FP8 compatibility / quantized model weights for 24 GB GPUs?

[jeolpyeoni]

Hi, thanks for great work! I've been experimenting with it and ran into a VRAM constraint.

The transformer is ~38 GB in bf16, which exceeds the 24 GB VRAM of common GPUs like the RTX A5000/3090/4090. A community fp8 quantization exists (https://huggingface.co/1038lab/Qwen-Image-Edit-2511-FP8) which brings it down to ~20 GB, potentially fitting within 24 GB.

A few questions:

1. Have you tested PixelSmile's LoRA weights with the fp8-quantized base model? Any noticeable quality degradation?
2. Are you planning to release an officially validated quantized version (fp8/int8) of the base model or LoRA?
3. Is there a recommended workaround for sub-24 GB inference in the meantime?

Thanks!

[Ammmob]

@jeolpyeoni Thanks for the questions!

We haven’t officially tested PixelSmile with fp8-quantized base models yet, so we can’t comment on the quality impact.

Our current release is validated under the bf16 setup. Reducing VRAM usage (e.g., via fp8/int8) is definitely an interesting direction, and something we may explore in the future.

For now, using community quantized models like the one you mentioned could be a practical workaround, though it hasn’t been fully validated on our side.