明明很好的repo，为啥star这么少 (Why does this repo have fewer stars than it deserves?)

[JasonLeeFdu]

1. huggingface版本问题
   Please pin and align compatible versions across diffusers / transformers / accelerate / peft, etc.

2. 双lora的切换模式问题
   The current “load → fuse → unfuse → unload” switching loop during inference becomes progressively slower, you can change it to pipeline.set_apater(), real fast.

[pureexe]

Thank you for your interest in our work and for the helpful feedback.

Version compatibility

The compatible versions of diffusers, transformers, accelerate, peft, etc., are explicitly pinned and defined in requirements.txt to ensure reproducibility of our experiments:
https://github.com/DiffusionLight/DiffusionLight-Turbo/blob/main/requirements.txt

Dual LoRA switching performance

We acknowledge the performance issue caused by the current load → fuse → unfuse → unload switching loop during inference.
However, we intentionally keep this implementation unchanged in this repository, as it corresponds exactly to the code used in our reported experiments. This helps ensure that results can be faithfully reproduced.

For users who prioritize inference speed, we provide an alternative implementation that uses pipeline.set_adapter(), which is significantly faster. This optimized version can be found here:
https://github.com/DiffusionLight/DiffusionLight-Turbo-diffusers

We appreciatee your suggestions and thank you again for taking the time to report these points.