Hugging Face Daily Papers · May 18, 2026 · 6 min read

InsightTok: Improving Text and Face Fidelity in Discrete Tokenization for Autoregressive Image Generation

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

InsightTok is a discrete visual tokenizer designed to improve the fidelity of text and faces, two of the most challenging yet perceptually important structures in autoregressive image generation.\nExisting visual tokenizers are typically trained with generic reconstruction objectives, which do not explicitly prioritize these fidelity-critical regions. InsightTok addresses this limitation through localized, content-aware perceptual supervision, enabling substantially better preservation of textual content and facial details under a compact discrete bottleneck.\nHighlights:\n<ul>\n<li>State-of-the-art text and face reconstruction among discrete visual tokenizers at the same compression rate, using 16× downsampling and a compact 16,384-entry codebook</li>\n<li>Minimal additional training overhead over a vanilla VQGAN-style tokenizer</li>\n<li>No changes required to downstream generative modeling. Readily compatible with standard autoregressive image generation pipelines</li>\n<li>Tokenizer improvements transfer effectively to downstream text-to-image generation, yielding clearer text and more faithful facial details</li>\n</ul>\n","updatedAt":"2026-05-18T01:56:58.775Z","author":{"_id":"63f1d16fbe95ed4c9a9418fe","avatarUrl":"/avatars/a1bdfa97323693808f2f16ec74698ed3.svg","fullname":"Yang Yue","name":"yueyang2000","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":3,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.779451310634613},"editors":["yueyang2000"],"editorAvatarUrls":["/avatars/a1bdfa97323693808f2f16ec74698ed3.svg"],"reactions":[],"isReport":false}},{"id":"6a0bc0bcaac8191f85fc270b","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":357,"isUserFollowing":false},"createdAt":"2026-05-19T01:45:32.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"This is an automated message from the [Librarian Bot](https://huggingface.co/librarian-bots). I found the following papers similar to this paper. \n\nThe following papers were recommended by the Semantic Scholar API \n\n* [MacTok: Robust Continuous Tokenization for Image Generation](https://huggingface.co/papers/2603.29634) (2026)\n* [A2BFR: Attribute-Aware Blind Face Restoration](https://huggingface.co/papers/2603.29423) (2026)\n* [GramSR: Visual Feature Conditioning for Diffusion-Based Super-Resolution](https://huggingface.co/papers/2604.25457) (2026)\n* [Semantic-Aware Prefix Learning for Token-Efficient Image Generation](https://huggingface.co/papers/2603.25249) (2026)\n* [RefDecoder: Enhancing Visual Generation with Conditional Video Decoding](https://huggingface.co/papers/2605.15196) (2026)\n* [Generative Refinement Networks for Visual Synthesis](https://huggingface.co/papers/2604.13030) (2026)\n* [MMCORE: MultiModal COnnection with Representation Aligned Latent Embeddings](https://huggingface.co/papers/2604.19902) (2026)\n\n\n Please give a thumbs up to this comment if you found it helpful!\n\n If you want recommendations for any Paper on Hugging Face checkout [this](https://huggingface.co/spaces/librarian-bots/recommend_similar_papers) Space\n\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: `@librarian-bot recommend`","html":"This is an automated message from the <a href=\"https://huggingface.co/librarian-bots\">Librarian Bot</a>. I found the following papers similar to this paper. \nThe following papers were recommended by the Semantic Scholar API \n<ul>\n<li><a href=\"https://huggingface.co/papers/2603.29634\">MacTok: Robust Continuous Tokenization for Image Generation</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2603.29423\">A2BFR: Attribute-Aware Blind Face Restoration</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2604.25457\">GramSR: Visual Feature Conditioning for Diffusion-Based Super-Resolution</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2603.25249\">Semantic-Aware Prefix Learning for Token-Efficient Image Generation</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2605.15196\">RefDecoder: Enhancing Visual Generation with Conditional Video Decoding</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2604.13030\">Generative Refinement Networks for Visual Synthesis</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2604.19902\">MMCORE: MultiModal COnnection with Representation Aligned Latent Embeddings</a> (2026)</li>\n</ul>\n Please give a thumbs up to this comment if you found it helpful!\n If you want recommendations for any Paper on Hugging Face checkout <a href=\"https://huggingface.co/spaces/librarian-bots/recommend_similar_papers\">this</a> Space\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: <code><a href=\"/librarian-bot\">@librarian-bot</a> recommend</code>\n","updatedAt":"2026-05-19T01:45:32.002Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":357,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6879501342773438},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.14333","authors":[{"_id":"6a06c06eb1a8cbabc9f09a74","user":{"_id":"63f1d16fbe95ed4c9a9418fe","avatarUrl":"/avatars/a1bdfa97323693808f2f16ec74698ed3.svg","isPro":false,"fullname":"Yang Yue","user":"yueyang2000","type":"user","name":"yueyang2000"},"name":"Yang Yue","status":"claimed_verified","statusLastChangedAt":"2026-05-18T09:44:09.362Z","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a75","name":"Fangyun Wei","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a76","name":"Tianyu He","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a77","name":"Jinjing Zhao","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a78","name":"Zanlin Ni","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a79","name":"Zeyu Liu","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a7a","name":"Jiayi Guo","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a7b","name":"Lei Shi","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a7c","name":"Yue Dong","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a7d","name":"Li Chen","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a7e","name":"Ji Li","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a7f","name":"Gao Huang","hidden":false},{"_id":"6a06c06eb1a8cbabc9f09a80","name":"Dong Chen","hidden":false}],"publishedAt":"2026-05-14T00:00:00.000Z","submittedOnDailyAt":"2026-05-18T00:00:00.000Z","title":"InsightTok: Improving Text and Face Fidelity in Discrete Tokenization for Autoregressive Image Generation","submittedOnDailyBy":{"_id":"63f1d16fbe95ed4c9a9418fe","avatarUrl":"/avatars/a1bdfa97323693808f2f16ec74698ed3.svg","isPro":false,"fullname":"Yang Yue","user":"yueyang2000","type":"user","name":"yueyang2000"},"summary":"Text and faces are among the most perceptually salient and practically important patterns in visual generation, yet they remain challenging for autoregressive generators built on discrete tokenization. A central bottleneck is the tokenizer: aggressive downsampling and quantization often discard the fine-grained structures needed to preserve readable glyphs and distinctive facial features. We attribute this gap to standard discrete-tokenizer objectives being weakly aligned with text legibility and facial fidelity, as these objectives typically optimize generic reconstruction while compressing diverse content uniformly. To address this, we propose InsightTok, a simple yet effective discrete visual tokenization framework that enhances text and face fidelity through localized, content-aware perceptual losses. With a compact 16k codebook and a 16x downsampling rate, InsightTok significantly outperforms prior tokenizers in text and face reconstruction without compromising general reconstruction quality. These gains consistently transfer to autoregressive image generation in InsightAR, producing images with clearer text and more faithful facial details. Overall, our results highlight the potential of specialized supervision in tokenizer training for advancing discrete image generation.","upvotes":31,"discussionId":"6a06c06fb1a8cbabc9f09a81","githubRepo":"https://github.com/LeapLabTHU/InsightTok","githubRepoAddedBy":"user","ai_summary":"InsightTok improves discrete visual tokenization for better text and face reconstruction through content-aware perceptual losses, enhancing autoregressive image generation quality.","ai_keywords":["discrete visual tokenization","autoregressive generators","tokenizer","discrete-tokenizer objectives","text legibility","facial fidelity","perceptual losses","codebook","downsampling","autoregressive image generation"],"githubStars":29,"organization":{"_id":"69719700e3846c07669d13ee","name":"Tsinghua-LeapLab","fullname":"Tsinghua-LeapLab","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/63987ffb2ceb55aabe0852f3/hflTWNTGxeJx83xNkYrDB.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66f0d2036a483077eed42bfb","avatarUrl":"/avatars/f7f3f726842c26b8e52c9bdd48774b8e.svg","isPro":false,"fullname":"Renping Zhou","user":"rpzhou","type":"user"},{"_id":"63f1d16fbe95ed4c9a9418fe","avatarUrl":"/avatars/a1bdfa97323693808f2f16ec74698ed3.svg","isPro":false,"fullname":"Yang Yue","user":"yueyang2000","type":"user"},{"_id":"647e83257f9ad5e44babe82a","avatarUrl":"/avatars/2d9593775c49856fe5dfa5bd23dfcda7.svg","isPro":false,"fullname":"yifan pu","user":"yifanpu001","type":"user"},{"_id":"688c5e3cc13bf6d45a41d834","avatarUrl":"/avatars/2a76025d16115c73eb0885b6bb68bdfe.svg","isPro":false,"fullname":"Liu","user":"Bei21","type":"user"},{"_id":"672b3682d40b55aa6f51041e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/qxRAQCawu_UR2BBpymBd6.png","isPro":false,"fullname":"chi bene chen","user":"BeneChen12138","type":"user"},{"_id":"649661df533c317c5c3a34af","avatarUrl":"/avatars/fedd6306422627f439439132d1629c32.svg","isPro":false,"fullname":"ZhiyangLiang","user":"ZhiyangLiang","type":"user"},{"_id":"64f7f119a92703ef65d9a717","avatarUrl":"/avatars/118524faab66cecba6d4da622034b44b.svg","isPro":false,"fullname":"Sirui Zhang","user":"zsr200901","type":"user"},{"_id":"683e9206b5052f5f8736e530","avatarUrl":"/avatars/e872b4877748b50ec606e0ebfa37040d.svg","isPro":false,"fullname":"Tianyu Li","user":"Litianyu2004","type":"user"},{"_id":"677921d46f370093aa0d26e4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/_lV-a6412BAICEBjrU1zM.png","isPro":false,"fullname":"LIU Zhening","user":"ZincL","type":"user"},{"_id":"64f8962bce75bb0fb50bdbdb","avatarUrl":"/avatars/c85537df848bda7ec92565f56cd32eed.svg","isPro":false,"fullname":"Jinjing Zhao","user":"Jinjing713","type":"user"},{"_id":"63987ffb2ceb55aabe0852f3","avatarUrl":"/avatars/343b796ff6b8906203904e8c620d7eb5.svg","isPro":false,"fullname":"Zanlin Ni","user":"nzl-thu","type":"user"},{"_id":"673f49bfeb9aded68bf73e38","avatarUrl":"/avatars/8bf8a28c5b4af354dea97d811d413362.svg","isPro":false,"fullname":"Teng Wang","user":"tengwang01","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"69719700e3846c07669d13ee","name":"Tsinghua-LeapLab","fullname":"Tsinghua-LeapLab","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/63987ffb2ceb55aabe0852f3/hflTWNTGxeJx83xNkYrDB.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.14333.md"}">

Papers

arxiv:2605.14333

InsightTok: Improving Text and Face Fidelity in Discrete Tokenization for Autoregressive Image Generation

Published on May 14

· Submitted by

Yang Yue on May 18

Tsinghua-LeapLab

Upvote

Authors:

Yang Yue ,

Abstract

InsightTok improves discrete visual tokenization for better text and face reconstruction through content-aware perceptual losses, enhancing autoregressive image generation quality.

AI-generated summary

Text and faces are among the most perceptually salient and practically important patterns in visual generation, yet they remain challenging for autoregressive generators built on discrete tokenization. A central bottleneck is the tokenizer: aggressive downsampling and quantization often discard the fine-grained structures needed to preserve readable glyphs and distinctive facial features. We attribute this gap to standard discrete-tokenizer objectives being weakly aligned with text legibility and facial fidelity, as these objectives typically optimize generic reconstruction while compressing diverse content uniformly. To address this, we propose InsightTok, a simple yet effective discrete visual tokenization framework that enhances text and face fidelity through localized, content-aware perceptual losses. With a compact 16k codebook and a 16x downsampling rate, InsightTok significantly outperforms prior tokenizers in text and face reconstruction without compromising general reconstruction quality. These gains consistently transfer to autoregressive image generation in InsightAR, producing images with clearer text and more faithful facial details. Overall, our results highlight the potential of specialized supervision in tokenizer training for advancing discrete image generation.

View arXiv page View PDF GitHub 29 Add to collection

Community

yueyang2000

Paper author Paper submitter 1 day ago

InsightTok is a discrete visual tokenizer designed to improve the fidelity of text and faces, two of the most challenging yet perceptually important structures in autoregressive image generation.

Existing visual tokenizers are typically trained with generic reconstruction objectives, which do not explicitly prioritize these fidelity-critical regions. InsightTok addresses this limitation through localized, content-aware perceptual supervision, enabling substantially better preservation of textual content and facial details under a compact discrete bottleneck.

Highlights:

State-of-the-art text and face reconstruction among discrete visual tokenizers at the same compression rate, using 16× downsampling and a compact 16,384-entry codebook
Minimal additional training overhead over a vanilla VQGAN-style tokenizer
No changes required to downstream generative modeling. Readily compatible with standard autoregressive image generation pipelines
Tokenizer improvements transfer effectively to downstream text-to-image generation, yielding clearer text and more faithful facial details