README.md

metadata

title: WorldGuide
emoji: 🧭
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
python_version: '3.10'
short_description: Goal-directed video world model, step by step
startup_duration_timeout: 1h

WorldGuide

Interactive closed-loop demo of WorldGuide: Goal-Directed Video World Model for Procedural Task Execution (code, weights).

Each press of the button runs one closed-loop step: the ContextPlanner (fine-tuned Qwen2.5-VL-7B) predicts the next atomic action from the latest observation, and the WorldGuide video DiT (HunyuanVideo-1.5 480p I2V based, mode-2 step memory over the last 2 steps) renders it as a 45-frame clip. Settings follow the authors' run_v5_qwen_closed_loop_memory.sh (45 frames, CFG 7.5, shift 5, seed 3208, step-1 memory from a 1-second environment preview). The default is 20 denoising steps instead of the authors' 30 to keep each ZeroGPU call around 3 minutes; set 30 under Advanced settings.

Example image credits

• examples/seafood_pancake_ingredients.jpg — "Korean pancake-Ingredients for Haemul pajeon" by draq (Flickr), via Wikimedia Commons, CC BY 2.0 (cropped/resized).
• examples/lego_bricks_desk.jpg — "Lego messy desk" by Pascal from Heidelberg, Germany, via Wikimedia Commons, CC BY 2.0 (cropped/resized).
• examples/strawberries_cutting_board.jpg — from linoyts/repo-to-space-example-inputs (CC0).