Shaojin Wu, Mengqi Huang*, Wenxu Wu, Yufeng Cheng, Fei Ding+, Qian He
Intelligent Creation Team, ByteDance
[08/18/2024] β¨ We open-sourced the UNO-1M dataset, which is a large and high-quality dataset (~1M paired images). We hope it can further benefit research.
[26/06/2025] π Congratulations! UNO has been accepted by ICCV 2025!
[04/16/2025] π₯ Our companion project RealCustom is released.
[04/10/2025] π₯ Update fp8 mode as a primary low vmemory usage support. Gift for consumer-grade GPU users. The peak Vmemory usage is ~16GB now. We may try further inference optimization later.
[04/03/2025] π₯ The demo of UNO is released.
[04/03/2025] π₯ The training code, inference code, and model of UNO are released.
[04/02/2025] π₯ The project page of UNO is created.
[04/02/2025] π₯ The arXiv paper of UNO is released.
In this study, we propose a highly-consistent data synthesis pipeline to tackle this challenge. This pipeline harnesses the intrinsic in-context generation capabilities of diffusion transformers and generates high-consistency multi-subject paired data. Additionally, we introduce UNO, which consists of progressive cross-modal alignment and universal rotary position embedding. It is a multi-image conditioned subject-to-image model iteratively trained from a text-to-image model. Extensive experiments show that our method can achieve high consistency while ensuring controllability in both single-subject and multi-subject driven generation.
π§ Requirements and InstallationInstall the requirements
# pip install -r requirements.txt # legacy installation command ## create a virtual environment with python >= 3.10 <= 3.12, like # python -m venv uno_env # source uno_env/bin/activate # or # conda create -n uno_env python=3.10 -y # conda activate uno_env # then install the requirements by you need # !!! if you are using amd GPU/NV RTX50 series/macos MPS, you should install the correct torch version by yourself first # !!! then run the install command pip install -e . # for who wanna to run the demo/inference only pip install -e .[train] # for who also want to train the model
then download checkpoints in one of the three ways:
hf_hub_download function in the code to your $HF_HOME(the default value is ~/.cache/huggingface).huggingface-cli download <repo name> to download black-forest-labs/FLUX.1-dev, xlabs-ai/xflux_text_encoders, openai/clip-vit-large-patch14, bytedance-research/UNO, then run the inference scripts. You can just download the checkpoint in need only to speed up your set up and save your disk space. i.e. for black-forest-labs/FLUX.1-dev use huggingface-cli download black-forest-labs/FLUX.1-dev flux1-dev.safetensors and huggingface-cli download black-forest-labs/FLUX.1-dev ae.safetensors, ignoreing the text encoder in black-forest-labes/FLUX.1-dev model repo(They are here for diffusers call). All of the checkpoints will take 37 GB of disk space.huggingface-cli download <repo name> --local-dir <LOCAL_DIR> to download all the checkpoints mentioned in 2. to the directories your want. Then set the environment variable AE, FLUX_DEV(or FLUX_DEV_FP8 if you use fp8 mode), T5, CLIP, LORA to the corresponding paths. Finally, run the inference scripts.AE, FLUX_DEV, T5, CLIP, LORA to the corresponding paths. Finally, run the inference scripts.For low vmemory usage, please pass the --offload and --name flux-dev-fp8 args. The peak memory usage will be 16GB. Just for reference, the end2end inference time is 40s to 1min on RTX 3090 in fp8 and offload mode.
python app.py --offload --name flux-dev-fp8
Start from the examples below to explore and spark your creativity. β¨
python inference.py --prompt "A clock on the beach is under a red sun umbrella" --image_paths "assets/clock.png" --width 704 --height 704 python inference.py --prompt "The figurine is in the crystal ball" --image_paths "assets/figurine.png" "assets/crystal_ball.png" --width 704 --height 704 python inference.py --prompt "The logo is printed on the cup" --image_paths "assets/cat_cafe.png" "assets/cup.png" --width 704 --height 704
Optional prepreration: If you want to test the inference on dreambench at the first time, you should clone the submodule dreambench to download the dataset.
git submodule update --init
Then running the following scripts:
# inference on dreambench ## for single-subject python inference.py --eval_json_path ./datasets/dreambench_singleip.json ## for multi-subject python inference.py --eval_json_path ./datasets/dreambench_multiip.json
# evaluated on dreambench ## for single-subject python eval/evaluate_clip_dino_score_single_subject.py --result_root <your_image_result_save_path> -save_dir <the_evaluation_result_save_path> ## for multi-subject python eval/evaluate_clip_dino_score_multi_subject.py --result_root <your_image_result_save_path> -save_dir <the_evaluation_result_save_path>
If you want to train on UNO-1M, you need to download the dataset from HuggingFace, extract and put it in ./datasets/UNO-1M. The directory will be like:
βββ datasets β βββ UNO-1M β βββ images β β βββ split1 β β β βββ object365_w1024_h1536_split_Bread_0_0_1_725x1024.png β β β βββ object365_w1024_h1536_split_Bread_0_0_2_811x1024.png β β β βββ ... β β βββ ... β βββ uno_1m_total_labels.json
Then run the training script:
# filter and format the dataset python uno/utils/filter_uno_1m_dataset.py ./datasets/UNO-1M/uno_1m_total_labels.json ./datasets/UNO-1M/uno_1m_total_labels_convert.json 4 # train accelerate launch train.py --train_data_json ./datasets/UNO-1M/uno_1m_total_labels_convert.json
We integrate single-subject and multi-subject generation within a unified model. For single-subject scenarios, the longest side of the reference image is set to 512 by default, while for multi-subject scenarios, it is set to 320. UNO demonstrates remarkable flexibility across various aspect ratios, thanks to its training on a multi-scale dataset. Despite being trained within 512 buckets, it can handle higher resolutions, including 512, 568, and 704, among others.
UNO excels in subject-driven generation but has room for improvement in generalization due to dataset constraints. We are actively developing an enhanced modelβstay tuned for updates. Your feedback is valuable, so please feel free to share any suggestions.
We open-source this project for academic research. The vast majority of images used in this project are either generated or licensed. If you have any concerns, please contact us, and we will promptly remove any inappropriate content. Our code is released under the Apache 2.0 License,, while our models are under the CC BY-NC 4.0 License. Any models related to FLUX.1-dev base model must adhere to the original licensing terms.
This research aims to advance the field of generative AI. Users are free to create images using this tool, provided they comply with local laws and exercise responsible usage. The developers are not liable for any misuse of the tool by users.
For the purpose of fostering research and the open-source community, we plan to open-source the entire project, encompassing training, inference, weights, etc. Thank you for your patience and support! π
./template).ComfyUI
We thanks the passionate community contributors, since we have reviced many requests about comfyui, but there aren't so much time to make so many adaptations by ourselves. if you wanna try our work in comfyui, you can try the above repos. Remember, they are slightly different, so you may need some trail and error to make find the best match repo for you.
If UNO is helpful, please help to β the repo.
If you find this project useful for your research, please consider citing our paper:
@article{wu2025less,
title={Less-to-More Generalization: Unlocking More Controllability by In-Context Generation},
author={Wu, Shaojin and Huang, Mengqi and Wu, Wenxu and Cheng, Yufeng and Ding, Fei and He, Qian},
journal={arXiv preprint arXiv:2504.02160},
year={2025}
}
RetroSearch is an open source project built by @garambo | Open a GitHub Issue
Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo
HTML:
3.2
| Encoding:
UTF-8
| Version:
0.7.4