Fast_deploy / ComfyUI-MiniMaxH3-Context-Loop /ComfyUI-MAINodes /docs /roadmaps /persistent_world_status.md
|
Download ComfyUI-MiniMaxH3-Context-Loop/ComfyUI-MAINodes/docs/roadmaps/persistent_world_status.md from dzmdd/Fast_deploy: direct link, hf CLI and curl.
- Browser
- Download file 2.83 kB
-
https://huggingface.co/dzmdd/Fast_deploy/resolve/main/ComfyUI-MiniMaxH3-Context-Loop/ComfyUI-MAINodes/docs/roadmaps/persistent_world_status.md
- Command line
-
hf download hf://dzmdd/Fast_deploy/ComfyUI-MiniMaxH3-Context-Loop/ComfyUI-MAINodes/docs/roadmaps/persistent_world_status.md
-
curl -L -o persistent_world_status.md https://huggingface.co/dzmdd/Fast_deploy/resolve/main/ComfyUI-MiniMaxH3-Context-Loop/ComfyUI-MAINodes/docs/roadmaps/persistent_world_status.md
2.83 kB
Persistent generative world: status
Updated 2026-08-13. The full design lives in
persistent_world_design.md: H3 Ref2VA as
the center of a persistent Gaussian world memory with agentic composition,
transactional world updates, history preservation, and a 3D -> H3 -> 3D
loop. Status: design only, nothing built.
What exists today
- The design thesis itself (1,900 lines, argued end to end).
- Building blocks already shipped in this pack that the design assumes:
the v2v source path (
H3 Video Fit), seam and audio clock discipline, rolling-window regeneration for long or large work, and the Ref2VA full-reference prompting contract.
Not built
Everything stateful: the persistent Gaussian store, world-update transactions, change detection, local coordinate systems, the QuerySplat-class lift, and any engine bridge.
First steps, cheapest first
- Single-scene loop probe: render a scene, lift it to splats, re-render from a new viewpoint via Ref2VA, and measure what survives (identity, layout, lighting). One scene, no persistence, answers whether the loop's core assumption holds.
- Confidence plumbing: cross-run divergence maps and oracle profiles as per-region confidence, feeding the design's confidence-is-first-class requirement.
- Dependencies from the alpha ledger: continuation UX and SAM-based
targeting (see
ATOS_ALPHA_STATUS.md) are the same machinery this design will lean on for world updates.
Mechanism unifications (2026-08-14 synthesis)
Three findings from the motion program supply machinery this design was missing:
- Per-token noise labels as the world's read/write interface. Store per-region confidence in the world memory; express it back to the model on re-render as partial denoise labels: well-observed regions arrive nearly finished and are preserved, uncertain regions arrive noisy and are invented. Confidence-weighted regeneration IS the transactional update semantics; a label ramp at the boundary is the handover between known and unknown world. This also resolves the "lift must not directly replace the world" tension: lift confidence maps to noise level, the model adjudicates.
- One world clock. The positional time axis is physical (frame- linear), and generation supports arbitrary time origins. Rendered clips can carry their true world-time origin and become temporally addressable in the world's coordinate frame instead of each starting at t=0.
- The budgeted-run shape generalizes. Budgeted pieces + a banked, resumable store + an explicit seam policy + priced pre-run reports: the rolling window proves the shape over time; spatial tiles and chained scene atoms are the same shape over space and sequence. Build them on that skeleton.