# Judging to Improve: A De-biased VLM-as-3D-Judge Protocol for Single-Image 3D Generation

arXiv: https://arxiv.org/abs/2606.20364 · HTML: https://mlsanity.com/paper/2606.20364/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 18 of its 20 references that are in the map.
- First posted on arXiv 18 Jun 2026.

## Builds on (most cited first, 18 of 18)

- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [DINOv2: Learning Robust Visual Features without Supervision](https://mlsanity.com/paper/2304.07193/) (2023)
- [Flow Matching for Generative Modeling](https://mlsanity.com/paper/2210.02747/) (2022)
- [Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow](https://mlsanity.com/paper/2209.03003/) (2022)
- [InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks](https://mlsanity.com/paper/2312.14238/) (2023)
- [InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models](https://mlsanity.com/paper/2504.10479/) (2025)
- [Large Language Models are not Fair Evaluators](https://mlsanity.com/paper/2305.17926/) (2023)
- [Diffusion Model Alignment Using Direct Preference Optimization](https://mlsanity.com/paper/2311.12908/) (2023)
- [Structured 3D Latents for Scalable and Versatile 3D Generation](https://mlsanity.com/paper/2412.01506/) (2024)
- [ORPO: Monolithic Preference Optimization without Reference Model](https://mlsanity.com/paper/2403.07691/) (2024)
- [3D-FUTURE: 3D Furniture shape with TextURE](https://mlsanity.com/paper/2009.09633/) (2020)
- [TripoSR: Fast 3D Object Reconstruction from a Single Image](https://mlsanity.com/paper/2403.02151/) (2024)
- [Flexible Isosurface Extraction for Gradient-Based Mesh Optimization](https://mlsanity.com/paper/2308.05371/) (2023)
- [SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement](https://mlsanity.com/paper/2408.00653/) (2024)
- [DreamDPO: Aligning Text-to-3D Generation with Human Preferences via Direct Preference Optimization](https://mlsanity.com/paper/2502.04370/) (2025)
- [DSO: Aligning 3D Generators with Simulation Feedback for Physical Soundness](https://mlsanity.com/paper/2503.22677/) (2025)
- [A Cross-Model VLM-Judge Protocol for Single-Image 3D Mesh Quality (and Why Cheap Proxies Fall Short)](https://mlsanity.com/paper/2606.18451/) (2026)

## Abstract (arXiv)

A companion study established a de-biased, cross-model VLM-as-3D-judge that reliably ranks single-image-to-3D mesh quality where cheap geometry and CLIP proxies fall short. This paper asks: can that judge's preferences specialize a strong open generator, TRELLIS, on one asset class (furniture), cheaply and without human labels? Taking the judge from ranking to optimization is where the work lives. Pushing a VLM judge into the training and evaluation loop exposes failure modes ranking never triggered, so our contribution is an optimization-grade hardening of the judge: a training judge (Qwen2.5-VL-7B) held distinct from an evaluation judge (InternVL3-8B) to break circularity; position-bias correction; and fixes for three failure modes (image overload, geometry-hiding splat renders, and reference-free judging that rewards clean-but-wrong outputs), with calibration evidence (clear-gap win-rate 0.83-1.0; base-vs-base ~0.5). Using this protocol as an independent evaluator, and working only from public models and data with lightweight parameter-efficient adaptation, we find our methods match the strong base rather than exceed it. Independent base samples carry essentially no learnable preference (0.94 order-flip rate), so signal must be engineered by quality-contrastive construction. Across six adaptation methods, two input regimes, and a severity sweep, the most targeted - conditioner repair under severe degradation - reaches parity (0.50) with the base, while no method clears the >=65% win-rate target. The result is mechanistic: clean inputs saturate the judge, flow-DIT fine-tuning washes out through the sampler, and conditioning repair is the locus that moves geometry. Win-rates are directional at n=8 objects. Matching a strong public-data base with cheap adaptation is itself informative: exceeding it needs more than lightweight PEFT on public data, and the judge protocol is reusable.