ymhao/hyperscene_traj

hyperscene_v0

Surface-normal control videos of annotated indoor scenes, with per-pixel instance ids, camera trajectories, six-section captions, LynnReal-Omni RGB renders, the scene meshes, and the packed MiniMax-H3 Ref2VA training split. Every clip is 832x480, 24 fps, 124 frames (5.17 s).

Selection

The clips are InsWindowFilt/select_final.json from this repository's earlier layout (commit b9305a2, since replaced), copied as select_final.json: 10,219 clips over 2,950 scenes that passed three filters in turn:

1. black fraction, edge density, inst_win_min >= 3 (24,560 of 46,309 renders);
2. trajectory dedup within a scene: position Hausdorff < 0.02 and mean per-frame heading < 5 deg;
3. first frame: n0 >= 5 instances of at least 0.1% area, c0 >= 0.6 of their area seen later.

One clip has one key everywhere: <scene>-<traj>, e.g. sk_00061-traj_05 (upstream writes sk_00061_traj_05).

Layout

select_final.json                  the clip list above, byte copy of upstream
raw/shards/raw-NNNNN.tar           10,219 clips in 18 shards of ~2 GiB, stock webdataset
raw/samples.jsonl                  key, scene, traj, shard, split, has_render
train/  val/                       Ref2VA training split, 2,764 / 33 clips (H3-Train format)
render/<scene>/<scene>.tar         5,154 LynnReal RGB videos, one tar per scene
scene/<scene>.mesh.meshopt.zst     2,950 scene meshes
tools/                             captioning tools (caption2step, video_caption); Hub only
gpt-6/                             archive of a normal-caption RGB experiment; Hub only
raw_v2_1016/                       20,044 more clips, new planner, normal + RGB + caption; see below

raw

Members of one sample K:

member	content
K.normal.mp4	surface-normal video, H.264. World-space normals, RGB = (n*0.5+0.5)*255; black = ray miss
K.traj.npz	pose (124,4,4) camera-to-world as upstream stores it (scene normalized to the unit sphere, Y up, OpenCV columns right/down/forward/position); camera_to_world (124,4,4) the same cameras in the mesh frame; intrinsics (3,3), fx = fy = 554.3622, cx = 415.5, cy = 239.5
K.instance.mkv	per-pixel instance id, FFV1 gray16le, lossless. Value = mesh instance id + 1; 0 = no instance (miss, or walls, floors and other unlabelled faces)
K.instance.json	instances: {id, frame_pixels} for every instance seen in some frame, sorted by id; frame_pixels is its pixel count in each of the 124 frames
K.meta.json	key, scene, traj, num_frames, fps, width, height, caption, scores, split, has_render

caption is the six-section Ref2VA prompt (template v2) without its trailing newline; null for the 5,065 clips never captioned. train/val were encoded from rewrite_prompt(caption) (H3-Train data/hyper/add_first_frame.py), which adds the <Picture 1> first-frame wording. scores holds every upstream InsWindowFilt field (blk, edge_density, inst_win_min, inst_win_max, inst_mean, inst_frame_min, inst_win, win, hit_frac, frames_scored, path, path_m, step_std, distinct_colors, window_start) plus n0, a0, c0 from the selection. Upstream defines only the filter fields and inst_win_min; step_std and distinct_colors are undefined there. split is train, val or null; has_render says whether render/ holds the clip.

instance.mkv was rasterized from scene/ with the cameras and intrinsics above (nvdiffrast, near plane 0.001). On a 31-clip check it agreed with normal.mp4 on ray misses everywhere except within 2 px of an edge.

train / val

The Ref2VA video-to-video split H3-Train trains on, restricted to the selection: the first-frame version (references: [picture, video]). Train clips also passed 10-50 visible instances and at most 3 per scene; val is 8 whole held-out scenes, unfiltered. Each shard shards/hyper-NNNNN.tar has a .utix index (h3_train.utix); each sample is a folder K/ with

• sample.safetensors: picture f32 (1,390,96), normal f32 (1,14430,96), rgb f32 (1,14430,96), H3 video-VAE rows of the first RGB frame, the normal video and the RGB target; encoder_hidden_states bf16 (1,L,5120), the Qwen3-VL conditioning of the caption, picture and control video;
• meta.json: scene, traj, geometry and source paths at packing time;
• rope.safetensors: the clip's rel_cam_rope coverage (coverage_entries, coverage_instance, coverage_pixels, visible_raw_ids, object_poses, object_extents, camera_extrinsics). Its file-level parameters are the manifest's rope block. Its object boxes are PCA-oriented, from before H3-Train switched to world-axis boxes.

Stock webdataset splits these folder members into separate samples; group them by folder (for example with rename_files) or read through h3_train.utix.

render

render/<scene>/<scene>.tar holds K.mp4 for the scene's selected clips that LynnReal-Omni rendered: the normal video turned into RGB with the clip's caption (weights weight/standard, seed 7, CRF 18), H.264 yuv420p, no audio.

scene

zstd-compressed pickle: meshopt-encoded vertices (V,3) f32 and faces (F,3) u32, instance_ids (K,) and instance_offsets (K+1,): instance k owns faces offsets[k]:offsets[k+1], faces past the last offset belong to none. Metres, Z up. H3-Train's data/hyperscene_process/fetch_instance_mesh.read_instance_mesh decodes it.

Built by

workspace/H3-Train/data/hyperscene_v0/ (render_instances, pack_raw, pack_split, pack_render, pack.sbatch), 2026-10-09.

raw_v2_1016

A second batch of clips from the same SketchUp scenes, made with the visibility-driven trajectory planner (coverage-driven walkthroughs, human eye height 1.6 m, arc-length-smoothed pacing) instead of the planner behind raw/. Same delivery spec: 832x480, 24 fps, 124 frames (5.17 s). The trajectories are not the ones in raw/: sk_00001-traj_01 here and in raw/ are different camera paths, so never join the two directories on key.

raw_v2_1016/shards/raw-NNNNN.tar    20,044 clips over 9,190 scenes in 34 shards of ~2 GiB, stock webdataset
raw_v2_1016/samples.jsonl           key, scene, traj, shard, has_rgb

Members of one sample K (<scene>-<traj>):

member	content
K.normal.mp4	surface-normal video, same renderer and encoding as raw/ (world-space normals, RGB = (n*0.5+0.5)*255, black = ray miss)
K.rgb.mp4	Cycles path-traced RGB of the textured scene along the same cameras (6,339 clips; absent otherwise). OptiX, 32 spp adaptive (0.01), OptiX denoiser (albedo+normal), AgX, exposure fixed per clip, rendered at 1664x960 and Lanczos-downsampled to 832x480, unsharp=3:3:0.6, H.264 CRF 16 yuv420p
K.traj.npz	pose, camera_to_world, intrinsics, identical conventions to raw/ (normalized Y-up OpenCV c2w; the same cameras in the mesh frame, metres, Z up; fx = fy = 554.3622, cx = 415.5, cy = 239.5)
K.meta.json	key, scene, traj, num_frames, fps, width, height, caption, caption_info, scores, has_rgb

caption is a single-call photoreal prompt written from 4 normal-map frames (work/caption_prompts.py SYSTEM, template v3, writer reasoning medium, temperature 0.3), split about evenly between deepseek/deepseek-v4.1-flash and moonshotai/kimi-k3 by a stable hash of the key. It is not the six-section Ref2VA prompt of raw/. caption_info keeps the rest of the captioner's output verbatim (model, negative, scene_type, ...); its camera and strength fields are leftovers of an earlier LTX recipe and say nothing about the clip.

Selection: every clip here has a caption and passed, on the delivered 124-frame window, blk <= 0.30, edge_density >= 0.02, inst_win_min >= 3, n0 >= 5, c0 >= 0.6 (same definitions as raw/; the blk/edge cut is stricter than raw/'s tiered one). No separate trajectory dedup was run: the planner targets content no earlier trajectory filmed. scores holds blk, edge_density, window_start, path, path_m, step_std, distinct_colors, inst_win, inst_win_min, inst_mean, inst_frame_min, inst_frame_max, n0, a0, c0.

No instance.mkv / instance.json and no render/ entries for this batch; scene/ covers only the scenes of raw/. Known issue shared with raw/: mirrored components in the source meshes have inverted face winding, so their normals point inward.

Appended 2026-10-17: shards raw-00023 to raw-00033 hold the 6,205 clips captioned after the first upload; shards raw-00000, raw-00005 to raw-00008 were rewritten in place to add the K.rgb.mp4 rendered since (same samples, same order, other members byte-identical; only has_rgb in meta.json changed). RGB rendering is still running, so later appends may rewrite more shards the same way. Five captions (caption_info.max_tokens == 3000) were written with a 3000-token cap after repeatedly running out of the default 1600.