Jiaxiang Cheng, Pan Xie*, Xin Xia, Jiashi Li, Jie Wu, Yuxi Ren, Huixia Li, Xuefeng Xiao, Min Zheng, Lean Fu (*Corresponding author)
AutoML, ByteDance Inc.
⭐ If ResAdapter is helpful to your images or projects, please help star this repo. Thanks! 🤗
We propose ResAdapter, a plug-and-play resolution adapter for enabling any diffusion model generate resolution-free images: no additional training, no additional inference and no style transfer.
Comparison examples between resadapter and dreamlike-diffusion-1.0.
[2024/03/30]
🔥 We release the ComfyUI-ResAdapter node.[2024/03/28]
🔥 We release the all resadapter weights.[2024/03/12]
🔥 We release the inference code and gradio demo.[2024/03/04]
🔥 We release the arxiv paper.
We provide a standalone example code to help you quickly use resadapter with diffusion models.
Comparison examples (640x384) between resadapter and dreamshaper-xl-1.0. Top: with resadapter. Bottom: without resadapter.
# pip install diffusers, transformers, accelerate, safetensors, huggingface_hub
import torch
from torchvision.utils import save_image
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler
generator = torch.manual_seed(0)
prompt = "portrait photo of muscular bearded guy in a worn mech suit, light bokeh, intricate, steel metal, elegant, sharp focus, soft lighting, vibrant colors"
width, height = 640, 384
# Load baseline pipe
model_name = "lykon-models/dreamshaper-xl-1-0"
pipe = AutoPipelineForText2Image.from_pretrained(model_name, torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True, algorithm_type="sde-dpmsolver++")
# Inference baseline pipe
image = pipe(prompt, width=width, height=height, num_inference_steps=25, num_images_per_prompt=4, output_type="pt").images
save_image(image, f"image_baseline.png", normalize=True, padding=0)
# Load resadapter for baseline
resadapter_model_name = "resadapter_v1_sdxl"
pipe.load_lora_weights(
hf_hub_download(repo_id="jiaxiangc/res-adapter", subfolder=resadapter_model_name, filename="pytorch_lora_weights.safetensors"),
adapter_name="res_adapter",
) # load lora weights
pipe.set_adapters(["res_adapter"], adapter_weights=[1.0])
pipe.unet.load_state_dict(
load_file(hf_hub_download(repo_id="jiaxiangc/res-adapter", subfolder=resadapter_model_name, filename="diffusion_pytorch_model.safetensors")),
strict=False,
) # load norm weights
# Inference resadapter pipe
image = pipe(prompt, width=width, height=height, num_inference_steps=25, num_images_per_prompt=4, output_type="pt").images
save_image(image, f"image_resadapter.png", normalize=True, padding=0)
We have released all resadapter weights, you can download resadapter models from Huggingface. The following is our resadapter model card:
Models | Parameters | Resolution Range | Ratio Range | Links |
---|---|---|---|---|
resadapter_v1_sd1.5 | 0.9M | 128 <= x <= 1024 | 0.5 <= r <= 2 | Download |
resadapter_v1_sd1.5_extrapolation | 0.9M | 512 <= x <= 1024 | 0.5 <= r <= 2 | Download |
resadapter_v1_sd1.5_interpolation | 0.8M | 128 <= x <= 512 | 0.5 <= r <= 2 | Download |
resadapter_v1_sdxl | 0.5M | 256 <= x <= 1536 | 0.5 <= r <= 2 | Download |
resadapter_v1_sdxl_extrapolation | 0.5M | 1024 <= x <= 1536 | 0.5 <= r <= 2 | Download |
resadapter_v1_sdxl_interpolation | 0.4M | 256 <= x <= 1024 | 0.5 <= r <= 2 | Download |
Hint1: We update the resadapter name format according to controlnet.
Hint2: If you want use resadapter with personalized diffusion models, you should download them from CivitAI.
Hint3: If you want use resadapter with ip-adapter, controlnet and lcm-lora, you should download them from Huggingface.
Hint4: Here is an installation guidance for preparing environment and downloading models.
If you want generate images in our inference script, you should download related models and fill in configs. Then you can directly run this script.
python main.py --config /path/to/file
Comparison examples (960x1104) between resadapter and dreamshaper-7. Top: with resadapter. Bottom: without resadapter.
Comparison examples (840x1264) between resadapter and lllyasviel/sd-controlnet-canny. Top: with resadapter, bottom: without resadapter.
Comparison examples (336x504) between resadapter and diffusers/controlnet-canny-sdxl-1.0. Top: with resadapter, bottom: without resadapter.
Comparison examples (864x1024) between resadapter and h94/IP-Adapter. Top: with resadapter, bottom: without resadapter.
Comparison examples (512x512) between resadapter and dreamshaper-xl-1.0 with lcm-sdxl-lora. Top: with resadapter, bottom: without resadapter.
- Replicate website: bytedance/res-adapter by (@ Chenxi)
- Huggingface space: ameerazam08/Res-Adapter-GPU-Demo by (@Ameer Azam)
- jiaxiangc/ComfyUI-ResAdapter (official comfyui node)
- blepping/ComfyUI-ApplyResAdapterUnet by (@ blepping)
helper_video_720p.mp4
Run the following script:
python app.py
- We recommend users to use interpolation version to generate lower-resolution images.
- We recommend users to use extrapolation version to generate higher-resolution images.
- We recommend users to use
resadapter_v1_sd1.5
andresadapter_v1_sdxl
for deploying resadapter to generate images with broader resolution. - We strongly recommend that you use the prompt corresponding to the personalized model, which helps to enhance the quality of the image.
- ResAdapter is developed by AutoML Team at ByteDance Inc, all copyright reserved.
- Thanks to the HuggingFace gradio team for their free GPU support!
- Thanks to the IP-Adapter, ControlNet, LCM-LoRA for their nice work.
If you find ResAdapter useful for your research and applications, please cite us using this BibTeX:
@article{cheng2024resadapter,
title={ResAdapter: Domain Consistent Resolution Adapter for Diffusion Models},
author={Cheng, Jiaxiang and Xie, Pan and Xia, Xin and Li, Jiashi and Wu, Jie and Ren, Yuxi and Li, Huixia and Xiao, Xuefeng and Zheng, Min and Fu, Lean},
booktitle={arXiv preprint arxiv:2403.02084},
year={2024}
}
For any question, please feel free to contact us via [email protected] or [email protected].