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..
Diffusion Feature Field for Text-based 3D Editing with Gaussian Splatting
Authors: Eunseo Koh, Sangeek Hyun, MinKyu Lee, Jiwoo Chung, Kangmin Seo, Jae-Pil Heo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our method achieves state-of-the-art performance in terms of CLIP similarity, and better preserves structural and semantic consistency compared to existing approaches. |
| Researcher Affiliation | Academia | Eunseo Koh Sangeek Hyun Min Kyu Lee Jiwoo Chung Kangmin Seo Jae-Pil Heo Sungkyunkwan University EMAIL |
| Pseudocode | No | The paper describes methods using mathematical formulations (e.g., equations 1-10) and provides high-level process diagrams (Figure 2), but it does not include a dedicated section or figure with structured pseudocode or algorithm blocks. |
| Open Source Code | No | We will release the code and data after accept. |
| Open Datasets | Yes | We conduct the experiments on 5 scenes from IN2N [9] and Dream Editor [43] |
| Dataset Splits | No | For each scene, we edit 20 views of the source scene and use the edited images as the reconstruction target to update 3DGS. For training details of 3DGS, we iteratively update the parameters of 3DGS for a total of 1000 iterations, while image editing is performed every 500 iterations. We use every image available in the training dataset for evaluation, although only use 20 views of images for editing. |
| Hardware Specification | No | The paper states, "The experiments compute resources would be in supplementary," indicating that specific hardware details are not provided within the main text of the paper. |
| Software Dependencies | No | For a fair comparison, we use Instruct Pix2Pix as an image editor, which is the most popular editor used by prior methods [4, 3]. During editing, we utilize classifier-free guidance with scales of 7.5 for textual conditions and 1.5 for image conditions, unless specified otherwise. We use Adam optimizer to update the 3DGS parameters and LPIPS [41] and L1 loss as reconstruction objectives, following IN2N [9]. |
| Experiment Setup | Yes | During editing, we utilize classifier-free guidance with scales of 7.5 for textual conditions and 1.5 for image conditions, unless specified otherwise. For training details of 3DGS, we iteratively update the parameters of 3DGS for a total of 1000 iterations, while image editing is performed every 500 iterations. We use Adam optimizer to update the 3DGS parameters and LPIPS [41] and L1 loss as reconstruction objectives, following IN2N [9]. For training our dual-encoder, we set λ1 = 1.0 and λ2 = 0.05. |