AnyGameRenderer Dresses Gray AI Prototypes in Real Graphics and Keeps 62.9% of Units Intact

Coding agents have learned to assemble small playable games in minutes, but the results look like "programmer art": gray boxes, colored geometric shapes, flat semantic blobs. The researchers behind AnyGameRenderer propose closing that gap with a video model that takes such a recording and redraws it as a full picture with characters, motion detail, and effects. Their key metric is entity preservation: temporal match rose from 50.9% to 62.9% relative to the base model, and the share of missed object-count changes fell from 34.7% to 21.3%.

Video Instead of a G-buffer

Most approaches to neural game rendering demand a lot of auxiliary data from the engine: G-buffers, per-object class labels, low-poly proxies. A game that an agent assembled in half an hour has no such luxury. AnyGameRenderer works with ordinary video. Gray blockouts, colored primitives, semantic maps, and wireframe meshes all work equally well. The model preserves the layout, camera, position, and timing of each entity, and fills in the style, characters, and effects on its own.

The claimed genre coverage is broad: tower defense, auto-battlers, top-down survival, turn-based combat, kart racing, and crowd runners. For these games, counting discipline matters most: if a wave had 12 enemies, the output must show 12, not 9 or 15.

Where Video Models Usually Lose Units

Generative video models draw well but count poorly. In a crowd, they merge neighboring characters, "forget" those who appear later, or add extras. That is tolerable for film but not for a game interface: players track a specific unit, and an object appearing or disappearing on the wrong frame breaks gameplay.

So the authors built an entity-level benchmark: 59 held-out cases, 753 entities, scored against ground-truth bounding boxes. The model is compared with the baseline on whether objects match over time, whether changes in count are caught, and how many entities are invented or left unrendered. AnyGameRenderer wins on all of these measures.

What the Model Was Fed

The foundation is MiniMax-H3 ref2va, a 33-billion-parameter video model that accepts a reference video as a condition. It was fine-tuned with LoRA on the sim10k corpus, roughly 10 thousand paired clips from three sources:

• multi-stream rendering data from 38 agent-built games exported from GEAR, an AI tool for game creation;
• two-stream data from Unreal Engine with programmatically controlled crowds, changes in unit count, and entity scales;
• real gameplay recordings, for which pairs were reconstructed through 3D-box detection.

The first source supplies the visual junk agents typically produce as input, the second allows precise control over counts and object appearance, and the third adds realism. It is a sensible mix: synthetic data provides exact labels, while real recordings guard against an overly sterile distribution.

The Numbers and Their Ceiling

• Temporal entity match: 50.9% → 62.9%.
• Missed object-count changes: 34.7% → 21.3%.
• Benchmark: 59 held-out cases, 753 entities.
• Base: MiniMax-H3 ref2va, 33B, LoRA fine-tuning.
• Data: about 10 thousand paired clips from three sources.

The improvement is noticeable, but 62.9% means more than a third of entities still fail to match the ground truth over time. That is not enough for production. For quick prototypes and pitches it is enough: you get visuals you can show to an investor or to players in an early playtest, without an artist.

The practical takeaway for the industry: the chain of "an agent writes the game, a video model draws it" is becoming plausible. Indie studios and prototype authors can use gray builds as a script for visualization. Developers of tools like GEAR should build clean clips with clear entity semantics into their export, and those building their own renderers should move from generic image-quality metrics to object-preservation metrics: that is where the line now runs between a pretty video and graphics that are usable in a game.