kerzgrr/diffusionminer

Diffusion Miner

An action-conditioned world model that learned a falling-sand mining sandbox. You hold A / D to walk, W or Space to jump, and the arrows to dig; it invents the next frame.

256×256 at 12 FPS. The attention prefix is 12 latent frames; a ConvGRU world memory scans 48. About 220M parameters. The codec is a frozen Tile VAE trained on this game, not SDXL. Needs a CUDA GPU with BF16.

This release is the teacher-stage instantaneous flow checkpoint. One step is the right setting — there is no distilled student.

Play it

pip install torch numpy pillow safetensors huggingface_hub

hf download kerzgrr/diffusionminer --local-dir checkpoints/diffusionminer
python checkpoints/diffusionminer/live_infer.py \
  --local-dir checkpoints/diffusionminer \
  --steps 1

If the weights are already in the current directory:

python live_infer.py --steps 1 --window-scale 3 --seed 2026

Click the window once so it has focus. Esc quits. Clicking the canvas drops a miner into the latent history if you want to poke it.

Write a short gameplay GIF with the scripted training miner:

python live_infer.py \
  --local-dir checkpoints/diffusionminer \
  --steps 1 \
  --seed-play \
  --autoplay \
  --frames 144 \
  --output diffusionminer.gif

Controls

key	action
A	move left
D	move right
W / Space	jump
← ↑ → ↓	dig that way

Action vector is [left, right, jump, dig_left, dig_right, dig_up, dig_down] as floats in {0,1}. Opposite move keys cancel. Only one dig direction is active at a time.

What's in the files

file	what
ema.safetensors	playable teacher weights
codec.pt	frozen Tile VAE
config.json	video / model / codec settings used at train time
livediffusion/	inference package used by live_infer.py
live_infer.py	live play launcher
demo.gif	scripted gameplay clip

If something breaks, it is probably a GPU that cannot do BF16, or running live_infer.py without the rest of this snapshot on sys.path.