ykeee/anycam-facezoom

AnyCam MultiCamVideo face-zoom camera-control bank — Wan2.2-Fun-5B (Wan2.2 VAE)

Pre-encoded training data for camera-controlled human video generation with a zooming target camera. Every clip is a MultiCamVideo scene seen through a target camera whose framing changes over the 81 frames (zoom in, zoom out from a close start, in→out, out→in) while the camera itself follows its real trajectory. The geometry condition is the camera-control condition of the earlier MultiCamVideo banks: the SAM 3D Body mesh fitted on cam10's frame 0 is FROZEN and rendered from the target camera's real per-frame extrinsics under the per-frame zoom window intrinsics, as a CANON normal map. Latents are for Wan2.2-Fun-5B-InP (Wan2.2 VAE, z=48, 16x spatial), shape (1, 48, 21, 30, 52) bf16 (81 frames, 832x480, tiled=False).

What is in a scene

data/shard-NNNN.tar are plain tars; members are <focal>__<scene>/...:

<key>/<slot>_<type>_<cam>/latents.pth      {rgb_latents, canon_latents}  bf16 (1,48,21,30,52)
<key>/<slot>_<type>_<cam>/clip.json        81 crop windows (x1,y1,x2,y2 in the 1280x1280 source), 81 K_output, zoom stats, sha256 of the source mp4 and of the stage-1 crop mp4
<key>/<slot>_<type>_<cam>/encode.json      recipe id, latent sha256, canon sha256, render check (lambda, placement check, mesh area / inside-window), host, code sha
<key>/<slot>_<type>_<cam>/render.json      condition render record
<key>/<slot>_<type>_<cam>/cond_preview.mp4 the condition frames, libx264 crf 14 (preview only; the latents were encoded from the lossless PNGs)
<key>/sources.json                         the 2 SOURCE cameras of the scene, BY REFERENCE to the 5-camera bank (see below)
<key>/scene.json                           stage-1 record: dropped clips with reasons, source video sha256
<key>/render_complete.json, encode_complete.json

type is one of in (zoom in), out (start close, zoom out), inout, outin; slot 0..3; cam the MultiCamVideo camera (cam10 = the static camera, cam01/02 pan/tilt, cam03..09 moving). Up to 4 target clips per scene: cam10 + 3 cameras drawn from those with a usable face track, the four types permuted over them. A scene whose 4 clips were all dropped (face left every window) is present with encode_complete.json {"empty": true} and no clip dirs.

Sources (not re-encoded)

sources.json names 2 source cameras per scene (cam10 + one random moving camera of the old 5-camera bank) and their items in ykeee/anycam-multicamvideo-480-5cam-wan22-5b-latents (scenes/<key>/camXX/latents.pth, same keys/shape/dtype). Those latents were encoded from the same original pixels with the same crop at zoom 1.0 (the full-width 1280x738 window), so a zoom clip's zoom-1.0 frames are bit-identical to its source's input pixels. Four-way training sample for (target clip, source cam):

trainer key	file
video_latents (target)	<clip>/latents.pth['rgb_latents']
video_cond_latents (source video)	5-cam bank scenes/<key>/<src>/latents.pth['rgb_latents']
pose_cond_latents (target geometry)	<clip>/latents.pth['canon_latents']
source_pose_cond_latents (source geometry)	5-cam bank scenes/<key>/<src>/latents.pth['canon_latents']

Any of the 2 sources can be paired with any of the scene's target clips (the target may be the same camera as the source).

How the windows were chosen (plan mcv_zoom_plan_v4.1, plan.jsonl in this repo)

• Tight end: window height = L x head height (L ~ U[1.3, 2.0], head = 50–77% of the output height), head top at 10% of the window; magnification relative to the zoom-1.0 window capped at 3x (stats.clamped); wide end = 0.8–1.0 of the zoom-1.0 window; wide/tight ratio >= 1.5. Window centre = the 81-frame median of the camera's real face track; the real face stays inside the window in >= 90% of the frames (stats.face_inside), otherwise another camera / a wider window was tried, else the clip was dropped.
• The 81 windows are interpolated between key rects (log-height, linear centre, random easing and holds, random turn frame for inout/outin). Per frame, K_output = [[fx*s, 0, (cx-x1)*s], [0, fy*s, (cy-y1)*s], [0,0,1]] with s = 832 / (x2-x1); depth is unchanged by the crop, so downstream geometry must use the per-frame K_output in clip.json.
• Pixels: cv2.warpAffine with the uniform scale s, INTER_LINEAR, BORDER_REPLICATE, from the original 1280x1280 frames.

Statistics (stage 1, all 13,327 scenes)

Clips per focal: f18 12,819 / f24 12,597 / f35 12,196 / f50 11,335 (48,947 total; 4,361 planned clips dropped, 180 scenes empty). Share of clips clamped at 3x: f18 94%, f24 82%, f35 57%, f50 33% (an 18 mm scene rarely reaches a head-filling close-up within 3x). Head height at the tight end as a fraction of the output height, p10/50/90: f18 .15/.29/.52, f24 .20/.38/.60, f35 .28/.52/.68, f50 .40/.58/.74.

Files

data/shard-NNNN.tar (grouped by completion order, ~250 scenes each), shards.json and manifest.json (uploaded LAST = complete) map every scene key to its shard with sha256 + size; plan.jsonl (the per-scene plan incl. windows), plan_summary.json, stage1_stats.json, encode_recipe.json (VAE file + sha256, crop + condition recipe, LUT/faces sha), code/ (the scripts). Condition renders use the SAM 3D Body (MHR) template; the frozen meshes come from the earlier camera-control bank.

Source and license

MultiCamVideo-Dataset (Bai et al., ReCamMaster) — the original videos' license applies to anything derived from them; SAM 3D Body / MHR assets under their own licenses. Access is gated: request access and state the intended use.