

[judian17]

https://drive.google.com/file/d/1MVaUpvCZ4K2lZSLozVFp9_62cGKaKNCD/view?usp=drive_link It seems that the model author has deleted it. Here is the Google Drive link I uploaded, and you should be able to download it via Google Drive.

[lmxyy]

I can't access the file on Google Drive. Did you set the correct permissions? You can grant me viewing access at lmxyy1999@gmail.com.

[judian17]

@lmxyy I'm sorry, but I am also not familiar with how to use Google Drive. I have already agreed to the file access request and sent the email. Also, I've noticed that the original website link for the model, which was previously inaccessible, is now available ,you can try it again.
https://civitai.com/models/101242/crystal-liu-or-liu-yifei-cn-actress-sd15-and-pony-and-flux

[judian17]

There is new information: I have installed the latest version of Nunchaku 0.13. Currently, using the auto option still results in the error "`utf-8' codec can't decode byte 0xc1 in position 0: invalid start byte". However, when I execute the following command, it successfully generates a usable LoRA:

python -m nunchaku.lora.flux.convert --lora-format auto --quant-path "E:\comfyui\ComfyUI-aki-v1.3\models\diffusion_models\svdqfluxdev2\transformer_blocks.safetensors" --lora-path "E:\comfyui\ComfyUI-aki-v1.3\models\loras\flux\LYF_Flux_young.safetensors" --lora-name lyfsvd

I am using Windows 11, and it might be a problem with reading temporary files?

[lmxyy]

It works pretty well on my side. It seems to be a windows issue. Are you able to run our LoRA example?

python examples/int4-flux.1-dev-lora.py

[lmxyy]

There is new information: I have installed the latest version of Nunchaku 0.13. Currently, using the auto option still results in the error "`utf-8' codec can't decode byte 0xc1 in position 0: invalid start byte". However, when I execute the following command, it successfully generates a usable LoRA:

python -m nunchaku.lora.flux.convert --lora-format auto --quant-path "E:\comfyui\ComfyUI-aki-v1.3\models\diffusion_models\svdqfluxdev2\transformer_blocks.safetensors" --lora-path "E:\comfyui\ComfyUI-aki-v1.3\models\loras\flux\LYF_Flux_young.safetensors" --lora-name lyfsvd

I am using Windows 11, and it might be a problem with reading temporary files?

This is possible. Maybe it is the temporary file that is not working on windows. I will work out a solution to address this.

[judian17]

I tested the Lora loading node of ComfyUI in different Python environments again and also tested this lora https://huggingface.co/aleksa-codes/flux-ghibsky-illustration , but the error persists. So, it's likely not an issue with the model or the Python version. Here is my log file, and I hope it will help resolve the problem.

Diagnostics-1740538249.log

[lmxyy]

Could you help print tmp_file.name to see what it is?