InstantX/flux-dev-de-distill-diffusers

Model Details

This is a diffusers version of nyanko7/flux-dev-de-distill that generates exactly the same results as the original. But please note that there is a certain degree of quality degradation in comparison with black-forest-labs/FLUX.1-dev in our test. We are training its fine-tuning models to check its effectiveness, feel free to let us know if you have any interesting findings.

Usage

import torch
from diffusers import FluxTransformer2DModel
from pipeline_flux_de_distill import FluxPipeline

model_path = "black-forest-labs/FLUX.1-dev"
    
transformer = FluxTransformer2DModel.from_pretrained(
    "InstantX/flux-dev-de-distill-diffusers",
    torch_dtype=torch.bfloat16
)

pipeline = FluxPipeline.from_pretrained(model_path, transformer=transformer, torch_dtype=torch.bfloat16).to("cuda")

prompt = "a tiny astronaut hatching from an egg on the moon"
negative_prompt = "bad photo"

image = pipeline(
    prompt=prompt,
    negative_prompt=negative_prompt,
    guidance_scale=3.5,
    num_inference_steps=24,
).images[0]

image.save("de-distill.jpg")

Other Resources

• MinusZoneAI/flux-dev-de-distill-fp8, TheYuriLover/flux-dev-de-distill-GGUF
• ashen0209/Flux-Dev2Pro, ostris/OpenFLUX.1

Acknowledgements

This project is co-sponsored by HuggingFace and fal.