3,221 clipsGEAR games
38 games built end-to-end by our AI game-creation tool, each exported with five aligned render passes, from blockout to final.
ALAYALAB
AnyGameRenderer
Entity-Consistent Generative Rendering Across Games and an Entity-Level Benchmark
Agents and vibe-coding tools can already build a playable game — the layout, logic, motion and camera of every scene. AnyGameRenderer turns that crude recording into the finished game, and one checkpoint covers all three axes of diversity: many game genres — shooters, racers, sports, fantasy and more; many input forms — blockouts, colored primitives, wireframes, semantic maps, coarse mockups; and many output styles — from Ghibli and LEGO to cyberpunk and watercolor — with every entity’s position, timing and count locked to the input.
One model renders shooters, racers, sports and fantasy alike. Drag the seam: left of the line is the crude input the model sees, right is what it returns.
The input does not have to be one fixed format. Wireframes, clay mockups, 2D semantic maps, bare primitives, engine whiteboxes — the model reads them all.
The same crude clip can be rendered into whatever world you ask for — whether it arrives as a blockout or a semantic map. One input, its rendered styles beside it, all playing together. Click through the samples.
Paired (crude → styled) clips from three sources, frame-aligned by construction.
38 games built end-to-end by our AI game-creation tool, each exported with five aligned render passes, from blockout to final.
Same simulation rendered twice — blockout and full — with scenarios scripted against zero-shot failure cases.
Real gameplay with detected 3D boxes re-rendered as crude proxy shapes; the original recording is the target.
TOTAL 10,330 paired clips · 124 frames @ 24 fps, 1344×768 · ≈ 14.8 hours of paired video
Frame-level metrics cannot tell whether each entity is rendered correctly, so we score entity correspondence directly: 59 held-out cases, 753 entities annotated with ground-truth boxes, same seed / steps / captions as the zero-shot base. The GT floor runs the real target footage through the same measurement pipeline.
| Metric | Zero-shot | Ours | GT floor |
|---|---|---|---|
| Temporal match ↑ | 50.9% | 62.9% | 80.9% |
| Position match ↑ | 57.1% | 61.9% | 79.3% |
| Unrendered input blocks ↓ | 4.1% | 2.9% | 0.1% |
| Hallucinated entities ↓ | 52.1% | 46.0% | 34.1% |
| Missed count changes ↓ | 34.7% | 21.3% | 0.0% |
A single checkpoint improves temporal match in every genre of the benchmark — first-person shooter, top-down, third-person, crowd and vehicle games — while styles stay fully promptable. Full protocol, per-source splits and qualitative comparisons are in the technical report.