FlashPortrait (Wan2.1 I2V-14B-720P): long identity-preserving portrait animation from a reference image and a driving clip, settings and memory modes (60 GB full load, about 10 GB offloaded)

FlashPortrait: long, identity-preserving portrait animation on Wan2.1 14B What it is. FlashPortrait (Fudan University, Microsoft Research Asia, Xi'an Jiaotong University, Tencent and Alibaba's Wan team; arXiv 2512.16900) is a portrait-animation model built on Wan-AI/Wan2.1-I2V-14B-720P. It animates a reference image of a person for arbitrarily long durations using a sliding context window with weighted blending of overlapping windows, which the authors say keeps identity stable. The authors state that their demo videos use no face-swapping or face-restoration post-processing. These are author claims, not checked here. What drives the motion: read carefully. The inference arguments are --validation_image_start (the reference image), --validation_driven_video_path and --prompt. The overview says the model computes identity-agnostic facial expression features with an off-the-shelf extractor, and the card speaks of emotion synchronization with the driven video. Yet the same sentence that names --validation_driven_video_path calls it the driven audio, the CFG note says "prompt and audio cfg", and the metadata tags the model "audio-driven avatar animation". The card is therefore inconsistent about whether the driving input is a video, audio or both; check the repository's example cases before relying on audio-only driving. Speed claim versus released code. The paper title promises 6x faster generation through Adaptive Latent Prediction. At this revision the to-do list marks both "Inference Code with Adaptive Latent Prediction" and "Multiple-GPU Inference Code" as not yet released, and the news item calls the release "a basic model checkpoint" with the acceleration part to follow. The card describes fast_infer.py as faster than infer.py with the same settings. Setup as documented. Install torch 2.6.0, torchvision 0.21.0 and torchaudio 2.1.1 from the CUDA 12.4 wheel index, then pip install -r requirements.txt; flash_attn is optional. Download FrancisRing/FlashPortrait to checkpoints/FlashPortrait and Wan-AI/Wan2.1-I2V-14B-720P to checkpoints/Wan2.1-I2V-14B-720P. Run bash inference.sh, with settings in infer.py or fast_infer.py: --wan_model_name, --transformer_path and --portrait_encoder_path point to the Wan2.1 weights, the FlashPortrait DiT and the portrait encoder. Recommended settings. Resolutions 512x512, 480x832, 832x480, 720x720, 720x1280 or 1280x720, set through max_size in infer.py; the authors say higher resolution gives higher quality. Use --num_inference_steps 30 to 50 (more steps, higher quality) and --context_overlap 10 to 40 (longer overlap, higher quality but slower). Set --text_cfg_scale and --emo_cfg_scale between 2 and 5, raising the emotion scale for closer emotion sync. Prompt structure: [description of the first frame]-[description of human behaviour]-[description of background, optional]. Memory. For a 10-second 720x1280 video at 25 fps the authors report about 60 GB of VRAM on an A100 with --GPU_memory_mode model_full_load, and about 10 GB with sequential_cpu_offload at slower speed. model_cpu_offload roughly halves memory versus full load, and model_cpu_offload_and_qfloat8 is also offered. For very long videos (10k+ frames) the 3D VAE decoder needs a lot of memory, and you can run the VAE on CPU. If memory runs out, reduce the number of frames or the resolution. Training is separate and needs about 50 GB (mixed resolution) or 40 GB (512x512 only), per the authors. Rights. The card metadata declares Apache-2.0; the Wan2.1 base carries its own terms, which were not reviewed. The card states no consent or likeness policy, and this record grants no likeness clearance; obtain permission for any real person you animate. Based on the FrancisRing/FlashPortrait card pinned to revision 2c3aac13d7627a35198e268eb59bf354404d5d80 and observed on 2026-10-28. Nothing was downloaded or run. Identity, speed and memory figures are the authors' statements; nothing was run here. The card is inconsistent about the driving input (video versus audio); the repository examples and code were not inspected. The 6x acceleration code was unreleased at this revision; released-code speed is unmeasured. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/FrancisRing/FlashPortrait/raw/2c3aac13d7627a35198e268eb59bf354404d5d80/README.md",
      "note": "Retained source snapshot observed 2026-10-28; TEXT.txt; UTF-8 bytes 16:258, 173:258, 2417:2589, 3686:4684, 4041:4684, 4862:4956, 5045:5131, 5079:5131, 5469:5710, 6119:6713, 6928:8665, 7416:8665, 8161:8665, 8509:8665, 8568:8665, 9460:9614, 17980:18195, 18418:18472, 16923:17345, 69:258; exact excerpt SHA256 cbc69dfe160d845b74823d5b133450ffdb15235"
    }
  ]
}