QwenImageTransformer2DModel.from_single_file

[dxqb]

Describe the bug

QwenImageTransformer2DModel.from_single_file requires the following arguments:

QwenImageTransformer2DModel.from_single_file(
  local_safetensors_file,
  config="Qwen/Qwen-Image",
  subfolder="transformer",
  torch_dtype = torch.bfloat16
)

if config and subfolder are not provided, it tries to load a SD15 configuration - see log below.
I cannot find specific documentation so this might also be expected behaviour - but it is different from the Flux and Chroma transformers. There a filename is enough.

Reproduction

see above

Logs

transformer = QwenImageTransformer2DModel.from_single_file(
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/workspace/OneTrainer/venv/lib/python3.12/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/workspace/OneTrainer/venv/src/diffusers/src/diffusers/loaders/single_file_model.py", line 383, in from_single_file
    diffusers_model_config = cls.load_config(
                             ^^^^^^^^^^^^^^^^
  File "/workspace/OneTrainer/venv/lib/python3.12/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/workspace/OneTrainer/venv/src/diffusers/src/diffusers/configuration_utils.py", line 419, in load_config
    raise EnvironmentError(
OSError: stable-diffusion-v1-5/stable-diffusion-v1-5 does not appear to have a file named config.json.

System Info

https://github.com/huggingface/diffusers.git@9b721db

Who can help?

No response

[DN6]

Hi @dxqb We usually do auto config fetching by running a heuristic against check against the original format checkpoint keys, but Qwen Image was originally released in a diffusers compatible format. Attempting the keep track of two sets of key types (diffusers and original) isn't really practical here. Auto config fetching will also be refactored to something more maintainable in the coming months)

Flux and SD checkpoints were not originally diffusers compatible, which is why we store their keys and try to infer the model config automatically.

[dxqb]

Hi @dxqb We usually do auto config fetching by running a heuristic against check against the original format checkpoint keys, but Qwen Image was originally released in a diffusers compatible format. Attempting the keep track of two sets of key types (diffusers and original) isn't really practical here. Auto config fetching will also be refactored to something more maintainable in the coming months)

Flux and SD checkpoints were not originally diffusers compatible, which is why we store their keys and try to infer the model config automatically.

Thanks!
Closing then