visionary-mlx

Visionary MLX

MLX video tokenizer, action-conditioned world model, and imagination-trained agent.

champion is the high-capacity configuration for the included 64px synthetic-video workload. It is not a state-of-the-art claim: release a trained checkpoint and its held-out report first.

visionary-mlx pipeline --config champion --env arena --run-dir runs/champion
visionary-mlx collect --env arena --data-dir data/eval --seed 100 --num-episodes 256
visionary-mlx benchmark --config champion --env arena --run-dir runs/champion \
  --eval-data-dir data/eval --name visionary-mlx-champion --out-dir releases/v0.2.0
python -m pytest

The benchmark command writes JSON and Markdown evidence from a separate, deterministic held-out rollout set. The shipped environments are synthetic (arena, pong, and balls), so their scores are not comparable to external vision or video benchmarks.

Release status

The champion run is complete. Release v0.2.0 contains tokenizer, dynamics, and agent safetensors, the exact configuration, a dream preview, checksums, and deterministic held-out benchmark reports. On the repository's separate synthetic arena evaluation set (32 batches), it achieved tokenizer PSNR 18.611 / MAE 0.02022 and open-loop dream PSNR 18.610 / MAE 0.02023. These are synthetic-workload results only, not an external state-of-the-art claim.

Long MLX runs checkpoint tokenizer, dynamics, and agent weights periodically; preview generation is non-fatal when local storage is exhausted.

Use a released checkpoint

The release directory contains model_config.json and three safetensors files (tokenizer, dynamics, and agent). Install MLX on Apple Silicon, then use the matching --config name and run directory:

# A dream requires a video/action rollout. Create one if you do not already
# have a compatible arena.npz dataset.
uv run visionary-mlx collect --config champion --env arena --data-dir data/eval \
  --num-episodes 256 --episode-len 32 --image-size 64 --seed 100
uv run visionary-mlx dream --config champion --run-dir releases/v0.2.0 \
  --data-dir data/eval --env arena --out dreams
uv run visionary-mlx benchmark --config champion --run-dir releases/v0.2.0 \
  --eval-data-dir data/eval --env arena --name visionary-mlx-champion