RenderFormer-V2: Neural Rendering with Heterogeneous Scene Primitives
RenderFormer-V2 is built to render mixed scene elements and difficult light transport without per-scene training.
The paper describes a transformer-based neural renderer that treats global light transport as a sequence-to-sequence problem. It extends the earlier RenderFormer with windowed attention and a rendering-informed attention sink for better scalability while preserving accuracy. The model supports environment maps, participating media, textured and displaced surfaces, caustics, volumetric scattering, and out-of-distribution materials. The authors say it was accepted to ECCV 2026 and include ablations of the new attention mechanism. HF Daily Papers' note
The paper describes a transformer-based neural renderer that treats global light transport as a sequence-to-sequence problem. It extends the earlier RenderFormer with windowed attention and a rendering-informed attention sink for better scalability while preserving accuracy. The model supports environment maps, participating media, textured and displaced surfaces, caustics, volumetric scattering, and out-of-distribution materials. The authors say it was accepted to ECCV 2026 and include ablations of the new attention mechanism. HF Daily Papers' note
score 5