Evaluating 3D asset generation by feeding renders as images to a small LLM judge

mervenoyann · x · 2026-09-07

Hugging Face's Merve shares a low-compute trick for evaluating 3D generation quality: instead of evaluating glb files directly, pass screenshots of renders as images to a smaller judge model (e.g., Gemma-4 or a shrunken Qwen) — a 9B model in bf16 needs only 17.9GB VRAM.

If compute is limited, you can restrict evaluation to single 3D assets rather than complex scenes and train a smaller Qwen to score them. The "render-then-judge" approach is directly reusable for indie 3D generative model work.

Original post →

Evaluating 3D asset generation by feeding renders as images to a small LLM judge

More from Multimodal