All four long-tail rehearsals · Roadwork image review
Cargo spill and retrieval · full scene
Scene approved · generated images awaiting human reviewThe full 47-second Higgsfield rehearsal spans world −7 to +40 seconds: approach, hazard, clearance and departure. Scrub the timeline, switch the calibrated camera views and inspect the approved critical actions alongside the generated photo edits.
Higgsfield revision 3 · 941 exported poses checked at 20 Hz · 3 critical actors · 39 NAVSIM source actors · one synchronized playback clip.
Loading full animated scene…
Critical moments
- Approximate NAVSIM mapped environment; native photo appearance will be preserved during inpainting.
- Critical actor roots follow the GPT6 continuous NAVSIM world; all source actors and cameras share one clip.
- The latest driver keyframes end outdoors beside the cab; a stale boarding event is corrected to match those poses.
Approximate offsets are acceptable. Check action, relative positions, facing, posture and occlusion. These blockout renders are position references; the real camera environment will come from the original photograph.
Rendered inspection views
The downloaded Higgsfield geometry and animation were rendered with Blender 5.2 locally after remote multi-preview queries reached a provider connection limit. The original GPU host also closes SSH connections. Numeric object names were aliased for import, camera principal-point shifts were restored, and daylight was added for visibility. The committed source is preserved. No video extraction or image generation is used here.
Original camera context
Temporal image generation
50 original-photo edits are planned from world -2 to +30 seconds, with a 1 Hz base and additional critical moments. Generated unique moments: 50; selected after assistant visual review: 50/50. Human output approval is pending.
Input 1 is the unwarped original NAVSIM camera photograph. Input 2 is the same-time view of this long Higgsfield scene, annotated with facing direction and instantaneous speed. Image 2.5 edits the original photo while preserving its camera framing and surroundings.
The source photograph's capture time is distinct from synthetic scene time. Camera choice and native pose differences are recorded per timestamp. Training admission remains separate from scene approval.
Temporal image gallery
50/50 selected moments. Open the full interactive image gallery, input references and attempt history →
Open editable Higgsfield project · Download full animated GLB · Editable Blender source ZIP · Review evidence · Exported contact checks