Inference

July 25, 2026 · View on GitHub

Inference

Minimal local inference for the Open Dreamer world model — roll out new frames from an MP4 and matching Minecraft VPT actions.

🎮 Live Demo · 🌐 Project & Training

_{REAL-TIME DEMO POWERED BY}

This repository is a small, self-contained rollout harness for Open Dreamer. Given a short video clip and a sequence of Minecraft/VPT-style actions, it encodes the clip into the model's latent space and generates the frames that follow — a local, scriptable way to drive the world model on your own footage.

Want to just play with the model instead? The live demo runs the real-time version in your browser, no setup required.

🎮 Try it now

The easiest way to experience Open Dreamer is the in-browser demo — step into a generated Minecraft world and play it in real time:

👉 Open the demo

Read on if you want to run rollouts locally.

⚙️ Quick Start

Run every command below from the repository root.

1. Install

uv sync

2. Provide a checkpoint

You need a trained Open Dreamer checkpoint (an Orbax checkpoint directory). Train one with the Open Dreamer training pipeline, or point --checkpoint_path at a checkpoint you already have. The examples below use /path/to/open-dreamer-checkpoint.

3. Verify JAX sees the GPU

Rollouts require a CUDA-visible JAX GPU; the script exits before loading the checkpoint if jax.devices("gpu") is empty.

uv run python - <<'PY'
import jax
import jax.numpy as jnp
print("backend", jax.default_backend())
print("gpu devices", jax.devices("gpu"))
x = jnp.ones((2048, 2048), dtype=jnp.float32)
y = (x @ x).block_until_ready()
print("result device", y.device)
PY

4. Download the sample data

download_vpt_sample.py fetches OpenAI's official Minecraft VPT 10.x contractor index, verifies the sample .mp4 is listed, and downloads the MP4 together with its paired .jsonl action file:

uv run python download_vpt_sample.py --overwrite

Expected files:

samples/vpt/cheeky-cornflower-setter-02e496ce4abb-20220421-092639.mp4
samples/vpt/cheeky-cornflower-setter-02e496ce4abb-20220421-092639.jsonl

5. Run a rollout

XLA_PYTHON_CLIENT_PREALLOCATE=false uv run python inference.py \
  --checkpoint_path /path/to/open-dreamer-checkpoint \
  --input_mp4 samples/vpt/cheeky-cornflower-setter-02e496ce4abb-20220421-092639.mp4 \
  --actions_path samples/vpt/cheeky-cornflower-setter-02e496ce4abb-20220421-092639.jsonl \
  --output_mp4 outputs/vpt_sample_rollout.mp4 \
  --num_context_frames 4 \
  --horizon 1 \
  --num_steps 4 \
  --use_ema

Larger rollouts

Increase --num_context_frames and --horizon for a longer output:

XLA_PYTHON_CLIENT_PREALLOCATE=false uv run python inference.py \
  --checkpoint_path /path/to/open-dreamer-checkpoint \
  --input_mp4 samples/vpt/cheeky-cornflower-setter-02e496ce4abb-20220421-092639.mp4 \
  --actions_path samples/vpt/cheeky-cornflower-setter-02e496ce4abb-20220421-092639.jsonl \
  --output_mp4 outputs/rollout.mp4 \
  --num_context_frames 16 \
  --horizon 64 \
  --num_steps 4 \
  --use_ema

Only the first --num_context_frames video frames are read and encoded. The action file must contain at least num_context_frames + horizon actions; the full action sequence is shifted before it is split into context and future actions.

Input format

The action file may be a JSON array or JSONL. Each entry is a VPT-style action dictionary with mouse and keyboard fields, for example:

{"mouse":{"dx":0.0,"dy":0.0,"buttons":[],"dwheel":0.0},"keyboard":{"keys":["key.keyboard.w"]}}

Input video frames must be RGB 368x640, or RGB 360x640 so they can be zero-padded to the trained 368x640 model shape.

📚 References

Open Dreamer: project page, blog post, and training pipeline
Dreamer 4: Training Agents Inside of Scalable World Models