Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation
Python agentic-video-generationmutli-agentmutli-agent-systemspatial-consistencyvideo-generationvideo-generator
# Client configuration. Copy to client/.env (git-ignored; ATAS_ENV_FILE may point elsewhere).
# --- video generation: Wan 3.0 on Alibaba Cloud DashScope
DASHSCOPE_API_KEY=
# WAN30_PRIME_MODEL=wan3.0-video-prime
# WAN30_VIDEO_MODEL=wan3.0-video
# --- language/vision model and gpt-image (planning, camera grounding, prompt authoring,
# film review, identity references, opening stills). Any OpenAI-compatible gateway.
ATAS_LLM_URL=https://api.openai.com/v1/responses
ATAS_LLM_API_KEY=
ATAS_LLM_MODEL=gpt-5.1
# ATAS_LLM_SERVICE_TIER= # optional Responses service_tier
# ATAS_IMAGE_BACKEND=auto # auto | images_api | responses_tool (gateways without /images)
# AVWM_IMAGE_BASE_URL= # images API base when it differs from the Responses gateway
# AVWM_IMAGE_API_KEY=
# ATAS_LLM_STREAM=1 # stream responses (avoids proxy timeouts on long calls)
# --- geometry service (geometry/), reachable from this machine
GEOMETRY_SERVICE_URL=http://127.0.0.1:18300
GEOMETRY_API_TOKEN=
# --- pipeline budgets
# ATAS_FILM_REPAIRS=2 # regenerations after a failed film review
# ATAS_CAMERA_REPLANS=2 # camera re-plans when a memory render is unusable
# ATAS_EVENT_SOURCE_RETRIES=1 # retakes of a recapture source that was not admitted
# ATAS_OPERATIONAL_ATTEMPTS=3 # retries of a shot on transient provider/service errors