You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is derived from the KlingTeam MultiCamVideo-Dataset. By requesting access you agree to use it only under the terms of that dataset's license and for research purposes.
Log in or Sign Up to review the conditions and access this dataset content.
AnyCam motion-control bank
Pre-encoded training data for camera-controlled human video generation with a moving target camera ("motion control": the target camera moves during the clip, so the geometry condition is rendered per frame instead of once). Every scene comes from the KlingTeam MultiCamVideo-Dataset (10 synchronized cameras per scene, 81 frames, 1280x1280; cam10 is static in every scene). For each scene this bank holds, for the static camera and 1-2 moving cameras:
- a per-frame SAM 3D Body fit of the person (81 frames per camera),
- the per-frame canonical-space normal map of that fit, rendered from the camera's own pose (the conditioning image of the AnyCam / FaceLift method), as 81 PNGs,
- Wan2.2 VAE latents (Wan2.2-Fun-5B VAE, bf16) of the cropped RGB clip and of the normal clip,
- the source pair: cam10's frame 0 repeated 81 times (RGB and normal), i.e. a still source image presented as a video, so the model sees (static source, moving target) pairs.
| scenes | |
|---|---|
| finished | 13,226 (f18_aperture10 3,343 / f24_aperture5 3,355 / f35_aperture2.4 3,317 / f50_aperture2.4 3,211) |
| planned | 13,359 |
failed and recorded in manifest.json |
133 (121: person leaves the static camera's view during the clip; 11: fewer than one moving camera keeps the person in view; 1: too few frames fitted) |
Nothing was dropped silently: every planned scene is either in data/ or listed under
failed_scenes in manifest.json with its reason.
Layout
data/shard-0000.tar ... data/shard-0052.tar plain tar (no gzip), 250 scenes each, ~9 GiB each (~470 GiB total)
manifest.json per-scene metadata for all finished scenes + the failed list (completion marker)
shards.json shard -> scene keys, size, sha256
plan.jsonl the 13,359-scene plan (static cam, planned moving cams + trajectory labels, seed)
code/ the exact scripts that produced the bank (fit_render.py, encode_motion.py, ...)
encode_recipe.json the encoder recipe: Wan2.2_VAE.pth of alibaba-pai/Wan2.2-Fun-5B-InP (sha256 recorded), encoder sha256, crop + source rules
Scene keys are <focal>__<scene> (e.g. f35_aperture2.4__scene2686), sorted; shard i holds keys
[250 i, 250 i + 250) of the sorted list of finished scenes. Each tar member is <key>/...:
<key>/
metadata.json plan record (plan.labels = trajectory type of all 10 cameras: static / arc / translation / pantilt / random;
plan.targets = the kept moving cameras with their type; plan.targets_planned[].feats = path length,
max rotation, sweep angle, orbit radius ...), per-camera crop/K/canon sha, encode recipe id +
latent sha256 per camera and for source, cameras tried
complete.json fit stage record; encode_complete.json encode stage record (latent sha256 per camera)
camera_extrinsics.json the original MultiCamVideo per-frame extrinsics of all 10 cameras (frame0..frame80)
camXX/ one dir per kept camera: cam10 (static) + the moving cameras
latents.pth {'rgb_latents', 'canon_latents'}: torch bf16, each [1, 48, 21, 30, 52]
canon.tar 81 PNGs 00000..00080: canonical-space normal map rendered from this camera, cropped like the RGB
fits.npz per-frame SAM 3D Body fit: status (int8, 0 = fitted), present, box, keypoints_2d/3d (70 mhr70 joints),
cam_t, focal, global_rot, body_pose, hand_pose, shape, expr, scale, K_original, K_output, crop_window
complete.json, encode.json, track.json per-camera fit / encode / detection-tracking records
source/
latents.pth {'rgb_latents', 'canon_latents'} of cam10 frame 0 repeated 81 times, same shape
canon_frame0.png the normal map of that frame
encode.json
Four-way conditioning
| trainer key | tensor |
|---|---|
video_latents (target) |
camXX/latents.pth['rgb_latents'] |
video_cond_latents (source video) |
source/latents.pth['rgb_latents'] |
pose_cond_latents (target geometry) |
camXX/latents.pth['canon_latents'] |
source_pose_cond_latents (source geometry) |
source/latents.pth['canon_latents'] |
Latent shape [1, 48, 21, 30, 52] = Wan2.2 VAE (z = 48, 4x temporal, 16x spatial) of an 81-frame
832x480 clip. Each moving camera of a scene forms one training pair with the scene's source/.
Crop rule
Per camera: the union of the person's detection boxes over all 81 frames, 5 % margin, expanded to
832:480 aspect, one fixed window for the whole clip (no per-frame re-framing), resized to 832x480.
metadata.json -> cameras[camXX].crop holds the window in original pixels, the scale, the affine and
the output intrinsics K_output; the same window was applied to the RGB frames and the rendered
normals, so the two latents of a camera are pixel-aligned.
Camera selection and labels
plan.jsonl draws, per scene, cam10 (static, always) plus 2 moving cameras with a type-weighted
draw over the trajectory labels (arc, translation, pantilt, random; labels in the plan
record). A planned camera in which the person is not fully visible for all 81 frames was rejected and
a replacement tried from the remaining cameras (moving_cams_planned, cams_rejected, cams_added,
tried in manifest.json). A scene was kept only if the static camera and at least one moving camera
keep the person in view for the whole clip.
Provenance
metadata.json -> encode.recipe_id is the id of encode_recipe.json (the Wan2.2-Fun-5B VAE file and its sha256, the encoder's source sha256, crop and source rules);
complete.json -> recipe_id names the fit/render recipe. The scripts in code/ are the ones that
ran. SAM 3D Body renders are not bit-reproducible (rasterisation jitter), so re-rendering a scene
gives a normal map that differs in a few hundred isolated pixels.
Verification
Each shard was uploaded with a local sha256 and accepted only after the Hub's LFS sha256 and size
matched; shards.json lists both. manifest.json was written last, after re-listing the repository
and finding every shard, so its presence is the completion marker.
Source and license
Derived from the KlingTeam MultiCamVideo-Dataset; use is subject to that dataset's license. Only derived products (fits, normal renders, VAE latents, crop metadata, extrinsics) are redistributed here; the original RGB videos are not included.
- Downloads last month
- 677