Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Machine Unlearning in 3D Generation: A Perspective-Coherent Acceleration Framework
Authors: Shixuan Wang, Jingwen Ye, Xinchao Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on the typical 3D generation models (Zero123 and Zero123XL), demonstrating that our approach achieves a 30% speedup, while effectively unlearning target concepts without compromising generation quality. ... We evaluate our approach across three key aspects: generation quality and efficiency, effectiveness of unlearning, and preservation of retained knowledge. The following metrics are used. (1) SSIM: Structural Similarity Index... (2) LPIPS: Learned Perceptual Image Patch Similarity... (3) PSNR: The difference in Peak Signal-to-Noise Ratio... (4) FID: The change in Fréchet Inception Distance... (5) Inference Time (Speedup Analysis): Measures the computational efficiency... |
| Researcher Affiliation | Academia | Shixuan Wang Jingwen Ye Xinchao Wang National University of Singapore EMAIL, EMAIL |
| Pseudocode | Yes | We propose a novel two-stage framework that integrates dynamic timestep skipping with directional unlearning, enabling efficient and precise removal of targeted concepts from a diffusion-based generative model. This section provides supplementary algorithms for our proposed framework, including Algorithm 1 and Algorithm 2. |
| Open Source Code | Yes | The code is available at: https://github.com/sxxsxw/Fast-3D-Unlearn-with-Skipacceleration |
| Open Datasets | Yes | Datasets. We conduct experiments on three types of data: (1) Ten 3D Minions models collected from the internet (denoted as Min10), with one used for training and the rest for testing; (2) Rendered 3D objects from Objaverse 1.0, including sculptures, traffic barriers, and fire hydrants; and (3) A subset of five Objaverse models rendered from 24 viewpoints, each with 35 images, totaling approximately 4,200 ground-truth images. |
| Dataset Splits | Yes | Datasets. We conduct experiments on three types of data: (1) Ten 3D Minions models collected from the internet (denoted as Min10), with one used for training and the rest for testing; |
| Hardware Specification | Yes | Table 4: Experimental Settings for Different Reference Angles and Unlearn Effects Parameter Reference Angles Experiment Target Forget Images Experiment GPU NVIDIA A100 80GB NVIDIA A6000 48GB |
| Software Dependencies | No | The paper mentions the Adam optimizer and uses diffusion models, but it does not specify versions for any programming languages, libraries, or frameworks (e.g., Python, PyTorch, CUDA, etc.). |
| Experiment Setup | Yes | Table 3: Hyperparameters for Fake Score and Generator Parameter Fake Score Generator λ 1.0 1.0 µ 0.01 0.01 Optimizer Adam Adam Learning Rate 4 10 6 6 10 6 β1 0.0 0.0 β2 0.999 0.999 ϵt 10 8 10 8 |