You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These are derived VAE latents computed from the MVHumanNet dataset (CUHK-SZ GAP-Lab). You must comply with the upstream dataset's licence and terms. State your affiliation and intended use below.
Log in or Sign Up to review the conditions and access this dataset content.
AnyCam — MVHumanNet 720p 4-shot encoded latents (Wan2.1 VAE, z=16)
Pre-encoded four-way VAE latents for camera-controlled multi-view video generation, computed from MVHumanNet (CUHK-SZ GAP-Lab) at 1280×720 with a 4-shot multi-camera cut convention and synthetic camera motion on both the source and target halves.
Why this set exists: it supplies the large camera rotations (65–180°, turning to the side or back) that the companion SynCam set does not contain — SynCam's camera arcs are frontal, ≤ ~65°. Measured source→target angles here reach 171°. It adds azimuth/yaw only, not pitch: both datasets' elevation distributions sit at roughly 0–30°.
Companion set, byte-format-identical: wlyu/anycam-syncam-720-4shot-latents. The two are designed to be
mixed; that is the whole reason the formats match.
Contents
49,097 samples. One data.pth per sample.
<outfit>-src<CAM>-tgt<CAM>-p6dd0ed/
tensors/data.pth 38,709,533 bytes the four latents, bf16
prompt.pth 8,390,365 bytes ONE shared file for the whole set
data.pth keys
All four are (1, 16, 21, 90, 160) bf16 — z=16, 21 = (81-1)/4 + 1 latent frames,
90×160 = 720/8 × 1280/8.
| key | content |
|---|---|
video_latents |
target-view video — what the model must produce |
video_cond_latents |
source-view video, frame-concatenated conditioning |
pose_cond_latents |
TARGET canonical-space normal map, channel-concatenated |
source_pose_cond_latents |
SOURCE canonical-space normal map, channel-concatenated |
prompt.pth holds {prompt_emb, negative_prompt_emb} once for the entire set, not per sample. It is
byte-identical to the SynCam set's, which is what makes a mixed run train both halves on the same
conditioning. A loader that guards its load with an existence check will silently serve samples with no
prompt_emb if it is absent.
Conventions that matter
Canonical-space normals, not view-space. Vertex colours come from a fixed rest-pose LUT, so front and back are distinguishable. The model reads camera rotation from the orientation difference between the source and target normals.
Every shot's geometry is its camera's frame 0, so all normals within a sample depict the same instant. Anchoring each shot at its own start frame instead folds the subject's pose change into the orientation difference the conditioning should read as viewpoint alone.
Cuts land on latent boundaries. The Wan VAE downsamples time 4×, so a cut at pixel frame s is clean
only when s ≡ 1 (mod 4); otherwise one latent frame blends two viewpoints.
Both halves have synthetic camera motion, drawn per shot and chained: the framing value is continuous across a cut while the rate and direction change. Zoom is expressed as the fraction of frame height the subject occupies (median ≈ 1.05, i.e. slightly taller than frame, so head or feet are often just clipped); subjects are deliberately more prominent than in the raw data, where they fill ≈ 59% of frame height.
Source and target share one background colour per sample and across cuts, so the only difference between the two views is viewpoint. Backgrounds are random solid colours chosen away from skin, the green screen, and the subject's clothing.
The -p6dd0ed suffix is the compatibility contract — md5(max_shots | seam grid | seed | resolution | motion version). Two conventions must never share a directory, because loaders typically glob a whole
tree and would mix them silently.
Known limitations
- ~5% of the source is unusable. 49,097 of a possible ~51,900 samples: some MVHumanNet scenes have fmask/image dimension mismatches or missing per-frame masks, and a sample is dropped whole rather than emitted partially.
- Pixel streams are not included. The companion SynCam repo ships four clean per-stream mp4s that let you re-encode at a different resolution without redoing the pipeline. This set's archived streams are an older single side-by-side file with text labels burned into the frames, which makes them unusable as a re-encode source, so they are withheld. Re-generating them is planned.
- No new elevation. See above — this set widens azimuth, not pitch.
Provenance
Derived from MVHumanNet (CUHK-SZ GAP-Lab), images_lr at 2048×1500 (48-camera group) or 1224×1024
(nominally "24cam", actually 16 cameras). 24 outfits are held out for large-angle evaluation and excluded
here. Geometry comes from SAM 3D Body (Meta) rendered at source resolution. Latents are from the Wan2.1
VAE (z=16).
- Downloads last month
- 3