AnyGameRenderer: Entity-Consistent Generative Rendering Across Games and an Entity-Level Benchmark

AnyGameRenderer takes a crude gameplay recording (the "programmer art" agents produce: gray boxes and colored cylinders on flat ground) and a text prompt, and outputs the same gameplay re-rendered in the prompted style, frame-aligned with the input.

This repository holds the LoRA (anygamerenderer_lora.safetensors, 2.4 GB) that turns MiniMax-H3 Ref2VA into AnyGameRenderer. Paper: https://arxiv.org/pdf/2610.31466 · Project page: https://alaya-lab.github.io/AnyGameRenderer/ · Inference code, the caption builder, and examples: https://github.com/AlayaLab/AnyGameRenderer.

• 🎮 Generalizes across genres: one checkpoint renders tower defense, auto-battlers, top-down survivors, turn-based battles, kart racing, crowd runners, arena survival, co-op games, and shooters.
• 🎯 Keeps entity correspondence: group counts stay correct, entities appear and disappear on the right frame, and small far-away entities are not lost.
• 📹 Its input is plain video: gray blockouts, colored primitives, semantic maps, wireframes, and coarse mockups all work, with no per-object labels or engine hooks.
• 🎨 Any prompted style: photoreal AAA, LEGO, Ghibli, crayon, synthwave, samurai, candy land … with the characters re-imagined to match.

Files

file	size	what
anygamerenderer_lora.safetensors	2.4 GB	LoRA on the MiniMax-H3 Ref2VA transformer: rank 128, alpha 128, on attn.qkv_proj, attn.out_proj, mlp.fc1, mlp.fc2 of every block (musubi-tuner networks.lora_minimax_h3 format).

The base weights (Ref2VA transformer, Qwen3-VL-32B text encoder, video and audio VAEs, ~123 GB) are not included; download them from Comfy-Org/MiniMax-H3 after accepting the MiniMax-H3 Community License.

Usage

git clone https://github.com/AlayaLab/AnyGameRenderer.git && cd AnyGameRenderer
bash setup_env.sh && conda activate agr
bash download_weights.sh                                   # base weights -> weights/
huggingface-cli download AlayaLab/AnyGameRenderer anygamerenderer_lora.safetensors --local-dir weights
bash run_example.sh                                        # renders the bundled examples (~17 min per 5 s clip on an 80 GB GPU)

Inputs are 1344×768, 24 fps clips of 2–15 s; captions are built from a short spec.json with make_caption.py (see the repo README and the anygamerenderer-caption skill for how to write one).

Citation

@article{liu2026anygamerenderer,
  title={AnyGameRenderer: Entity-Consistent Generative Rendering Across Games and an Entity-Level Benchmark},
  author={Liu, Ruicong and Ouyang, Liangyang and Lan, Fengbo and Wang, Zhixiang and Zhang, Kaipeng},
  journal={arXiv preprint arXiv:2610.31466},
  year={2026}
}

License

The LoRA is a derivative of the MiniMax-H3 weights and is distributed under the MiniMax-H3 Community License.