Alaya Lab's InfiniDream extends Infinigen via Claude Opus 4.8

Alaya Lab, with National Yang Ming Chiao Tung University and National Taiwan University, posted InfiniDream to Hugging Face on October 20. The framework wraps Claude Opus 4.8 around Infinigen, the procedural scene engine built on "over 40,000 lines of code," and uses the VLM to extend its asset factories rather than only call them.

The paper reports that "InfiniDream achieves the highest scores across all four perceptual metrics at both the object and scene levels." Evaluation covers "100 object-level tasks across six categories and 20 scene-level tasks at resolutions of 1024×1024 and 960×540, respectively." On the scene benchmark, InfiniDream posts a CLIP score of 29.64 and DINOv3 of 0.361, versus 28.85 and 0.321 for the strongest baseline, VLM (Harness), and 25.21 and 0.182 for stock Infinigen. Other baselines include VIGA, BlenderRAG and Code2Worlds.

The human study is small. Thirty participants compared results on 10 object and 10 scene tasks. Within that pool, InfiniDream was picked 77.5% of the time overall at the object level and 83.2% at the scene level, with scene semantic alignment reaching 91.0%.

The core trick is called Symbolically Constrained Editing. The authors describe the resulting Factory Brief as containing "a Code Structure Map identifying key functions and dependencies, an Editable Surface consisting of asset-specific modules dynamically identified through depth-limited dependency analysis, and Linked Infrastructure providing read-only access to shared low-level algorithms (e.g., node graphs) via symbolic links." A separate Deterministic Frustum-Guided Grid Placement module handles scene layout, and the scene-level DINOv3 reference images are generated by Gemini 3 Pro Image, shared across all methods and never shown to any system during generation.

Only Claude Opus 4.8 is tested as the backbone; a qualitative comparison with GPT-6 Astra appears in the appendix as "supplementary evidence beyond the main benchmark suite." This is the 44th computer-vision item we've tracked in the past 90 days, following yesterday's note on ZJU's Scene-Designer in the same agentic 3D scene generation lane.