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

AnyGameRenderer. Entity-Consistent Generative Rendering Across Games and an Entity-Level Benchmark. By Ruicong Liu and four others, from Alaya Lab. Published on the twenty-second of October. Its code repository is public and real: about thirty source files.

AI coding agents can now build small playable games in minutes. What they produce, though, usually looks like programmer art: gray blockouts, colored geometric primitives, flat semantic blobs. The game works, and it looks like a placeholder. Dressing it up is not just a matter of style, either, because a game's visuals carry information. If a renderer moves an enemy, drops one from a crowd, or makes one appear a frame late, the player is now looking at a different game.

AnyGameRenderer is a video-to-video generative renderer that turns such footage into high-quality visuals, with characters, motion detail and effects. Throughout, it keeps the layout, the camera, and every entity's position and timing intact. That last property is the one the authors stress: entity correspondence. Group counts stay correct, and entities appear and disappear on the right frame.

The input is deliberately plain. The renderer takes ordinary video as its game state, with no engine G-buffers, no per-object class labels, no low-poly proxies. Gray blockouts, colored primitives, semantic maps and wireframes all work. Whatever the agent happened to draw the game with, the renderer reads it as the state of play. It is meant to generalize across the genres agents build: tower defense, auto-battlers, top-down survivors, turn-based battles, kart racing and crowd runners. That is a wide spread, from racing karts to whole crowds of runners, and the same renderer is asked to handle all of it.

The system is a LoRA fine-tune of MiniMax-H3 ref2va, a thirty-three billion parameter video generation model conditioned on reference video. The training pairs come from three sources. The first is multi-stream rendering data from thirty-eight agent-made games, exported from GEAR, an AI game-creation tool.

The second is two-stream rendering data from Unreal Engine, where crowds, count changes and entity scales are controlled programmatically. The third is real gameplay recordings, with the pairs reconstructed through 3D-box detection. Between them, the sources cover games that agents actually made, controlled cases where the number and size of entities can be varied on purpose, and footage from games people really play. Each pair links a plain version of a clip to a rendered one, which is what the fine-tune learns to map between.

To test correspondence directly, the authors propose an entity-level benchmark: fifty-nine held-out cases containing seven hundred and fifty-three entities, scored against ground-truth boxes. Compared with the base model, AnyGameRenderer raises temporal match from fifty point nine percent to sixty-two point nine percent.

Missed count changes fall from thirty-four point seven percent to twenty-one point three percent. Fewer entities are hallucinated, and fewer are left unrendered. Each of those measures asks the same question from a different side: is every thing in the original footage still there, once, at the right moment?

The work carries two contributions under one name. One is a renderer that takes the plainest possible input and produces finished-looking game footage. The other is a benchmark that refuses to grade on looks alone, and instead counts entities and checks their timing, because for a game, getting the picture right means getting the state right.