Qwen Image Fine Tuning Metadata and Multiple GPU Fine Tuning Fix

[FurkanGozukara]

This fixes inaccurate metadata of saved Qwen Fine Tuned models

This is massively important @kohya-ss since editing metadata takes massive time on RunPod when using with SwarmUI

I hope you can merge asap

I tested and both resolution and arch fix works auto recognized

[kohya-ss]

I don't think it's a good idea to add arguments to build_metadata just for Qwen-Image-Edit-2509... I'll think about what to do.

[kohya-ss]

In my understanding, isn't the resolution of Qwen-Image-Edit (1024,1024)?

[FurkanGozukara]

@kohya-ss both are 1328x1328

i also even trained it and generating 1328x1328 here example : #669

also please add metadata it really makes difference when using easiness

[kohya-ss]

According to the official implementation in Diffusers, it is 1024x1024.

https://github.com/huggingface/diffusers/blob/9c3b58dcf16ebd027fd3d85ec703ad5a142b1e1e/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit_plus.py#L566-L569

[FurkanGozukara]

@kohya-ss i see. i just updated my pull request to include this new strategy tested it works

e.g.

--meta_arch "qwen-image-edit-plus" --meta_res "512x512"

you can modify delete edit as you wish merge my pull request i just want it to be better :)

and this is super important because i will train qwen image edit model like qwen image base model so i need to provide accurate arch when saving

i will also train 1024x1024 and compare with 1328x1328

[FurkanGozukara]

@kohya-ss i have another proposal

can you add 2 args that we can provide arbitrary that will be used when writing metadata?

like

--meta_res "1328x1328"
--meta_arch "qwen-image-edit-plus"

i prefer this solution

i can make pull request for this if you say ok

[FurkanGozukara]

@kohya-ss I just fixed 2 errors of Multiple GPU training with Qwen Image full fine tuning

LoRA working fine

You can read error details here #672

[kohya-ss]

@kohya-ss I just fixed 2 errors of Multiple GPU training with Qwen Image full fine tuning

If you have multiple features to add, please make multiple pull requests.

Edit:

Otherwise testing will be difficult.
Please make this pull request the one for the metadata, and make a separate pull request for the multi-gpu fixes.

Once this pull request becomes for one feature, I will merge and make some minor fixes, and merge it into the mainline.

[FurkanGozukara]

seperated into 2 different pull requests

#674

#673

[FurkanGozukara]

@kohya-ss I just fixed 2 errors of Multiple GPU training with Qwen Image full fine tuning

If you have multiple features to add, please make multiple pull requests.

Edit:

Otherwise testing will be difficult. Please make this pull request the one for the metadata, and make a separate pull request for the multi-gpu fixes.

Once this pull request becomes for one feature, I will merge and make some minor fixes, and merge it into the mainline.

sure thank you so much

seperated into 2 pull requests