Noise Conditional Variational Score Distillation
Authors: Xinyu Peng, Ziyang Zheng, Yaoming Wang, Han Li, Nuowen Kan, Wenrui Dai, Chenglin Li, Junni Zou, Hongkai Xiong
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate NCVSD through extensive experiments, including class-conditional image generation and inverse problem solving. By scaling the test-time compute, our method outperforms teacher diffusion models and is on par with consistency models of larger sizes. Additionally, with significantly fewer NFEs than diffusion-based methods, we achieve record-breaking LPIPS on inverse problems. The source code is available at https: //github.com/xypeng9903/ncvsd. ... In this section, we evaluate proposed method by testing the distilled generative denoisers on i) class-conditional image generation on Image Net-64 64 and Image Net-512 512 datasets (Deng et al., 2009), and ii) plug-and-play inverse problem solving with Pn P-GD on FFHQ-256 256 dataset (Karras et al., 2019). |
| Researcher Affiliation | Collaboration | 1Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China 2Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, China 3Meituan Inc, China. Correspondence to: Ziyang Zheng <EMAIL>, Yaoming Wang <EMAIL>, Wenrui Dai <EMAIL>. |
| Pseudocode | Yes | For further details and pseudo code or a single iteration of the NCVSD training iteration, please refer to Appendix B.2 and Algorithm 3. ... The pseudocode for multi-step sampling is presented in Algorithm 2. ... The pseudocode for the Pn P-GD procedure is provided in Algorithm 1. |
| Open Source Code | Yes | The source code is available at https: //github.com/xypeng9903/ncvsd. |
| Open Datasets | Yes | In this section, we evaluate proposed method by testing the distilled generative denoisers on i) class-conditional image generation on Image Net-64 64 and Image Net-512 512 datasets (Deng et al., 2009), and ii) plug-and-play inverse problem solving with Pn P-GD on FFHQ-256 256 dataset (Karras et al., 2019). |
| Dataset Splits | Yes | In this section, we evaluate proposed method by testing the distilled generative denoisers on i) class-conditional image generation on Image Net-64 64 and Image Net-512 512 datasets (Deng et al., 2009), and ii) plug-and-play inverse problem solving with Pn P-GD on FFHQ-256 256 dataset (Karras et al., 2019). |
| Hardware Specification | Yes | A100 GPU hours 650 1100 1450 650 1100 1450 300 |
| Software Dependencies | No | The paper does not explicitly state specific version numbers for software dependencies such as programming languages, libraries (e.g., PyTorch, TensorFlow), or other key software components used in their experiments. It only mentions using the 'EDM2 codebase' without a specific version. |
| Experiment Setup | Yes | Table 3. Hyperparameter details for training and inference. ... Training details Effective batch size 2048 ... Learning rate max ... Adam β1 0.9 Adam β2 0.99 ... Table 4. Hyperparameter details for Pn P-GD. ... Annealing schedule details Steps (N) 50 ... EMA schedule details EMA threshold (σema) 0.2 ... Likelihood step details Energy strength β 1e-4 ULA step 100 ULA C1 0.1 ULA C2 0.1 |