Async Multi-Agent State Machine (Pure Python)

This is a minimal, framework-free skeleton for a multi-agent "video world" system:

• WorldState holds a dictionary of ObjectState entries keyed by object id
• Director broadcasts SceneObjective messages to ObjectAgent inbox queues (asyncio.Queue)
• Each ObjectAgent runs an async state machine (idle → acting → stopped) and uses an injected JSONLLM to propose the next 24 frames as structured JSON

Run the demo

python3 -m mvg.demo

Blender (Headless) World Executor

blender/world_executor.py reads a per-frame WorldState buffer (JSONL or JSON), applies object transforms, and renders a visible headless frame sequence.

If a compositor tree is available, it also writes technical passes; otherwise it falls back to RGB frames so the demo still produces visible output.

Integrated Demo: Agents → Director → Physics → WorldState Buffer

This produces a world.jsonl buffer (one frame per line) that you can feed into the Blender executor.

python3 -m mvg.demo_pipeline --out-jsonl /tmp/mvg_world.jsonl --frames 24

Example JSONL (one line per frame):

{"frame":1,"objects":{"car_1":{"id":"car_1","position":[0.0,0.0],"velocity":[1.0,0.0],"current_action":"moving"}}}

Run headless:

blender -b --factory-startup -P blender/world_executor.py -- \
  --worldstate-jsonl /abs/path/world.jsonl \
  --output-dir /abs/path/out \
  --frame-start 0 --frame-end 23

Outputs:

• out/rgb/rgb_####.png when running without a compositor tree
• out/depth/depth_####.exr
• out/normal/normal_####.exr
• out/mask/object_index_####.exr (object-id segmentation via Blender Object Index pass)

PyTorch V2V Inference (Temporal KV Cache)

scripts/v2v_infer.py is a checkpoint-agnostic PyTorch inference harness that:

• conditions generation on Depth + Object Mask (ControlNet-style)
• conditions on a text prompt
• carries the previous frame latent as a simple “Temporal KV Cache” for temporal consistency

It expects a Torch file (.pt) with tensors:

{"depth": Float[T,H,W], "mask": Long[T,H,W]}

Run:

python3 scripts/v2v_infer.py \
  --gbuffer-pt /abs/path/gbuffer.pt \
  --prompt "photoreal, cinematic lighting" \
  --out-dir /abs/path/out \
  --checkpoint /abs/path/model.pt

Notes:

• The included model is a minimal ControlNet-style stub; photorealistic output requires your trained weights/model swap.
• Output frames are written as .ppm to avoid extra image dependencies.