Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2603.08064 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914387440173056 |
|---|---|
| author | Jia, Zexi Luo, Pengcheng Zhong, Yijia Zhang, Jinchao Zhou, Jie |
| author_facet | Jia, Zexi Luo, Pengcheng Zhong, Yijia Zhang, Jinchao Zhou, Jie |
| contents | Most evaluations of generative models rely on feature-distribution metrics such as FID, which operate on continuous recognition features that are explicitly trained to be invariant to appearance variations, and thus discard cues critical for perceptual quality. We instead evaluate models in the space of discrete visual tokens, where modern 1D image tokenizers compactly encode both semantic and perceptual information and quality manifests as predictable token statistics. We introduce Codebook Histogram Distance (CHD), a training-free distribution metric in token space, and Code Mixture Model Score (CMMS), a no-reference quality metric learned from synthetic degradations of token sequences. To stress-test metrics under broad distribution shifts, we further propose VisForm, a benchmark of 210K images spanning 62 visual forms and 12 generative models with expert annotations. Across AGIQA, HPDv2/3, and VisForm, our token-based metrics achieve state-of-the-art correlation with human judgments. We will release all code and datasets to facilitate future research, with the code publicly available at https://github.com/zexiJia/1d-Distance. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_08064 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Evaluating Generative Models via One-Dimensional Code Distributions Jia, Zexi Luo, Pengcheng Zhong, Yijia Zhang, Jinchao Zhou, Jie Computer Vision and Pattern Recognition Most evaluations of generative models rely on feature-distribution metrics such as FID, which operate on continuous recognition features that are explicitly trained to be invariant to appearance variations, and thus discard cues critical for perceptual quality. We instead evaluate models in the space of discrete visual tokens, where modern 1D image tokenizers compactly encode both semantic and perceptual information and quality manifests as predictable token statistics. We introduce Codebook Histogram Distance (CHD), a training-free distribution metric in token space, and Code Mixture Model Score (CMMS), a no-reference quality metric learned from synthetic degradations of token sequences. To stress-test metrics under broad distribution shifts, we further propose VisForm, a benchmark of 210K images spanning 62 visual forms and 12 generative models with expert annotations. Across AGIQA, HPDv2/3, and VisForm, our token-based metrics achieve state-of-the-art correlation with human judgments. We will release all code and datasets to facilitate future research, with the code publicly available at https://github.com/zexiJia/1d-Distance. |
| title | Evaluating Generative Models via One-Dimensional Code Distributions |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2603.08064 |