hugging-apps/abot-world-0-5b-lf

Interactive demo of ABot-World-0 (5B, Long-forcing) — a real-time, action-conditioned world model fine-tuned from Wan2.2-TI2V-5B (paper · code).

Pick a starting frame, press Start world, and steer the rollout live:

• —W A S D — move & turn
• —I J K L — pan the camera

The model rolls the world out block-by-block conditioned on your held keys, and frames stream to the browser over a WebSocket at real-time pace. Each session gets ~150 s of ZeroGPU time (~2 minutes of exploration).

How it works

• —Engine: one pipe.stream(actions=<callable>) loop over the Modular Diffusers `ABotWorldStreamingBlocks` preset — the pipeline polls the held keys once per generated block and yields every block's decoded frames.
• —Serving: gradio.Server (FastAPI-native) — /start_game / /stop_game API endpoints, a WebSocket /ws streaming binary JPEG frames, and /upload_seed + /example_seed endpoints for the i2v seed frame.
• —Weights: `YiYiXu/ABot-World-0-5B-LF-Diffusers` (diffusers mirror of acvlab/ABot-World-0-5B-LF).
• —Preset starting-world images come from the Apache-2.0 ABot-World repo.