--- title: "A Zipfian Analysis of Visual Token Distributions for AI-Generated Images — Shin 2026" description: "The paper asks whether photorealistic text-to-image systems reproduce the deeper statistical structure of real photographs or only their… 9 references." authors: ["Andrew Shin"] year: 2026 venue: "Proceedings of the 4th Workshop on Advances in Language and Vision Research (ALVR)" source: "https://aclanthology.org/2026.alvr-main.2/" doi: "10.18653/v1/2026.alvr-main.2" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/shin-2026-zipfian-analysis-visual-token-distributions/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/shin-2026-zipfian-analysis-visual-token-distributions/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/3/) / A Zipfian Analysis of Visual Token Distributions for AI-Generated Images # A Zipfian Analysis of Visual Token Distributions for AI-Generated Images **Authors:** Andrew Shin · **Year:** 2026 · **Venue:** Proceedings of the 4th Workshop on Advances in Language and Vision Research (ALVR) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.alvr-main.2/) · [PDF](https://aclanthology.org/2026.alvr-main.2.pdf) · [DOI 10.18653/v1/2026.alvr-main.2](https://doi.org/10.18653/v1/2026.alvr-main.2) · [OpenAlex](https://openalex.org/W7166890279) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary The paper asks whether photorealistic text-to-image systems reproduce the deeper statistical structure of real photographs or only their appearance. Borrowing tools from statistical linguistics, the author treats images as documents made of visual tokens and fits Zipf-style power laws to their frequency distributions. The corpus contains 50,000 images: 40,000 generated images from the Rapidata Image Generation Alignment Dataset, 10,000 each from DALL-E 3, Midjourney, Stable Diffusion 3 and Flux 1.1 Pro, plus 10,000 ImageNet-1k validation photos. Two tokenizations are used. Basic Image Features (BIFs) capture low-level geometry, and a K-means vector-quantized codebook captures texture. At the geometric level, generated and real images are statistically indistinguishable. Texture tokens, however, fit a log-normal distribution better than a pure power law, and for AI images they decay more steeply. This means synthetic images under-represent rare textures. Each generator shows its own entropy level, and longer, more specific prompts reduce texture diversity further. The results suggest statistical fingerprints that could support forensic detection, and they question the assumption that more detailed prompts yield more realistic output. ## Key points - **Tokenization:** - BIFs come from Gaussian derivative filters (σ = 1.0), with each pixel assigned to one of 7 symmetry classes. - Texture words are non-overlapping 4×4 pixel patches, quantized with a shared 512-word codebook built by mini-batch K-means. - Power-law fits use maximum likelihood estimation for α and Kolmogorov–Smirnov distance for x_min. A likelihood-ratio test (R) compares the power-law fit with a log-normal alternative. - **Aggregate results:** - For BIFs, AI and real images are statistically indistinguishable (p = 0.85 and 0.70; α 1.76 vs. 1.82). - For VQ texture tokens, both real and AI distributions favor log-normal (R \< 0). - AI images have a steeper slope (α 2.29 vs. 1.95 for real) and a lower x_min (1.42M vs. 1.93M), meaning their distribution reaches its decaying tail sooner. - **Per-model texture slopes:** - DALL-E 3 is steepest (α 2.44), which the author speculatively links to LLM-based prompt rewriting. - Stable Diffusion 3 (2.29) and Midjourney (2.25) fall in between. The author tentatively attributes this to VAE compression for SD3 and aesthetic RLHF for Midjourney. - Flux 1.1 Pro is closest to real images (2.17), possibly because of its flow-matching design. - **Prompt complexity:** In a comparison of 5,000 images per group, long prompts (>15 words) gave α 2.33 versus 1.93 for short prompts (\<6 words), along with a higher x_min. - **Limitations noted by the author:** - Texture tokens come from shallow pixel-patch clustering rather than semantic tokens. - The analysis uses a single filter scale. - The images were selected for aesthetic alignment, which may introduce bias. - Word count serves as the only proxy for prompt complexity. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) ## Institutions - [Keio University](https://md.research.elamur.ai/free/en/universities/keio-university/) — University, Tokyo, Japan · 3 works · 678 citations ## Cite Cite as: Andrew Shin. 2026. A Zipfian Analysis of Visual Token Distributions for AI-Generated Images. Proceedings of the 4th Workshop on Advances in Language and Vision Research (ALVR). The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{shin2026zipfian, title = {{A Zipfian Analysis of Visual Token Distributions for AI-Generated Images}}, author = {Andrew Shin}, booktitle = {Proceedings of the 4th Workshop on Advances in Language and Vision Research (ALVR)}, year = {2026}, doi = {10.18653/v1/2026.alvr-main.2}, url = {https://doi.org/10.18653/v1/2026.alvr-main.2} } ``` - [← previous A Transparent Model of Syntactic and Semantic Cue-based Retrieval](https://md.research.elamur.ai/free/en/papers/yue-2026-transparent-model-syntactic-semantic-cue/) - [next → Abjad-Kids: An Arabic Speech Classification Dataset for Primary Education](https://md.research.elamur.ai/free/en/papers/snoubara-2026-abjad-kids-arabic-speech-classification/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/3/) ## References 9 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Frank Benford. 1938. The law of anomalous numbers. Proceedings of the American Mathematical Society. [find](https://scholar.google.com/scholar?q=Frank+Benford%2E+1938%2E+The+law+of+anomalous+numbers%2E+Proceedings+of+the+American+Mathematical+Society%2E) 2. Chameleon Team, Mingda Chen, Jacob Kahn, and Shang-Wen Li. 2024. Chameleon: Mixed-modal early-fusion foundation models. ArXiv, abs/2405.09818. [our card](https://md.research.elamur.ai/free/en/papers/team-2024-chameleon-mixed-modal-early-fusion/) 3. James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † Wesam-Manassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. Ping-Rui Tsai, Chi hsiang Wang, Yu-Che… — [arXiv:2501.18620](https://arxiv.org/abs/2501.18620) [find](https://scholar.google.com/scholar?q=James+Betker%2C+Gabriel+Goh%2C+Li+Jing%2C+%E2%80%A0+TimBrooks%2C+Jianfeng+Wang%2C+Linjie+Li%2C+%E2%80%A0+LongOuyang%2C+%E2%80%A0+JuntangZhuang%2C+%E2%80%A0+JoyceLee%2C+%E2%80%A0+YufeiGuo%2C+%E2%80%A0+Wesam%2DManassra%2C+%E2%80%A0+PrafullaDhariwal%2C+%E2%80%A0+CaseyChu%2C+%E2%80%A0+YunxinJiao%2C+and+Ad) 4. George Kingsley Zipf. 1932. Selected Studies of the Principle of Relative Frequency in Language. Harvard University Press, Cambridge, MA. [find](https://scholar.google.com/scholar?q=George+Kingsley+Zipf%2E+1932%2E+Selected+Studies+of+the+Principle+of+Relative+Frequency+in+Language%2E+Harvard+University+Press%2C+Cambridge%2C+MA%2E) 5. David M. Chan, Rodolfo Corona, Joonyong Park, Cheol Jun Cho, Yutong Bai, and Trevor Darrell. 2024. Analyzing the language of visual tokens. ArXiv, abs/2411.05001. — [arXiv:2411.05001](https://arxiv.org/abs/2411.05001) [find](https://scholar.google.com/scholar?q=David+M%2E+Chan%2C+Rodolfo+Corona%2C+Joonyong+Park%2C+Cheol+Jun+Cho%2C+Yutong+Bai%2C+and+Trevor+Darrell%2E+2024%2E+Analyzing+the+language+of+visual+tokens%2E+ArXiv%2C+abs%2F2411%2E05001%2E+%E2%80%94+%5BarXiv%3A2411%2E05001%5D%28https%3A%2F%2Farxiv%2Eor) 6. Michael Crosier and Lewis D. Griffin. 2007. Zipf’s law in image coding schemes. In British Machine Vision Conference. [find](https://scholar.google.com/scholar?q=Michael+Crosier+and+Lewis+D%2E+Griffin%2E+2007%2E+Zipf%E2%80%99s+law+in+image+coding+schemes%2E+In+British+Machine+Vision+Conference%2E) 7. Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. 2024. Scaling rectified flow transformers for high-resolu… — [arXiv:2403.03206](https://arxiv.org/abs/2403.03206) [find](https://scholar.google.com/scholar?q=Patrick+Esser%2C+Sumith+Kulal%2C+A%2E+Blattmann%2C+Rahim+Entezari%2C+Jonas+Muller%2C+Harry+Saini%2C+Yam+Levi%2C+Dominik+Lorenz%2C+Axel+Sauer%2C+Frederic+Boesel%2C+Dustin+Podell%2C+Tim+Dockhorn%2C+Zion+English%2C+Kyle+Lacey%2C+Alex) 8. H. S. Heaps. 1978. Information retrieval, computational and theoretical aspects. [find](https://scholar.google.com/scholar?q=H%2E+S%2E+Heaps%2E+1978%2E+Information+retrieval%2C+computational+and+theoretical+aspects%2E) 9. Black Forest Labs, Stephen Batifol, A. Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Muller, Dustin Podell, [find](https://scholar.google.com/scholar?q=Black+Forest+Labs%2C+Stephen+Batifol%2C+A%2E+Blattmann%2C+Frederic+Boesel%2C+Saksham+Consul%2C+Cyril+Diagne%2C+Tim+Dockhorn%2C+Jack+English%2C+Zion+English%2C+Patrick+Esser%2C+Sumith+Kulal%2C+Kyle+Lacey%2C+Yam+Levi%2C+Cheng+Li%2C+)