AnyGameRenderer Turns Agent-Made Blockouts Into Final Visuals

A video-to-video renderer called AnyGameRenderer takes the gray boxes and colored cylinders that coding agents produce for playable games and repaints them, frame by frame, into finished game visuals with characters, motion detail and effects. The authors — Liangyang Ouyang, Fengbo Lan, Zhixiang Wang and Kaipeng Zhang — posted the paper on 22 October, pitching it as the rendering half of the engine-plus-generator split that CWM, PWM, Matrix-Game 4.0, StreamRender-H3 and SolarRenderer have all taken up in recent months.

The system is a rank-128 LoRA on MiniMax-H3 ref2va, a 33B reference-video-conditioned video model, trained with flow matching on 10,330 paired clips of 124 frames at 24 fps and 1344×768. The data mix is three roughly equal parts: 3,221 clips from 38 agent-made games exported via the GEAR game-creation tool, 3,910 Unreal Engine clips scripted around correspondence failures (count changes in same-color groups, entity sizes from 20 to 600 pixels, mixed camera views), and 3,199 real gameplay recordings turned into crude inputs by 3D-box detection. Training ran three epochs in about 48 hours on 16 NVIDIA H200 GPUs.

The headline claim is on an entity-level benchmark the authors introduce — 59 held-out cases with 753 annotated entities, scored with SAM3 open-vocabulary segmentation and Hungarian matching. "temporal match rises from 50.9% to 62.9%, missed count changes fall from 34.7% to 21.3%," they report, with position match moving from 57.1% to 61.9% and unrendered rate from 4.1% to 2.9%. The hallucination rate — entities in the output that are not in the input — only drops from 52.1% to 46.0%, and on the real-gameplay split temporal match improves by less than five points.

The paper lands in a run of video-world-model research we've been tracking all week, including Nvidia's JCA and HKU's MIST, part of 45 video-generation stories in the last 90 days. The authors explicitly do not argue for the engine-plus-renderer split; they say they "build a rendering layer that is more general in what it accepts and more faithful in what it keeps."