Flash Sd3

⚡ 🚀 Flash Diffusion: FlashSD3 ⚡

Flash Diffusion is a diffusion distillation method that accelerates image generation with diffusion models. It's capable of generating high - resolution 1024x1024 images in just 4 steps, based on the SD3 model. Check out our live demo and official Github repo to see it in action.

🚀 Quick Start

The model can be used using the StableDiffusion3Pipeline from the diffusers library directly, which allows reducing the number of required sampling steps to 4 steps.

⚠️ Important Note First, you need to install a specific version of diffusers by running:

pip install git+https://github.com/initml/diffusers.git@clement/feature/flash_sd3

Then, you can run the following to generate an image:

import torch
from diffusers import StableDiffusion3Pipeline, SD3Transformer2DModel, FlashFlowMatchEulerDiscreteScheduler
from peft import PeftModel

# Load LoRA
transformer = SD3Transformer2DModel.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    subfolder="transformer",
    torch_dtype=torch.float16,
)
transformer = PeftModel.from_pretrained(transformer, "jasperai/flash-sd3")

# Pipeline
pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    transformer=transformer,
    torch_dtype=torch.float16,
    text_encoder_3=None,
    tokenizer_3=None
)

# Scheduler
pipe.scheduler = FlashFlowMatchEulerDiscreteScheduler.from_pretrained(
  "stabilityai/stable-diffusion-3-medium-diffusers",
  subfolder="scheduler",
)

pipe.to("cuda")

prompt = "A raccoon trapped inside a glass jar full of colorful candies, the background is steamy with vivid colors."

image = pipe(prompt, num_inference_steps=4, guidance_scale=0).images[0]

✨ Features

Flash Diffusion is a diffusion distillation method proposed in Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation by Clément Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin from Jasper Research. This model is a 90.4M LoRA distilled version of SD3 model that is able to generate 1024x1024 images in 4 steps.

🔧 Technical Details

The model was trained for ~50 hours on 2 H100 GPUs.

💡 Usage Tip The model could perform much better on text if distilled on a dataset of images containing text. Feel free to try it yourself.

📄 License

This model is released under the Creative Commons BY - NC license.

📚 Documentation

Citation

If you find this work useful or use it in your research, please consider citing us:

@misc{chadebec2024flash,
      title={Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation}, 
      author={Clement Chadebec and Onur Tasar and Eyal Benaroche and Benjamin Aubin},
      year={2024},
      eprint={2406.02347},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}