ATAS
A world-state-centric video generation agent framework. It compiles a script into shot requests with explicit spatial and temporal references, and combines retrieval-augmented generation mechanisms of scene memory and event memory to achieve cross-shot consistency and replayable reconstruction of events.
Timeline (last 90 days)
ATAS transfers the retrieval-augmented generation (RAG) paradigm to the video domain: retrieval before generation is equivalent to preparing a set description and event archive for the current shot, with the difference being that the units of video retrieval are geometric structures and dynamic observations.
The paper "Any Time, Any Space" proposes a framework named ATAS (Any Time, Any Space), which upgrades video generation from shot-by-shot stitching to scheduling around world state.
ATAS compiles the script into shot requests with explicit references, and each shot request is annotated with three types of information: shooting location (location), camera parameters (camera), and event time (event-time).
Before generation, ATAS retrieves relevant evidence from two independently maintained memories, scene memory and event memory; scene memory stores registered static geometric information, and event memory stores dynamic observation records indexed along three dimensions: scene, event, and time.
After ATAS completes generation, observations are written to memory only when both review and state checks pass.
ATAS's experimental results show that on four different video generators, ATAS improved all reported cross-shot consistency metrics and achieved the highest overall benchmark score for every backbone model.
ATAS outperforms the compared agent methods and 3D-guided methods on the spatial reproduction task for target images, and leads on three criteria for event recapture.
ATAS received the highest scores on all six human evaluation criteria.
ATAS's film delivery and memory admission are decoupled, so even if a segment is not accepted into memory, the final output is still completed independently.
ATAS's project page and code have been made public at https://lhmd.top/atas/.
2 more timeline events
All Facts (12)
ATAS transfers the retrieval-augmented generation (RAG) paradigm to the video domain: retrieval before generation is equivalent to preparing a set description and event archive for the current shot, with the difference being that the units of video retrieval are geometric structures and dynamic observations.
50%UnverifiedThe paper "Any Time, Any Space" proposes a framework named ATAS (Any Time, Any Space), which upgrades video generation from shot-by-shot stitching to scheduling around world state.
50%UnverifiedATAS compiles the script into shot requests with explicit references, and each shot request is annotated with three types of information: shooting location (location), camera parameters (camera), and event time (event-time).
50%UnverifiedBefore generation, ATAS retrieves relevant evidence from two independently maintained memories, scene memory and event memory; scene memory stores registered static geometric information, and event memory stores dynamic observation records indexed along three dimensions: scene, event, and time.
50%UnverifiedAfter ATAS completes generation, observations are written to memory only when both review and state checks pass.
50%UnverifiedATAS's experimental results show that on four different video generators, ATAS improved all reported cross-shot consistency metrics and achieved the highest overall benchmark score for every backbone model.
50%UnverifiedATAS outperforms the compared agent methods and 3D-guided methods on the spatial reproduction task for target images, and leads on three criteria for event recapture.
50%UnverifiedATAS received the highest scores on all six human evaluation criteria.
50%UnverifiedATAS's film delivery and memory admission are decoupled, so even if a segment is not accepted into memory, the final output is still completed independently.
50%UnverifiedThe article argues that ATAS represents a paradigm shift, namely viewing video generation as a world state management problem rather than a next-frame prediction problem.
50%UnverifiedThe Emmy Awards are presented by the Academy of Television Arts & Sciences (ATAS) and have historically been broadcast on a rotating basis by traditional broadcast networks such as NBC, CBS, and ABC
50%UnverifiedATAS's project page and code have been made public at https://lhmd.top/atas/.
50%