The results of flux_minimal_inference are much worse than webui forge
Opened this issue · 1 comments
When I fixed the seed, there was a big gap between the image obtained using flux_minimal_inference.py inference and the image obtained using webui forge. The quality of the image obtained by flux_minimal_inference.py was lower than that of forge. What is the reason? I am training flux. When lora is used, the image obtained by sample may also be of low quality. The machine used both times is the same
This is the script I run. In flux_minimal_inference.py I set both target_height and target_width to 512
python ./flux_minimal_inference.py \ --ckpt_path=$pretrained_model_name_or_path \ --clip_l=$clip_l \ --t5xxl=$t5xxl \ --ae=$ae \ --prompt="coffee table,pendant lamp,l-shaped sofa,lazy sofa" \ --output_dir=$output_dir \ --flux_dtype=fp8 \ --seed=1 \ --steps=20 \ --guidance=3.5 \ --cfg_scale=1.0 \ --offload
The results obtained are as follows
Below is the result I got in forge
The implementation is basically the same as the official BFL repository. Could you please try it out with the official repository?