NitroFusion

NitroFusion: High-Fidelity Single-Step Diffusion through Dynamic Adversarial Training

Dar-Yen Chen, Hmrishav Bandyopadhyay, Kai Zou, Yi-Zhe Song

[arXiv Paper] [Project Page]

News

04 Dec 2024: Paper is released on arXiv, and the project page is now public.
30 Nov 2024: Our single-step text-to-image demo is publicly available on 🤗 Hugging Face Space.
29 Nov 2024: Released two checkpoints: NitroSD-Realism and NitroSD-Vibrant.

Online Demos

NitroFusion single-step Text-to-Image demo hosted on 🤗 Hugging Face Space

Model Overview

nitrosd-realism_unet.safetensors: Produces photorealistic images with fine details.
nitrosd-vibrant_unet.safetensors: Offers vibrant, saturated color characteristics.
Both models support 1 to 4 inference steps.

Usage

First, we need to implement the scheduler with timestep shift for multi-step inference:

from diffusers import LCMScheduler
class TimestepShiftLCMScheduler(LCMScheduler):
    def __init__(self, *args, shifted_timestep=250, **kwargs):
        super().__init__(*args, **kwargs)
        self.register_to_config(shifted_timestep=shifted_timestep)
    def set_timesteps(self, *args, **kwargs):
        super().set_timesteps(*args, **kwargs)
        self.origin_timesteps = self.timesteps.clone()
        self.shifted_timesteps = (self.timesteps * self.config.shifted_timestep / self.config.num_train_timesteps).long()
        self.timesteps = self.shifted_timesteps
    def step(self, model_output, timestep, sample, generator=None, return_dict=True):
        if self.step_index is None:
            self._init_step_index(timestep)
        self.timesteps = self.origin_timesteps
        output = super().step(model_output, timestep, sample, generator, return_dict)
        self.timesteps = self.shifted_timesteps
        return output

We can then utilize the diffuser pipeline:

import torch
from diffusers import DiffusionPipeline, UNet2DConditionModel
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
# Load model.
base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
repo = "ChenDY/NitroFusion"
# NitroSD-Realism
ckpt = "nitrosd-realism_unet.safetensors"
unet = UNet2DConditionModel.from_config(base_model_id, subfolder="unet").to("cuda", torch.float16)
unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device="cuda"))
scheduler = TimestepShiftLCMScheduler.from_pretrained(base_model_id, subfolder="scheduler", shifted_timestep=250)
scheduler.config.original_inference_steps = 4
# # NitroSD-Vibrant
# ckpt = "nitrosd-vibrant_unet.safetensors"
# unet = UNet2DConditionModel.from_config(base_model_id, subfolder="unet").to("cuda", torch.float16)
# unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device="cuda"))
# scheduler = TimestepShiftLCMScheduler.from_pretrained(base_model_id, subfolder="scheduler", shifted_timestep=500)
# scheduler.config.original_inference_steps = 4
pipe = DiffusionPipeline.from_pretrained(
    base_model_id,
    unet=unet,
    scheduler=scheduler,
    torch_dtype=torch.float16,
    variant="fp16",
).to("cuda")
prompt = "a photo of a cat"
image = pipe(
    prompt=prompt,
    num_inference_steps=1,  # NotroSD-Realism and -Vibrant both support 1 - 4 inference steps.
    guidance_scale=0,
).images[0]

License

NitroSD-Realism is released under cc-by-nc-4.0, following its base model DMD2.

NitroSD-Vibrant is released under openrail++.

ChenDY
/

NitroFusion