RectifID

Code for RectifID: Personalizing Rectified Flow with Anchored Classifier Guidance.

New Updates

• [2024-06] Our code now supports SDXL. See single_person_sdxl.py and the results.

  

Requirements

• A GPU with a minimum of 12.5GB memory.
• Install insightface, onnxruntime-gpu, onnx2torch, kornia, deepface, and the dependencies of PeRFlow.
• Download antelopev2 and put the .onnx models in this folder.
• To use 2-rectified flow, add the parameter callback_on_step_end to InstaFlow/code/pipeline_rf.py according to this code.

Getting Started

• For single-face personalization, run the following command:

  python single_person.py --model FLOW_NAME --ref REF_IMAGE --prompt PROMPT_LIST --out OUT_IMAGE_LIST

• For single-object customization,, run the following command:

  python single_object.py --model FLOW_NAME --ref REF_IMAGE --ref_name REF_NAME --prompt PROMPT_LIST --out OUT_IMAGE_LIST

• Alternatively, you can explore the jupyter notebooks on your own:

  • single-person: single_person_perflow.ipynb and single_person_2rflow.ipynb
  • single-object: single_object_perflow.ipynb and single_object_2rflow.ipynb
  • multi-person: multi_person_perflow.ipynb
  • multi-object: multi_object_perflow.ipynb

• For face-centric generation, please include the trigger word "person" in your prompt, and optionally append ", face" at the end.

Results

The following results were obtained with PeRFlow based on Stable Diffusion 1.5:

Single-person	Single-object

Multi-person	Multi-object

Please see the paper appendix for more results.

Citation

If you find this code useful, please consider citing:

@article{sun2024rectifid,
  title={RectifID: Personalizing Rectified Flow with Anchored Classifier Guidance},
  author={Sun, Zhicheng and Yang, Zhenhao and Jin, Yang and Chi, Haozhe and Xu, Kun and Xu, Kun and Chen, Liwei and Jiang, Hao and Zhang, Di and Song, Yang and Gai, Kun and Mu, Yadong},
  journal={arXiv preprint arXiv:2405.14677},
  year={2024}
}

Acknowledgement

This code is based on PeRFlow and InstaFlow, thanks to their efforts in open-sourcing.

Related projects worth checking out: DOODL, FlowGrad, DNO, InitNO, DSG, D-Flow, RFDS, ReNO.