ATAS.monster

Method / 01

A persistent world
behind every shot.

ATAS turns a high-level screenplay into bounded requests, retrieves the right spatial and temporal evidence, and accepts new evidence only when it passes validation.

The pipeline

Compile. Query. Generate. Check.

The central object is a world state: a pair of memories that can be queried before the next shot and updated after an observation is judged reliable.

01Compile

Split the screenplay into bounded shot requests with a scene, anchor, camera, time and optional event.

02Query

Retrieve scene geometry and event evidence that match the requested space and moment.

03Generate

Condition a short-video generator on the local shot prompt and retrieved context.

04Validate

Keep the clip for delivery, updating memory only when the observation passes its checks.

Two complementary memories

Different evidence needs different persistence.

Static geometry answers “where is this?” Dynamic evidence answers “what happened, and when?” ATAS keeps these roles separate so a visually similar frame does not stand in for the wrong event.

⌂Scene memory / S

Recover the place.

Accepted static geometry is registered to a shared scene coordinate frame. A later request can render that geometry from a new camera for a spatial revisit.

  • Registered point-cloud evidence
  • Shared coordinate frame
  • Dynamic regions excluded
◷Event memory / E

Recall the moment.

Dynamic observations stay tied to a scene, event identity, source observation and time interval. A later request can ask for the earlier event from a different view.

  • Frame-level RGB-D and masks
  • Dynamic point-cloud evidence
  • Scene and time indexed

The admission rule

A film clip is not automatically a world fact.

ATAS separates film delivery from memory admission. This prevents a plausible but misaligned generation from contaminating later shots.

f
Film review

Is this clip good enough to deliver?

r
Registration

Does the reconstructed observation align to the scene frame?

a
Admission

Does it have valid provenance and state scope?

f = r = a = 1→ update the next world state

A request has coordinates

Every shot carries enough context to be found again.

ATAS does not treat a screenplay as a flat list of prompts. It compiles each request into a structured description of local intent, scene, spatial anchor, camera, time and event.

bounded shot request
q = { description, scene,
  spatialAnchor, camera,
  timeInterval, event? }

Read the result carefully

A memory system still has a horizon.

The paper reports gradual quality decline as requests move farther from the latest accepted state. Registration quality, stale evidence and exact action-state reproduction remain open problems.

See the reported limitations