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..
Towards Reliable Identification of Diffusion-based Image Manipulations
Authors: Alex Costanzino, Woody Bayliss, Juil Sock, Marc Gorriz Blanch, Danijela Horak, Ivan Laptev, Philip H.S. Torr, Fabio Pizzati
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate these techniques to significantly improve both the accuracy of our method and its generalisation to a large number of diffusion models. To support realistic evaluation, we further introduce BBC-PAIR, a new comprehensive benchmark, with images tampered by 28 diffusion models. Our experiments show that RADAR achieves excellent results, outperforming the state-of-the-art in detecting and localising image edits made by both seen and unseen diffusion models. |
| Researcher Affiliation | Collaboration | Alex Costanzino University of Bologna Woody Bayliss BBC R&D Juil Sock BBC R&D Marc Gorriz Blanch BBC R&D Danijela Horak BBC R&D Ivan Laptev MBZUAI Philip Torr University of Oxford Fabio Pizzati MBZUAI |
| Pseudocode | No | The paper describes the methodology using textual descriptions and diagrams (e.g., Figure 1, Figure 2, Figure 3) to illustrate the architecture and process. There are no explicit sections or blocks labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | No | Further information about our code, data and models, including separate licensing terms, will be publicly available at https://alex-costanzino.github.io/radar/. The Neur IPS Paper Checklist also states: 'the code for both method and data generation will be released upon eventual acceptance.' |
| Open Datasets | Yes | For the creation of BBC-PAIR, we gathered data from the following datasets: Open Images-v7: https://storage.googleapis.com/openimages/web/index. html, released under Apache 2.0 license; Coco Glide: https://github.com/grip-unina/Tru For released under MIT license; SID-Set: https://huggingface.co/datasets/saberzl/SID_Set released under Creative Commons Attribution 4.0 International License; Safire MS-Expert: https://www.kaggle.com/datasets/qsii24/ safire-safirems-expert-multi-source-dataset released under CC BY-NC 4.0 license. |
| Dataset Splits | Yes | BBC-PAIR-ID We consider 15,000/100 randomly sampled Open Images-v7 [50] as Dorig for train/test, and we use the data generation pipeline in Section 3.1 with 10 inpainters in I... We generate 150,000/1,000 images for training/testing, processing each image in Dorig with all I I. |
| Hardware Specification | Yes | Training details. We train RADAR for 120 epochs with batch-size 16, dropout (p = 0.1) and NAdam optimiser [44] (lr=10 4 and decay of 10 5) using 4 NVIDIA A100 80GB GPUs, optimizing the fusion block, γ and ϕ while keeping ES and EG frozen. More details are given in the Appendix. [...] Machine configuration Experiments were conducted on a high-performance computing server equipped with four NVIDIA A100 GPUs (80 GB each, PCIe) and a 96-core CPU, with a total of 866 GB of RAM. The system ran with NVIDIA Driver version 550.144.03 and CUDA Driver version 12.4. |
| Software Dependencies | Yes | Machine configuration Experiments were conducted on a high-performance computing server equipped with four NVIDIA A100 GPUs (80 GB each, PCIe) and a 96-core CPU, with a total of 866 GB of RAM. The system ran with NVIDIA Driver version 550.144.03 and CUDA Driver version 12.4. |
| Experiment Setup | Yes | Training details. We train RADAR for 120 epochs with batch-size 16, dropout (p = 0.1) and NAdam optimiser [44] (lr=10 4 and decay of 10 5) using 4 NVIDIA A100 80GB GPUs, optimizing the fusion block, γ and ϕ while keeping ES and EG frozen. More details are given in the Appendix. |