[Request] Qwen-image-edit-4bit

#1
by jspsoli - opened

I tried to use your model for training with 16Gb on Ai-Toolkit - and it was successful with some minor changes in the source code and at low resolution (512) + vae and text_encoder caching.

If you can, please make a conversion for the newly released Edit model as well.

That’s great. I am able to generate at full recommended settings except that you may have to enable CPU offloading. This is because the final stage of vae and outputting the image takes a spike of GPU .

I saw the news on image edit. I will quantize and upload within a few hours but I will not be able to test it this week.

@jspsoli
Done: ovedrive/qwen-image-edit-4bit since you can code you should be fine. Please do let me know if this works out or if there are any issues.

Sign up or log in to comment