# Mining Attribute Subspaces for Efficient Fine-tuning of 3D Foundation Models

Source: [Mining Attribute Subspaces for Efficient Fine-tuning of 3D Foundation Models](<https://www.sota2.com/research/paper/69ddc9f17d0172140ec7989f>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-28T05:00:05.404Z

Authors: Yu Jiang, Hanwen Jiang, Ahmed Abdelkader, Wen-Sheng Chu, Brandon Y. Feng, Zhangyang Wang, Qixing Huang

Published: 2026-04-11T08:33:13.000Z

## Abstract

With the emergence of 3D foundation models, there is growing interest in fine-tuning them for downstream tasks, where LoRA is the dominant fine-tuning paradigm. As 3D datasets exhibit distinct variations in texture, geometry, camera motion, and lighting, there are interesting fundamental questions: 1) Are there LoRA subspaces associated with each type of variation? 2) Are these subspaces disentangled (i.e., orthogonal to each other)? 3) How do we compute them effectively? This paper provides answers to all these questions. We introduce a robust approach that generates synthetic datasets with controlled variations, fine-tunes a LoRA adapter on each dataset, and extracts a LoRA sub-space associated with each type of variation. We show that these subspaces are approximately disentangled. Integrating them leads to a reduced LoRA subspace that enables efficient LoRA fine-tuning with improved prediction accuracy for downstream tasks. In particular, we show that such a reduced LoRA subspace, despite being derived entirely from synthetic data, generalizes to real datasets. An ablation study validates the effectiveness of the choices in our approach.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Point Map Estimation on 2K2K (Cross-domain)](<https://www.sota2.com/research/sota/point-map-estimation-on-2k2k-cross-domain.md>): Point Map Estimation; 2K2K (Cross-domain); Accuracy (Acc): 2.513.

- [Face Reconstruction on Real Face Dataset](<https://www.sota2.com/research/sota/face-reconstruction-on-real-face-dataset.md>): Face Reconstruction; Real Face Dataset; Absolute Relative Error: 2.151.

- [Point Map Estimation on THuman 2.1 (In-domain)](<https://www.sota2.com/research/sota/point-map-estimation-on-thuman-2-1-in-domain.md>): Point Map Estimation; THuman 2.1 (In-domain); Accuracy (Acc): 3.392.

- [Face Reconstruction on Synthetic Face Dataset](<https://www.sota2.com/research/sota/face-reconstruction-on-synthetic-face-dataset.md>): Face Reconstruction; Synthetic Face Dataset; Accuracy: 5.921.

- [Point Map Estimation on ClearPose (In-domain)](<https://www.sota2.com/research/sota/point-map-estimation-on-clearpose-in-domain.md>): Point Map Estimation; ClearPose (In-domain); Accuracy: 1.764.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.10095>)

[PDF](<https://arxiv.org/pdf/2604.10095.pdf>)
