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..
CPO: Condition Preference Optimization for Controllable Image Generation
Authors: Zonglin Lyu, Ming Li, Xinxin Liu, Chen Chen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that CPO significantly improves controllability over the state-ofthe-art Control Net++ across multiple control types: over 10% error rate reduction in segmentation, 70 80% in human pose, and consistent 2 5% reductions in edge and depth maps. |
| Researcher Affiliation | Academia | Zonglin Lyu Ming Li Xinxin Liu Chen Chen Institute of Artificial Intelligence University of Central Florida Orlando, FL 32816 EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical formulations for loss functions (Eq. 5, 6, 10, 11) and provides a data curation process in prose, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | 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: We will open-source everything, but we may not be able to submit it anonymously at this point. |
| Open Datasets | Yes | Following Control Net++ [11], we use ADE20K [22] and COCO-Stuff [23] for segmentation-to-image generation, and Multi Gen-20M [10] for edge and structure-based conditions, including Canny edges, HED, Lineart, and depth maps. In addition, we include COCO-Pose [46] and Human Art [47] for pose-to-image generation. ... New Dataset. We curate and will open-source our Condition Preference (CPO) Dataset. |
| Dataset Splits | No | The paper mentions using COCO-Pose for training and testing on COCO-Pose and Human Art, but does not provide specific training/test/validation split percentages, sample counts, or explicit splitting methodology for the datasets used in their experiments. |
| Hardware Specification | Yes | The training with batch size 16 and 10K steps takes approximately 8 hours to finish with 2 H100 GPUs. |
| Software Dependencies | No | The paper mentions in its NeurIPS checklist that 'required versions of packages' will be provided with the open-source code, but these specific software versions are not explicitly listed within the paper's text. |
| Experiment Setup | Yes | We include our training details on hyperparameters in Tab. 9. The training with batch size 16 and 10K steps takes approximately 8 hours to finish with 2 H100 GPUs. Readers may assume linear scaling when increasing the number of steps and batch size. Non-mentioned hyperparameters are in their default setup, such as optimizer configuration. |