jasperai/flash-sdxl

⚡ Flash Diffusion: FlashSDXL ⚡

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 108M LoRA distilled version of SDXL model that is able to generate images in 4 steps. The main purpose of this model is to reproduce the main results of the paper. See our live demo and official Github repo.

How to use?

The model can be used using the DiffusionPipeline from diffusers library directly. It can allow reducing the number of required sampling steps to 4 steps.

from diffusers import DiffusionPipeline, LCMScheduler

adapter_id = "jasperai/flash-sdxl"

pipe = DiffusionPipeline.from_pretrained(
  "stabilityai/stable-diffusion-xl-base-1.0",
  use_safetensors=True,
)

pipe.scheduler = LCMScheduler.from_pretrained(
  "stabilityai/stable-diffusion-xl-base-1.0",
  subfolder="scheduler",
  timestep_spacing="trailing",
)
pipe.to("cuda")

# Fuse and load LoRA weights
pipe.load_lora_weights(adapter_id)
pipe.fuse_lora()

prompt = "A raccoon reading a book in a lush forest."

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

Combining Flash Diffusion with Existing LoRAs 🎨

FlashSDXL can also be combined with existing LoRAs to unlock few steps generation in a training free manner. It can be integrated straight to Hugging Face pipelines. See an example below.

from diffusers import DiffusionPipeline, LCMScheduler
import torch

user_lora_id = "TheLastBen/Papercut_SDXL"
trigger_word = "papercut"

flash_lora_id = "jasperai/flash-sdxl"

# Load Pipeline
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    variant="fp16"
)

# Set scheduler
pipe.scheduler = LCMScheduler.from_config(
    pipe.scheduler.config
)

# Load LoRAs
pipe.load_lora_weights(flash_lora_id, adapter_name="flash")
pipe.load_lora_weights(user_lora_id, adapter_name="lora")

pipe.set_adapters(["flash", "lora"], adapter_weights=[1.0, 1.0])
pipe.to(device="cuda", dtype=torch.float16)

prompt = f"{trigger_word} a cute corgi"

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

Combining Flash Diffusion with Existing ControlNets 🎨

FlashSDXL can also be combined with existing ControlNets to unlock few steps generation in a training free manner. It can be integrated straight to Hugging Face pipelines. See an example below.

import torch
import cv2
import numpy as np
from PIL import Image

from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel, LCMScheduler
from diffusers.utils import load_image, make_image_grid

flash_lora_id = "jasperai/flash-sdxl"

image = load_image(
    "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png"
).resize((1024, 1024))

image = np.array(image)

image = cv2.Canny(image, 100, 200)
image = image[:, :, None].repeat(3, 2)
canny_image = Image.fromarray(image)

# Load ControlNet
controlnet = ControlNetModel.from_pretrained(
    "diffusers/controlnet-canny-sdxl-1.0",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    controlnet=controlnet,
    torch_dtype=torch.float16,
    safety_checker=None,
    variant="fp16"
).to("cuda")

# Set scheduler
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)

# Load LoRA
pipe.load_lora_weights(flash_lora_id)
pipe.fuse_lora()

image = pipe(
    "picture of the mona lisa",
    image=canny_image,
    num_inference_steps=4,
    guidance_scale=0,
    controlnet_conditioning_scale=0.5,
    cross_attention_kwargs={"scale": 1},
).images[0]
make_image_grid([canny_image, image], rows=1, cols=2)

Training Details

The model was trained for 20k iterations on 4 H100 GPUs (representing approximately a total of 176 GPU hours of training). Please refer to the paper for further parameters details.

Metrics on COCO 2014 validation (Table 3)

• FID-10k: 21.62 (4 NFE)
• CLIP Score: 0.327 (4 NFE)

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}
}

License

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