News

Use the ew versio for commuity: SDXS-512-DreamShaper. It has better quality ad is faster.

SDXS-512-0.9

SDXS is a model that ca geerate high-resolutio images i real-time based o prompt texts, traied usig score distillatio ad feature matchig. For more iformatio, please refer to our research paper: SDXS: Real-Time Oe-Step Latet Diffusio Models with Image Coditios. We ope-source the model as part of the research.

SDXS-512-0.9 is a old versio of SDXS-512. I order to avoid some possible commercial ad copyright risks, the SDXS-512-1.0 ad SDXS-1024-1.0 will ot be available shortly, ad as a alterative we will provide ew versios with differet teacher DM or offlie DM. Watch our repo for ay updates.

Model Iformatio:

• Teacher DM: SD Turbo
• Offlie DM: SD v2.1 base
• VAE: TAESD

The mai differeces betwee this model ad versio 1.0 are i three aspects:

1. This versio employs TAESD, which may produce low-quality images whe weight_type is float16. Our image decoder is ot compatible with the curret versio of diffusers, so it will ot be provided ow.
2. This versio did ot perform the LoRA-GAN fietue metioed i the implemetatio details sectio, which may result i slightly iferior image details.
3. This versio replaces self-attetio with cross-attetio i the highest resolutio stages, which itroduces miimal overhead compared to directly removig them.

Diffusers Usage

import torch
from diffusers import StableDiffusioPipelie, AutoecoderKL

repo = "IDKiro/sdxs-512-0.9"
seed = 42
weight_type = torch.float32     # or float16

# Load model.
pipe = StableDiffusioPipelie.from_pretraied(repo, torch_dtype=weight_type)

# use origial VAE
# pipe.vae = AutoecoderKL.from_pretraied("IDKiro/sdxs-512-0.9/vae_large")

pipe.to("cuda")

prompt = "portrait photo of a girl, photograph, highly detailed face, depth of field, moody light, golde hour"

# Esure usig 1 iferece step ad CFG set to 0.
image = pipe(
    prompt, 
    um_iferece_steps=1, 
    guidace_scale=0,
    geerator=torch.Geerator(device="cuda").maual_seed(seed)
).images[0]

image.save("output.pg")

Cite Our Work

@article{sog2024sdxs,
  author    = {Yuda Sog, Zehao Su, Xuawu Yi},
  title     = {SDXS: Real-Time Oe-Step Latet Diffusio Models with Image Coditios},
  joural   = {arxiv},
  year      = {2024},
}

News Use the new version for community: SDXS-512-DreamShaper. It has better quality and is faster. S