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..
CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation
Authors: Bowen Song, Zecheng Zhang, Zhaoxu Luo, Jason Hu, Wei Yuan, Jing Jia, Zhengxu Tang, Guanyang Wang, Liyue Shen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments to compare our proposed sampling approach with other methods on both sampling controllability and sampled data quality. Results show that our CCS method achieves more precisely controlled sampling while maintaining superior sample quality and diversity, enhancing the applications of precise image editing. |
| Researcher Affiliation | Collaboration | 1Department of Electrical Engineering and Computer Science, University of Michigan 2Trace Root.AI 3Department of Statistics, Rutgers University 4Department of Computer Science, Rutgers University |
| Pseudocode | Yes | Our CCS algorithm is described in Algorithm 1. Algorithm 2 Controller Tuning (CT) ... Alg. 3 demonstrates using P-CCS for constrained sampling based on Stable Diffusion. Alg. 4 demonstrates using P-CCS for precise image editing based on Stable Diffusion. Alg. 5 demonstrates using P-CCS for improving the sample quality, instead of controllability. |
| Open Source Code | Yes | The code is available at https://github.com/efzero/diffusioncontroller. |
| Open Datasets | Yes | We perform extensive experiments on both pixel diffusion models on the FFHQ [22] and CIFAR-10 [25] dataset and latent diffusion models on the Celeba-HQ and f Mo W dataset [5]. |
| Dataset Splits | Yes | Experimental setup: We FFHQ-256 [22] and Celeb A-HQ [43] test set images as target images. We use ADM (a pixel diffusion model) for FFHQ, and Stable Diffusion for Celeb A-HQ. Baselines: ... For each experiment, we first sample 50 images as target images from each validation dataset from FFHQ [22], CIFAR-10[25], and Celeba-HQ [43]. |
| Hardware Specification | Yes | The table reports the controller tuning NFE per batch, and sampling NFE per batch for each baseline. ... using Stable Diffusion tested on one A40 GPU. |
| Software Dependencies | No | We take the pretrained models for FFHQ and CIFAR-10 from the improved/guided diffusion repos [30, 13] for the pixel diffusion experiments, and the Stable Diffusion 1.5 [31] for latent diffusion experiments. The paper mentions specific models but does not list ancillary software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | For LDMs, we set t0 = 45, where T = 50 due to DDIM inversion performing worse with classifier-free guidance than unconditional models. We set the r MSE target to be 0.12, 0.11 for FFHQ and CIFAR-10 respectively, and 0.07 for Stable Diffusion experiments to test diverse control targets. The tolerance is set to be 0.01 in all cases. |