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..
TRIM: Scalable 3D Gaussian Diffusion Inference with Temporal and Spatial Trimming
Authors: Zeyuan Yin, Xiaoming Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments and analysis demonstrate that TRIM significantly improves both the efficiency and quality of 3D generation. 39th Conference on Neural Information Processing Systems (NeurIPS 2025). |
| Researcher Affiliation | Academia | Zeyuan Yin Xiaoming Liu Department of Computer Science and Engineering, Michigan State University, East Lansing, MI, USA EMAIL |
| Pseudocode | Yes | Algorithm 1 Data Synthesis for Latent Selector Training |
| Open Source Code | Yes | Answer: [Yes] Justification: We provide all experiments details to reproduce our results, including datasets, models, parameter settings and code for main experiments. |
| Open Datasets | Yes | We adopt Diff Splat [1] as our main backbone model, trained on the G-Objaverse dataset [28]. For the text-to-3D generation task, we evaluate on T3Bench [29], which consists of 300 descriptive prompts about a single object, a single object with surrounding context, or multiple objects. ... For the image-to-3D generation task, we randomly select 300 objects from the Google Scanned Objects (GSO) dataset [33]. |
| Dataset Splits | Yes | The dataset was then split into a 7:3 ratio for training and testing, respectively. |
| Hardware Specification | Yes | All results are reported based on RTX A6000 GPU. |
| Software Dependencies | No | Our 3D generation model with the backbone of Stable-Diffusion-3.5-Medium uses the original flow matching Euler ODE solver with 28 steps in the main experiments. |
| Experiment Setup | Yes | Our 3D generation model with the backbone of Stable-Diffusion-3.5-Medium uses the original flow matching Euler ODE solver with 28 steps in the main experiments. The classifier-free guidance scale is set to 7 for text-to-3D generation and 2 for image-to-3D generation. ... We employ the Adam W optimizer with a learning rate of 0.001, a weight decay of 0.01, and a cosine weight decay schedule. Training is conducted with a batch size of 64 for 20 epochs. |