Skip to content

armgit / atas

Public Imported from GitHub
0

Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation

2.9 KB
Raw Blame

Geometry service HTTP API

The client (client/atas/world/geometry_client.py) is the only caller. All /v1 routes accept Authorization: Bearer $GEOMETRY_API_TOKEN when a token is configured. Identifiers (run_id, story_id, scene_id, shot_id, event_id) are [A-Za-z0-9_-]+.

Health

  • GET /healthz → {"status": "ok"} while the process serves requests.
  • GET /readyz → 200 with "ready": true when the Vista4D code, its Python, the Pi3X and SAM3 weights, a writable hot root and nvidia-smi are all present; 503 and the failing check otherwise.

Jobs

POST /v1/jobs creates a job and returns {"job_id": ...}. Identical requests return the same job (a request fingerprint); retry_failed: true re-queues a failed one. Poll GET /v1/jobs/{job_id} until status is succeeded or failed. While a job waits for a card its stage is waiting_gpu; admission is by free GPU memory.

Reconstruction (multipart form)

Fields: video (the reviewed clip), run_id, story_id, shot_id, scene_id, event_id, duration, world_time_start, dynamic_keywords (JSON list of SAM3 class nouns observed in the clip), entity_registry, review_lineage, publication_mode=reviewed_candidate_v1.

The service reconstructs the clip (Pi3X + SAM3, dense per-frame RGB-D and dynamic masks), then either builds the scene's initial world (first observation) or registers the clip to the committed scene world with anchored fusion. The result's fuse block reports registered (the paper's r) and, when registered, a candidate_job_id with publication_status: pending_review. Nothing is published by reconstruction alone.

Render (JSON, type: "render", camera_contract: "semantic-camera-v1")

Renders the committed scene memory along a planned camera: geometry_layers: "l1" for ordinary shots, "l1+l2" with render_mode: "recapture", source_shot_id, source_event_id, source_time_start/end for a historical event (static L1 and the source clip's time-local dynamic points share one depth buffer). camera_plan comes from the visual planner after /v1/camera-context; camera_relation is continuous, cut_recompose or recapture. Fetch the result with GET /v1/jobs/{job_id}/artifacts/memory_input.mp4.

Camera context

POST /v1/camera-context returns the observation the visual planner grounds subjects in (source shot, frame index, crop and sampling metadata, current world version) and a context_id. A render whose plan carries a stale context_id is refused.

World publication

  • POST /v1/world-publication publishes a registered candidate as the next world version. It is the state-update check (a): the film review must certify the exact reconstructed bytes, the review and reconstruction lineage must agree, the candidate must belong to the requested scene, and its parent version must still be current. A stale parent returns 409.
  • POST /v1/world-rebase re-registers such a candidate against the newer world.