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..

GeoRemover: Removing Objects and Their Causal Visual Artifacts

Authors: Zixin Zhu, Haoxiang Li, Xuelu Feng, He Wu, Chunming Qiao, Junsong Yuan

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method achieves state-of-the-art performance in removing both objects and their associated artifacts on two popular benchmarks. The project page is available at https://buxiangzhiren.github.io/Geo Remover.
Researcher Affiliation Collaboration Zixin Zhu1,2 Haoxiang Li2 Xuelu Feng1 He Wu2 Chunming Qiao1 Junsong Yuan1 1University at Buffalo 2Pixocial Technology EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes the methodology using textual descriptions and diagrams (e.g., Figure 2, Figure 3), but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes The project page is available at https://buxiangzhiren.github.io/Geo Remover. We plan to publicly release the full code, pretrained models, and data processing scripts upon acceptance. Although we do not include the code or data in the submission to preserve anonymity, the paper provides sufficient experimental details (Sections 4.1 4.3) to enable reproduction. We will include full implementation and reproducibility instructions in the camera-ready version.
Open Datasets Yes We use the training set from the RORD [34] dataset as our primary training data. ... For evaluation, we follow prior works such as Smart Eraser [21] and Omni Eraser [9], We use both RORD-Val and Removal Bench [9] as our primary benchmarks. ... To evaluate our method s ability to remove causal visual artifacts, we construct the Caus Rem benchmark, consisting of 200 realworld images (100 with shadows, 100 with reflections).
Dataset Splits Yes We use the training set from the RORD [34] dataset as our primary training data. RORD is a large-scale real-world object removal dataset consisting of 516,705 images captured under 3,447 unique indoor scenes. ... For evaluation, we follow prior works such as Smart Eraser [21] and Omni Eraser [9], We use both RORD-Val and Removal Bench [9] as our primary benchmarks.
Hardware Specification Yes Stage 1 is trained for 17,000 steps on 8 NVIDIA H100 GPUs, taking approximately 24 hours, while Stage 2 requires around 60 hours for the same number of steps.
Software Dependencies No The paper mentions software components like Depth Anything [32], FLUX.1-Filldev [8], and Lo RA [33], but does not provide specific version numbers for these tools.
Experiment Setup Yes All images are processed at a resolution of 1024 1024. For both stages, we use a batch size of 24, a learning rate of 1 10 4, and a guidance scale of 1.0. The text prompt a beautiful scene is used during both training and inference. Stage 1 is trained for 17,000 steps on 8 NVIDIA H100 GPUs, taking approximately 24 hours, while Stage 2 requires around 60 hours for the same number of steps.