ReCamMaster: re-render an existing video along a new camera trajectory (Wan2.1 port) - 10 cam_type presets, 81-frame captioned input, setup and undocumented gaps

ReCamMaster: re-render an existing video along a new camera trajectory What it does. The README says ReCamMaster re-captures in-the-wild videos with novel camera trajectories: you give it an existing clip and choose a camera move, and it generates the same scene from the new camera path. This is video-to-video re-rendering, not camera control for image-to-video generation; for that, see art_01M5MC6W980D86FMNC8Y6ZWWKM and the ComfyUI Wan Fun Camera template record art_01M5ZYWQTKFBBB93SVMVAMRKKG. Which model you get. The paper (arXiv:2503.11647; the README title calls it an ICCV'25 Oral and Best Paper Finalist) used an internal text-to-video model that the authors say they cannot release. The public repository is a reference implementation migrated to Wan2.1, and the authors warn that it may not reach the paper's performance or the demo results. For the best version the README asks you to upload your video through a form it links, which sends your footage to the authors. Setup, as documented. 1. DiffSynth-Studio needs Rust and Cargo to compile extensions; then clone the repository and run pip install -e . inside it. 2. Download the base Wan2.1 models with python download_wan2.1.py. The README does not name the Wan2.1 variant for inference; its training commands point at Wan2.1-T2V-1.3B files (DiT, umt5-xxl text encoder and Wan2.1 VAE). 3. Download step20000.ckpt from the ReCamMaster-Wan2.1 Hugging Face repository and place it in models/ReCamMaster/checkpoints. 4. Test the bundled examples with python inference_recammaster.py --cam_type 1. Your own video. - Prepare data like the example_test_data folder: N mp4 videos, each with at least 81 frames, plus a metadata.csv holding each video's path and caption. The README points to Wan2.1's prompt-extension guidance for writing captions. - Run python inference_recammaster.py --cam_type 1 --dataset_path path/to/your/data. Camera presets (cam_type). 1 Pan Right; 2 Pan Left; 3 Tilt Up; 4 Tilt Down; 5 Zoom In; 6 Zoom Out; 7 Translate Up (with rotation); 8 Translate Down (with rotation); 9 Arc Left (with rotation); 10 Arc Right (with rotation). The README says you can also generate new trajectories, but it does not document the format. Its dataset stores per-frame camera extrinsics as camera_extrinsics.json, and the validation split holds 10 trajectories in that form; the README does not say these are the cam_type presets. Not documented in the README: output resolution and fps for inference, inference GPU memory, the custom-trajectory format and its coordinate convention, and a maximum input length. The data-processing command uses 81 frames at 832x480, and the training example uses eight GPUs at batch size 1 each; neither is stated as an inference requirement. Training data. The MultiCamVideo dataset is rendered with Unreal Engine 5: 13.6K scenes filmed by 10 cameras each (136K videos), 1280x1280, 81 frames at 15 fps, with characters and animations collected from Fab and Mixamo. The README states no dataset licence. Related existing-video camera methods named by the README (pointers only, not reviewed): GCD, ReCapture, Trajectory Attention, GS-DiT, Diffusion as Shader, TrajectoryCrafter and GEN3C. Rights and identity. The repository's LICENSE at the same commit is MIT (copyright 2025 Kuaishou Visual Generation and Interaction Center; archived separately, sha256 7fe4712b0d5bbb743f3299b61d686fa4aa6455f5feed3f5f07bf7f32fb132c64) and covers the code. The Hugging Face weights repository declares Apache-2.0. Wan2.1's own licence is not restated here. You need rights to any footage you re-render. The README links use the former KwaiVGI names; Hugging Face now resolves the weights repository to KlingTeam/ReCamMaster-Wan2.1 (revision 4f3f7391743dfd25067ab27e0f1eb8928d11b91d) and GitHub resolves the code to KlingAIResearch/ReCamMaster (commit fcf98bc86e876bb534518cd99e8a65b282f0f16e). README statements only; nothing was installed or run, and no output quality was checked. Custom trajectory format, inference resolution/fps and GPU memory are not documented in the README. Code licence from the separately archived LICENSE; weights licence from the Hugging Face card; Wan2.1 and dataset terms not restated. 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://raw.githubusercontent.com/KlingAIResearch/ReCamMaster/fcf98bc86e876bb534518cd99e8a65b282f0f16e/README.md",
      "note": "Retained source snapshot observed 2026-10-29; TEXT.txt; UTF-8 bytes 3052:3376, 1719:2167, 5439:5676, 5745:5963, 6251:6331, 6444:6497, 7087:7168, 6783:7076, 7170:7299, 7239:7299, 8840:9294, 8373:8448, 14858:14915, 16620:16664; exact excerpt SHA256 b6c483f7fb469c3aa17e404dc5049bbd9ded760b34b76f0ec8777845ea9a9e94,d87e78f51a680ac9fb7ca8a77454f8759d"
    }
  ]
}