AnyGameRenderer brings entity-consistent rendering to simple game footage
Ruicong Liu and four coauthors present AnyGameRenderer, an arXiv paper describing a video-to-video generative renderer that upgrades simple game footage with characters, motion detail and effects while preserving layout, camera movement, entity positions and timing. The system accepts plain video without engine buffers, per-object class labels or low-poly proxies. It is implemented through LoRA fine-tuning of the 33B-parameter MiniMax-H3 ref2va model and trained on about 10,000 paired clips from agent-made games exported through GEAR, Unreal Engine and real gameplay recordings.
The authors also introduce an entity-level benchmark containing 59 held-out cases and 753 entities, scored against ground-truth boxes. They report that, compared with the base model, AnyGameRenderer raises temporal matching from 50.9% to 62.9% and reduces missed count changes from 34.7% to 21.3%, providing a way to assess entity correspondence in generative game rendering.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.31466 (cs)
[Submitted on 22 Oct 2026]
Title:AnyGameRenderer: Entity-Consistent Generative Rendering Across Games and an Entity-Level Benchmark
Authors:Ruicong Liu, Liangyang Ouyang, Fengbo Lan, Zhixiang Wang, Kaipeng Zhang
View a PDF of the paper titled AnyGameRenderer: Entity-Consistent Generative Rendering Across Games and an Entity-Level Benchmark, by Ruicong Liu and 4 other authors