3D Quality Assessment

3D-PAQA

Towards preference-aligned quality assessment for real 3D assets with natural artifacts, criterion-wise diagnosis, and scalable supervision.

JiHyuk Byun and Seon Joo Kim

Workshops on Image Processing and Image Understanding (IPIU), 2026

Paper Code Coming Soon Dataset Coming Soon

Why 3D-PAQA?

Conventional 3D quality assessment often measures controlled synthetic distortions. Real 3D repositories are different: assets vary in geometry, texture, material, plausibility, artifacts, and the quality judgments people actually prefer.

3D-PAQA reframes the task as direct per-asset scoring. Given an arbitrary 3D asset, it predicts a quality vector across human-relevant criteria without a pristine reference or expensive pairwise A/B comparison.

Distortion-centric 3D-QA

  • Synthetic degradations
  • Reference-oriented fidelity
  • Severity of distortion

3D-PAQA

  • Natural artifacts
  • Human-aligned preference
  • Absolute quality of each asset
3D-PAQA annotation overview with exemplar anchors and MLLM prompts
Exemplar-anchored visual and textual prompting stabilizes criterion-wise annotations for 3D assets.
260K+3D assets for preference-aligned supervision
10semantic domains sampled from real asset collections
6holistic and component-level quality criteria
33Mparameter 3D evaluator distilled from annotations

What the project contributes

New task

3D-PAQA evaluates the quality of an individual 3D asset through a human-aligned quality vector.

Scalable labels

Exemplar anchors and relative ranking turn subjective MLLM judgments into stable large-scale annotations.

Efficient evaluator

A compact 3D-native model predicts criterion-wise quality signals for evaluation and curation workflows.

Qualitative results

Explore examples ranked by criterion. Each row moves from stronger to weaker predicted quality and keeps the paired Ours and human scores from the qualitative analysis.

Criteria