search

XLabs-AI / x-flux

Apache License 2.0
1.65k stars 118 forks source link

Adapters not working in huggingface #80

Open Xuzzo opened 3 months ago

Xuzzo commented 3 months ago

Some of your LoRA adapters are not loaded in Huggingface. The easiest way to try this is by generating images e.g. here

Prompt: "A cute corgi lives in a house made out of sushi, anime" LoRA: flux-anime CFG Scale 3.5 Steps 28 height 512 width 256 seed: 0

The image with LoRA Scale 0.9

image

The image with LoRA scale 0 image (1)

In other words, the LoRA weights have no effect

I have also tried directly with the code

base_model = "black-forest-labs/FLUX.1-dev"
pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16)

pipe.load_lora_weights("XLabs-AI/flux-lora-collection", weight_name="anime_lora.safetensors")

image = pipe(
            prompt="A cute corgi lives in a house made out of sushi, anime",
            num_inference_steps=28,
            guidance_scale=3.5,
            width=512,
            height=256,
            generator=torch.Generator(device='cpu').manual_seed(0),
            joint_attention_kwargs={"scale": 0.9},
).images[0]

and including the pipe.load_lora_weights or not makes no difference