Unofficial PyTorch Implementation for FlashFace. The work is a ReferenceNet absed zero shot Identity Personalization.

This project is the minimal implementation of the flashface and still work in process.

The generate result based on our pretrained model with prompt: a woman with a flower in her hair, white dress, looking at viewer, flower, hair ornament, realistic, blue background, hair flower, simple background, upper body

From left to right means use 1 to 4 faces. Face blend

• [2024-04-07]: add inference code and upload pretrained model.
• [2024-04-06]: Init repo and upload training code.

torch>2.0
transformers==4.34.1
diffusers==0.22.1
accelerate==0.23.0

prepare jsonl file for data. Each line should be a json string with following key and value:

• path: for ground truth image path.
• size: (width, height) tuple of ground truth image size.
• caption: caption for ground truth image.
• ref: list of reference face path.

{"path": "path/to/image.jpg", "size": [512, 512], "caption": "a woman holding flowers, white dress, looking at viewer, black hair, black eyes, realistic", "ref": ["path/to/face1.jpg", "path/to/face2.jpg", "path/to/face3.jpg"]}

PRETRAINED_MODEL=""
accelerate launch --multi_gpu --main_process_port=21634 --mixed_precision=fp16 train.py \
    --pretrained_model_name_or_path=$PRETRAINED_MODEL \
    --output_dir output \
    --metafiles data.jsonl \
    --clip_skip 2 \
    --proportion_empty_prompts 0.1 \
    --proportion_empty_face 0.0 \
    --save_steps 20000 \
    --resolution=512 \
    --learning_rate=5e-6 \
    --train_batch_size=8 \
    --dataloader_num_workers=6 \
    --num_train_epochs=20 \
    --mixed_precision=fp16 \
    --seed 42

We train a model with about 400M samples on 8 x A100-80G with total batch size 64. Please download it from here ! For inference, use insightface to crop align face firstly, and then modify inference.py to run. At least 8GB VRAM required.

python inference.py

• magic-research/magic-animate

@misc{zhang2024flashface,
      title={FlashFace: Human Image Personalization with High-fidelity Identity Preservation}, 
      author={Shilong Zhang and Lianghua Huang and Xi Chen and Yifei Zhang and Zhi-Fan Wu and Yutong Feng and Wei Wang and Yujun Shen and Yu Liu and Ping Luo},
      year={2024},
      eprint={2403.17008},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}