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

arXiv: https://arxiv.org/abs/2604.10095 · HTML: https://mlsanity.com/paper/2604.10095/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 34 of its 37 references that are in the map.
- First posted on arXiv 11 Apr 2026.

## Builds on (most cited first, 30 of 34)

- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [On the Opportunities and Risks of Foundation Models](https://mlsanity.com/paper/2108.07258/) (2021)
- [SAM 2: Segment Anything in Images and Videos](https://mlsanity.com/paper/2408.00714/) (2024)
- [DUSt3R: Geometric 3D Vision Made Easy](https://mlsanity.com/paper/2312.14132/) (2023)
- [VGGT: Visual Geometry Grounded Transformer](https://mlsanity.com/paper/2503.11651/) (2025)
- [Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World](https://mlsanity.com/paper/1703.06907/) (2017)
- [Gradient Descent Finds Global Minima of Deep Neural Networks](https://mlsanity.com/paper/1811.03804/) (2018)
- [GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection](https://mlsanity.com/paper/2403.03507/) (2024)
- [MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motion](https://mlsanity.com/paper/2410.03825/) (2024)
- [PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models](https://mlsanity.com/paper/2404.02948/) (2024)
- [Function4D: Real-time Human Volumetric Capture from Very Sparse Consumer RGBD Sensors](https://mlsanity.com/paper/2105.01859/) (2021)
- [ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs](https://mlsanity.com/paper/2311.13600/) (2023)
- [LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias](https://mlsanity.com/paper/2410.17242/) (2024)
- [Generalization in diffusion models arises from geometry-adaptive harmonic representations](https://mlsanity.com/paper/2310.02557/) (2023)
- [LoRA-GA: Low-Rank Adaptation with Gradient Approximation](https://mlsanity.com/paper/2407.05000/) (2024)
- [MUSt3R: Multi-view Network for Stereo 3D Reconstruction](https://mlsanity.com/paper/2503.01661/) (2025)
- [Implicit Style-Content Separation using B-LoRA](https://mlsanity.com/paper/2403.14572/) (2024)
- [RayZer: A Self-supervised Large View Synthesis Model](https://mlsanity.com/paper/2505.00702/) (2025)
- [High-fidelity 3D Human Digitization from Single 2K Resolution Images](https://mlsanity.com/paper/2303.15108/) (2023)
- [K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs](https://mlsanity.com/paper/2502.18461/) (2025)
- [ClearPose: Large-scale Transparent Object Dataset and Benchmark](https://mlsanity.com/paper/2203.03890/) (2022)
- [MatSynth: A Modern PBR Materials Dataset](https://mlsanity.com/paper/2401.06056/) (2024)
- [A Purely Algebraic Justification of the Kabsch-Umeyama Algorithm](https://mlsanity.com/paper/1902.03138/) (2019)
- [E-RayZer: Self-supervised 3D Reconstruction as Spatial Visual Pre-training](https://mlsanity.com/paper/2512.10950/) (2025)
- [MegaSynth: Scaling Up 3D Scene Reconstruction with Synthesized Data](https://mlsanity.com/paper/2412.14166/) (2024)
- [LoRA3D: Low-Rank Self-Calibration of 3D Geometric Foundation Models](https://mlsanity.com/paper/2412.07746/) (2024)
- [The Implicit Values of A Good Hand Shake: Handheld Multi-Frame Neural Depth Refinement](https://mlsanity.com/paper/2111.13738/) (2021)
- [GP3: A 3D Geometry-Aware Policy with Multi-View Images for Robotic Manipulation](https://mlsanity.com/paper/2509.15733/) (2025)
- [Shakes on a Plane: Unsupervised Depth Estimation from Unstabilized Photography](https://mlsanity.com/paper/2212.12324/) (2022)
- [LiON-LoRA: Rethinking LoRA Fusion to Unify Controllable Spatial and Temporal Generation for Video Diffusion](https://mlsanity.com/paper/2507.05678/) (2025)

## Abstract (arXiv)

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.