AlayaRenderer-Flash

Generative World Renderer at the Speed of Play

AlayaRenderer-Flash is the real-time version of AlayaRenderer. The game engine keeps simulating the world and exports G-buffers (albedo, depth, metallic, normal, roughness). AlayaRenderer-Flash renders them into RGB frames in a style you set with a text prompt, at playback rate. This repository has the inference code, the released SuperTuxKart model, and the playable demo behind the live site.

Features

• Streams indefinitely. An autoregressive renderer with a compressed long/mid/short latent history.
• Renders in 4 steps. Distilled from a 50-step teacher, with classifier-free guidance folded into the model.
• Uses lightweight codecs. A grouped G-buffer encoder and a stateful tiny decoder replace the Wan VAE.
• Runs in real time. The model renders at about 31 FPS on one H200, and the full game loop holds 30 FPS.
• Plays live. SuperTuxKart integration with a web UI, prompt-driven style switching, and keyboard control.

Installation

You need Linux, an NVIDIA GPU with driver 570 or newer (CUDA 12.8), and uv.

git clone https://github.com/AlayaLab/AlayaRenderer-Flash.git
cd AlayaRenderer-Flash
uv sync                       # Python 3.12 + PyTorch 2.11 (CUDA 12.8) + locked dependencies
uv sync --extra demo          # also install OpenCV for the live SuperTuxKart demo

To use it as a library in an existing environment: uv pip install "git+https://github.com/AlayaLab/AlayaRenderer-Flash.git".

Quickstart

uv run python examples/fetch_samples.py      # two recorded SuperTuxKart G-buffer windows
uv run alayarender-flash render samples/window-00.npy samples/window-01.npy \
    --prompt "heavy snow at night, warm street lights, cinematic" -o snow.mp4

The weights download from the Hugging Face Hub on first use. To use local copies instead, pass --checkpoint / --base-model or set ALAYA_CHECKPOINT / ALAYA_BASE_MODEL. On GPUs with less than about 24 GB, add --text-encoder-cpu.

From Python:

from alayarenderer_flash.render import OfflineRenderer, load_gbuffer_stream
from alayarenderer_flash.runtime import RuntimeConfig
from alayarenderer_flash.weights import resolve_base_model, resolve_checkpoint

config = RuntimeConfig(base_model=resolve_base_model(), checkpoint=resolve_checkpoint())
renderer = OfflineRenderer(config)
frames = renderer.render(load_gbuffer_stream(["samples/window-00.npy"]), "golden hour, warm haze")

G-buffer inputs are uint8 arrays of shape (T, 5, 448, 832, 3) in the order albedo, depth, metallic, normal, roughness. They are rendered causally: 13 frames for the first chunk, then 16 per chunk.

Live demo

Play SuperTuxKart through the renderer in your browser:

cp examples/deploy.env.example deploy.env          # set the STK binary and weight paths
uv run alayarender-flash serve --config deploy.env doctor --json
uv run alayarender-flash serve --config deploy.env start

See docs/demo.md for the full setup. The demo needs a SuperTuxKart build with G-buffer export, which is not distributed here; see integrations/supertuxkart. docs/architecture.md describes how the pieces fit together.

Model zoo

Model	Domain	Steps	Resolution	Weights
AlayaRenderer-Flash (checkpoint-600)	SuperTuxKart	4	832×448	🤗 AlayaLab/AlayaRenderer-Flash

It needs the base model Wan-AI/Wan2.1-T2V-1.3B-Diffusers for the text encoder, tokenizer and VAE configuration.

Results

Evaluation set from the technical report: 131 clips × 150 frames at 832×448 on one H200.

Method	Steps	SCLIP-I ↑	Boundary SSIM ↑	tLPIPSwarp ↓	FPS ↑	VRAM (GB) ↓
AlayaRenderer	50	0.836	0.308	0.124	0.56	30.1
AlayaRenderer-Flash	4	0.847	0.430	0.155	31.54	16.2

Development

uv run pre-commit install
uv run ruff format . && uv run ruff check .
uv run pytest -q

See CONTRIBUTING.md.

Citation

@article{lin2026alayarendererflash,
  title   = {Generative World Renderer at the Speed of Play},
  author  = {Lin, Guixu and Huang, Zheng-Hui and Yang, Siqi and Yang, Ming-Hsuan and Zhang, Kaipeng and Wang, Zhixiang},
  journal = {arXiv preprint arXiv:2607.18703},
  year    = {2026}
}

License

The code is licensed under the Apache License 2.0; attributions are in NOTICE, and third-party code, dependencies and media are listed in THIRD_PARTY_LICENSES.md. Exceptions:

• SuperTuxKart (GPL-3.0-or-later) is not part of this repository: its source, binaries, modifications and game assets are not distributed here, nor are images rendered from SuperTuxKart scenes.
• src/alayarenderer_flash/decoders/taehv.py is MIT-licensed (TAEHV); see LICENSES/TAEHV-MIT.txt.
• The model weights are released under the license in the model repository; downloads are pinned to a fixed revision in src/alayarenderer_flash/weights.py.

Acknowledgements

This project builds on Wan 2.1, Helios, TAEHV, Diffusers, and SuperTuxKart.