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..

InstaInpaint: Instant 3D-Scene Inpainting with Masked Large Reconstruction Model

Authors: Junqi You, Chieh Lin, Weijie Lyu, Zhengbo Zhang, Ming-Hsuan Yang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on two standard 3D inpainting benchmarks with diverse and challenging real-world scenes. In Fig. 2, we highlight that Insta Inpaint achieves state-of-the-art performance on both speed and quality axes. We also provide ablations of the key design choices on the masking strategies and the encoding design.
Researcher Affiliation Academia Junqi You1,2 Chieh Hubert Lin1 Weijie Lyu1 Zhengbo Zhang3 Ming-Hsuan Yang1 1 UC Merced 2 Shanghai Jiao Tong University 3 Singapore University of Technology and Design
Pseudocode No The paper describes its methodology in Section 3, titled 'Methodology', which uses descriptive text and figures (e.g., Figure 4: 'Overall pipeline of Masked Finetuning') to explain the process. There are no explicitly labeled pseudocode or algorithm blocks.
Open Source Code No 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: The code release requires additional review. We will put our best effort into releasing the codes and the pretrained model.
Open Datasets Yes Implementations. Insta Inpaint is trained on DL3DV [9]. which is one of the largest open-source realworld 3D multi-view datasets... Evaluation Datasets. We evaluate Insta Inpaint on two real-world datasets, SPIn-Ne RF [30] and LLFF [45].
Dataset Splits Yes We split each scene videos into 15-frame video clips and randomly select one at each iteration.
Hardware Specification Yes For all experiments, we train Insta Inpaint for 12800 iterations in the second stage with 8 A6000 GPUs, which costs about 40 hours.
Software Dependencies No The paper mentions several models and frameworks (e.g., GS-LRM, ViT-based transformer, Adam optimizer, COLMAP) and programming concepts, but it does not list specific software dependencies with their version numbers (e.g., Python version, PyTorch version, CUDA version, or any specific library versions).
Experiment Setup Yes We set the image patch size to 8 and the token dimension to be 1024. In the first stage, we train the model on a resolution of 256 256 for 80K iterations. In the second stage, we finetune Insta Inpaint on a resolution of 512 512... Insta Inpaint takes 4 input views and is supervised on 8 views in both training stages. For all experiments, we train Insta Inpaint for 12800 iterations in the second stage... We use Adam optimizer with a learning rate of 8e 5 and a total batch size of 80.