4DAnyone × Rerun

Upload a clip of a person. GVHMR recovers the SMPL-X motion, a Wan 2.2 diffusion transformer generates six synchronized views around the subject, and every phase streams into the embedded Rerun viewer while it happens.

What you watch

The run is one four-link Gradio event chain, and the viewer's layout switches with it:

1. Motion — the source clip beside its detections. GVHMR's tracked boxes and ViTPose keypoints appear on the frame timeline as each stage completes, and the posed SMPL-X body lands at the end.
2. Diffusion — a 2×3 grid of per-view previews on a diffusion_step timeline. Turbo mode takes four steps; after each one the clean latent estimate is decoded through TAEW2.2 at eight frames per view. Every preview is stamped on both diffusion_step and frame.
3. Result — the camera rig in the canonical human world, the six generated MP4s in a grid, and the frame timeline playing on a loop.

Files

• app.py — the Hugging Face entrypoint. Standard library only; it installs the pinned Pixi CLI when absent and runs the locked app task.
• fdanyone_app.py — the whole application: validation, the ZeroGPU callbacks, the Rerun choreography, and the Gradio interface.
• download_assets.py — every CPU-side download, idempotent, run before Gradio starts.
• sync_vendor.sh and PROVENANCE.md — how fdanyone/ got here and from where.
• pixi.toml and pixi.lock — the complete CUDA 13, PyTorch 2.12.1, Rerun 0.36.1, Gradio 6.20.0 environment, locked for linux-64 and linux-aarch64.

Streaming mechanics

prepare_run and generate_run are single blocking calls that report progress through synchronous hooks. The @spaces.GPU callback therefore runs each pipeline call on a worker thread and yields BinaryStream.read() as the hooks fill an explicit RecordingStream. Yielding from inside a hook is impossible, and yielding only after the call returns would leave the viewer blank for minutes. The recording is an explicit object rather than a thread-local one, which is what makes cross-thread logging safe.

spaces.GPU runs its callback in a forked child process, and that shapes the rest. A RecordingStream cannot cross the fork — the SDK refuses to flush one whose pid has changed — so every link opens its own stream under the run's token and the viewer merges same-token streams into one recording. Run state cannot cross it either: the chain passes a picklable RunSpec, and motion and generation share one link because PreparedRun holds decoded frames and a live barrier that no pickle can carry into another worker. Inside that child the pipeline runs with inline_workers=True, because a grandchild of its own would not hold the GPU allocation.

Fixed policy

• Turbo mode: four denoising steps, FP8 DiT, the pinned Wan2.2 Turbo-LoRA, the TAEW2.2 tiny decoder, and NVDEC skeleton decode.
• Six views on one ring pitched 15° above the subject, spanning 360°. Six or fewer targets skip the RCP proposal stage entirely.
• The frozen 121-frame, 1280×704 contract. Inputs shorter than 121 frames from the chosen start time are rejected on CPU, before any GPU work.
• The fixed prompt ships as an exported embedding, so the 11 GB UMT5-XXL encoder is never downloaded and never loaded.

Run it locally

On a CUDA 13 host with HF_TOKEN set for the private asset mirror:

CONDA_OVERRIDE_CUDA=13.0 pixi run --frozen app

FDANYONE_MODEL_DIR (default models) and FDANYONE_DATA_DIR (default data) move the weights and the scratch tree. pixi run test runs the unit tests for the pure helpers.

Compatibility notes

• gradio==6.20.0, gradio-rerun==0.36.1, and rerun-sdk==0.36.1 belong together. Gradio 6.21 through 6.26 carry a frontend regression that remounts custom components on every output update, which stops the Rerun WebViewer mid-stream.
• transformers is 5.x, not the source repository's 4.57. BiRefNet is loaded through AutoModelForImageSegmentation with trust_remote_code, and its custom code only needs PretrainedConfig and PreTrainedModel. Transformers 4.x caps huggingface_hub below 1.0, which Gradio 6.20 forbids.
• Start the app through pixi run, never by calling the environment's python directly. Turbo mode compiles the DiT with Triton, and conda-forge's Triton finds a Blackwell-capable ptxas through $CONDA_PREFIX, which only the activation sets. Without it the compile fails with Cannot find ptxas-blackwell.
• SageAttention is built from source: no wheel exists for sm_120 or sm_121. TORCH_CUDA_ARCH_LIST=12.0 is set for linux-64 only, because a Space build host has no GPU to detect; linux-aarch64 detects its GB10 natively.

Credits

4DAnyone by Ant Research. GVHMR by ZJU3DV. TAEHV by Ollin Boer Bohan.