Automated judges for 3D generation, whether image-scoring models or multimodal large language models (MLLMs, language models that also read images), look at a 3D asset through a handful of rendered pictures. Those pictures are almost always taken from cameras placed on a fixed sphere around the object, with no regard for where the surface actually faces. A patch of texture seen edge-on occupies only a few blurry, squashed pixels. The judge then cannot tell whether the asset's texture is poor or whether it was simply shown badly. This proposal treats evaluation-view selection as a texture-precision problem, borrowing a 25-year-old graphics idea. We choose a small set of cameras so that every visible surface point is seen at least once with a projected pixel no larger than the asset's own texel. We then test whether MLLM judges become more accurate and less confused when shown these views instead of uniform ones, at the same number of judge calls.
The central claim is narrow. At a fixed budget of $K$ rendered views, a precision-covering view set will let an MLLM judge detect localized texture defects more often and agree with humans more closely than uniform spherical views. It will also sharply cut one specific error, which we call attribution error: penalizing a clean asset for blur that comes only from the viewpoint. We will build a controlled 3D benchmark that separates real texture defects from viewpoint-induced degradation, so that this claim can be checked rather than asserted.
What exists and where the gap is
Classical texture acquisition already knows how to pick good views. On the Precision of Textures defines a texture as having precision $\epsilon$ when every surface point is covered by a projected pixel no larger than $\epsilon$, and casts viewpoint choice as set cover. In that paper, placing cameras along simplified triangle normals reached 100% coverage on a 1,634-triangle scene, against 97% for spherical sampling. Reconstruction and texturing systems use the same principle in weaker form. Differentiable Stereopsis weights views by viewing angle when compositing color. Mesh-based 3D Textured Urban Mapping shows that equal weighting of views causes stretched textures. TEXTure marks regions painted from oblique angles for later refinement. Pixel-GS shows that averaging gradients across views, rather than weighting them by pixel coverage, lets weak views dilute strong ones. These papers supply the method: all of them use angle and footprint to decide which pixels to trust when building a texture. None of them uses these quantities to decide which pixels a judge should see.
On the judging side, T3Bench is the closest prior work. It scores 161 views on a level-2 icosahedron with ImageReward or CLIP and smooths the scores over neighboring views. For text alignment, it captions 12 fixed views with BLIP and has GPT-4 merge the captions. It reports Spearman correlation above 0.75 with human experts. Its cameras ignore the mesh geometry entirely, and its language model never sees pixels. Two further papers motivate the proposal. MLLM-as-a-Judge finds that MLLM judges crowd scores into 4–5, fail to penalize blur, and are most reliable in pairwise comparison, where GPT-4V reaches about 79% agreement with humans. Mitigating Perceptual Judgment Bias shows that judges accept fluent text that contradicts the image, raises perceptual-error detection from 69.5% to 85.7% with targeted training, and states that its judge remains weak at 3D viewpoint reasoning. That last admission also constrains our design. Instead of asking a judge to reason about foreshortening, which it does poorly, we remove foreshortening from its input through geometry.
Core mechanism
For a surface point $p$ and camera $v$ at distance $d_v(p)$ with focal length $f$, pixel pitch $s$, and angle $\theta_v(p)$ between the viewing ray and the surface normal, the projected pixel footprint along the most stretched direction is approximately
$$\epsilon_v(p) \approx \frac{s\, d_v(p)}{f \cos\theta_v(p)},$$
following the precision analysis of Nielsen and de Mauroy. Let $\tau$ be the asset's texel size on the surface. For a neural or Gaussian asset converted to a mesh as T3Bench does, $\tau$ is taken from the extracted texture. We choose the view set $V$, with $|V| = K$, by greedily maximizing precision-covered area:
$$\max_{V} \; \mathrm{Area}\{\, p : \min_{v \in V,\ p \text{ visible in } v} \epsilon_v(p) \le \tau \,\}.$$
Candidate cameras come from clustered face normals, as in the Simplified Triangle Normal heuristic, plus the canonical views needed for semantic context. Greedy selection gives the usual $(1-1/e)$ guarantee for this coverage objective. Each selected view is also rendered with a per-pixel footprint map. In one variant, pixels with $\epsilon_v > \tau$ are dimmed so the judge knows not to trust them. The judge itself is unchanged. It receives the $K$ images and a rubric, and it gives pairwise preferences, the most reliable protocol according to Chen et al., along with a list of localized defects.
Research Program
The first component is a controlled benchmark built from Google Scanned Objects, whose 1,030 real textured meshes, many with printed product text, give us ground-truth textures. For each object we create defective variants by editing the texture in small regions. The edits include local blur, a misspelled label, a shifted seam, a duplicated face-like patch that mimics the Janus artifact (a generated object showing the same face on several sides), and a color patch swap. Each defect is placed on surfaces of varying orientation. For each clean object we also create "viewpoint-degraded" presentations: the same flawless asset, shown through cameras that see key regions only at grazing angles. The pairs are clean versus defective (can the judge find the flaw?) and clean-from-bad-views versus defective-from-good-views (does the judge blame the viewpoint on the asset?). This is the 3D counterpart of the perturbation design in PPJD, with orientation as the controlled variable.
The second component is a human-anchored study on real generated assets. We take the 300 T3Bench prompts, generate assets with several text-to-3D systems (including MVDream and ProlificDreamer, which T3Bench identifies as the leaders), and collect human pairwise preferences in a free-orbit 3D viewer, so humans are not limited to any fixed view set. Judges are one proprietary MLLM and one open 7B-class MLLM (Qwen2.5-VL-7B), both run under each view policy.
Baselines
The naive reference point is the T3Bench alignment pipeline as published: 12 canonical views, BLIP captions, and a GPT-4 merge-and-recall judgment. It shows what happens when the judge never sees pixels. Its reported figure is Spearman above 0.75 with human experts (T3Bench). That figure was measured against T3Bench's own human ratings rather than ours, so we re-run the pipeline on our human preferences.
The strongest existing automated metric is T3Bench's quality score: 161 icosahedron views scored by ImageReward with regional graph smoothing. We run it on exactly our assets and human labels. It is the strongest competitor because it already sees many views and was validated against humans. To beat it, our method must achieve higher pairwise agreement with human preferences on the T3Bench-generated set. It must also clearly outperform this metric on localized defect detection in the GSO benchmark, where global image scorers are expected to be weak. For MLLM judging in general, the reference is GPT-4V's roughly 79% pairwise human agreement on 2D tasks (MLLM-as-a-Judge). No reported result on 3D assets was found on Lacuna for MLLM pairwise judging.
The direct competitor for the key ingredient is the same MLLM judge, with the same prompt, the same $K$, and the same render resolution, given uniform views: $K$ points spread evenly on the sphere, and separately $K$ random views. Every condition gets the same number of judge calls, the same image tokens, and the same prompt-tuning effort. The prompt is tuned once on a held-out split under the uniform policy, which favors the baseline. Beating it means a statistically significant gain in defect detection and human agreement at each $K$ from 4 to 16, together with lower attribution error.
Two ablations isolate the mechanism. The first is "distance-only": uniform view directions, but each camera zoomed so its footprint matches $\tau$ at the fronto-parallel limit. It corrects the $d$ term in the footprint equation but not the $\cos\theta$ term. The second compares precision-covering views with and without the footprint-map annotation, which tests whether telling the judge which pixels to trust helps beyond choosing better cameras.
Why not just…?
Why not just render many uniform views and let the MLLM judge all of them? This is the real threat. With 161 views, some camera will view any region nearly head-on, so precision coverage may be achieved by brute force. We will run this dense condition, taking the minimum score across views (flag a defect if any view shows it). If it matches precision-covering views, our contribution shrinks to a cost reduction, which we will quantify as judge calls needed to reach equal accuracy. We also expect dense views to raise attribution error, because the grazing views remain in the set and the MLLM, which cannot reason reliably about foreshortening according to Park et al., may count their blur as defects. The clean-from-bad-views pairs measure exactly this.
Why not just zoom in or raise render resolution? Higher resolution and closer cameras fix the distance term in the footprint but not the obliquity term. A texture viewed at 80° is compressed by a factor of about six along one axis however many pixels the renderer uses, because the texture itself is minified and filtered. The distance-only ablation tests this directly. If it closes most of the gap, orientation matters less than we claim, and we will report that.
Why not just weight T3Bench's per-view ImageReward scores by incident angle? This is cheap, uses no language model, and repairs the equal-weighting problem that Pixel-GS identifies in optimization. We include it as an additional baseline. The reason to expect it to fall short is that ImageReward and CLIP produce one holistic score per image, so they cannot flag a misspelled label or a duplicated face. The defect-type breakdown in the GSO benchmark will show which defects benefit from language-model judgment and which do not.
Why not just show the judge the flattened texture image (the UV map)? A UV atlas breaks surfaces apart along seams and distorts them. As the TEXTure line of work shows, it hides whether neighboring regions match in 3D, and many neural and Gaussian assets have no clean atlas at all. We run it as a baseline on GSO meshes, where atlases exist. We expect it to do reasonably on isolated blur but fail on seam and Janus defects.
Decisive checks
The proposal succeeds if, at matched $K$, precision-covering views improve defect detection on the GSO benchmark and pairwise human agreement on T3Bench-generated assets. The gains must hold for both judges, and the precision-covering views must also cut attribution error relative to both uniform and dense sampling. It fails if dense uniform sampling matches it on detection and attribution at any practical budget, or if the distance-only ablation recovers the gain. In either case, orientation-aware selection is not doing the work. Specular and glossy assets violate the matte-surface (Lambertian) assumption behind the footprint model. We report them separately rather than hide them.
References
- On the Precision of Textures
- Differentiable Stereopsis: Meshes from multiple views using differentiable rendering
- Mesh-based 3D textured urban mapping
- TEXTure: Text-Guided Texturing of 3D Shapes
- Pixel-GS: Density Control with Pixel-Aware Gradient for 3D Gaussian Splatting
- T3Bench: Benchmarking Current Progress in Text-to-3D Generation
- MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark
- Mitigating Perceptual Judgment Bias in Multimodal LLM-as-a-Judge for Mathematical and Scientific Reasoning Evaluation
- Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items