82 new long-tail 3D rehearsals

The remaining consolidated catalog is ready for scene review. Image generation is awaiting human approval.

Open the interactive 82-scene catalog →

FLUIDAGENTIC / NAVSIM · Roadwork image review

Four long-tail full scene rehearsals

Scenes approved · temporal images ready for review

Review wheelchair crossing, police checkpoint, cargo spill and retrieval, and windblown hat retrieval using the same long-scene pipeline as roadwork. Each approved rehearsal covers 47 seconds. Each temporal image plan includes a 1 Hz base and additional critical moments. Review the generated image sequences below.

Generation pipeline

Original NAVSIM camera photograph → Image 2.5 inpainting, conditioned by a same-time view from the full Higgsfield scene. Each guide includes posture, facing direction and current speed. Clear geometry offsets are allowed when the action and relationships remain right.

New unique image moments: 200; selected after assistant visual review: 200/200. Earlier roadwork selected moments: 50. Combined selected total: 250. Completed attempts and unique samples are counted separately. The earlier roadwork batch and its generation history remain available. All playback data is hosted with this site; it does not depend on localhost or a GPU server.

Four-scene manifest

Completed image phase · 250 selected samples

ScenarioSelected momentsCompleted attemptsFirst attempt passes
Wheelchair crossing505149/50 (98.0%)
Police checkpoint506139/50 (78.0%)
Cargo spill and retrieval507132/50 (64.0%)
Windblown hat retrieval505249/50 (98.0%)
Workers clearing cones (earlier batch)5054 documented, including 4 reused critical outputs42/46 new moments (91.3%); critical pilot rate unavailable

The four new scenarios have 200 selected moments from 235 completed image attempts. First attempt pass rate: 84.5%. Attempt acceptance: 201/235 (85.5%); selected yield: 85.1%. The accepted alternate hat attempt is retained but its timestamp counts once. Six transport failures produced no image and are excluded from image acceptance.

All image outputs await human review. Training admission remains closed. These are independent edits of original NAVSIM photos; exact background pixels and temporal appearance continuity are not proven. Token usage, billed cost and average cost per image are unavailable from the native tool. Image 2.5 was requested; actual backend model readback is not exposed.

Completion and integrity evidence · 250 selected sample index