We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find that the converged diffusion loss scales with the amount of structured language in the prompt. To quantify structured language, we adapt two complementary measures: a white-box likelihood metric (GPG) and a black-box attribute metric (ED). Across controlled training runs, the converged diffusion loss decreases approximately linearly with GPG and follows a power law with ED. Guided by these scaling properties, we improve \\emph{diffusability} by constructing structured prompts with semantic and geometric annotations derived from images, and improve \\emph{promptability} by training a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. The resulting system outperforms all evaluated open-weight models on nearly every compositional, reasoning, and world-knowledge benchmark, while matching or surpassing the strongest closed-weight models on most evaluations.</p>\n","updatedAt":"2026-08-03T01:53:25.785Z","author":{"_id":"644e3e5f030210812f413073","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/uW8TKV2sds97lFBwnt6JK.jpeg","fullname":"Zilong Chen","name":"heheyas","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":14,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.894237220287323},"editors":["heheyas"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/uW8TKV2sds97lFBwnt6JK.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.29679","authors":[{"_id":"6a6ff463bbe824e6bcc465b2","user":{"_id":"644e3e5f030210812f413073","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/uW8TKV2sds97lFBwnt6JK.jpeg","isPro":true,"fullname":"Zilong Chen","user":"heheyas","type":"user","name":"heheyas"},"name":"Zilong Chen","status":"claimed_verified","statusLastChangedAt":"2026-08-03T07:54:50.270Z","hidden":false},{"_id":"6a6ff463bbe824e6bcc465b3","name":"Chaorui Deng","hidden":false},{"_id":"6a6ff463bbe824e6bcc465b4","name":"Kunchang Li","hidden":false},{"_id":"6a6ff463bbe824e6bcc465b5","name":"Hongyi Yuan","hidden":false},{"_id":"6a6ff463bbe824e6bcc465b6","name":"Haoqi Fan","hidden":false}],"publishedAt":"2026-07-31T00:00:00.000Z","submittedOnDailyAt":"2026-08-03T00:00:00.000Z","title":"Scaling Properties of Text Conditioning in Visual Generation","submittedOnDailyBy":{"_id":"644e3e5f030210812f413073","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/uW8TKV2sds97lFBwnt6JK.jpeg","isPro":true,"fullname":"Zilong Chen","user":"heheyas","type":"user","name":"heheyas"},"summary":"We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find that the converged diffusion loss scales with the amount of structured language in the prompt. To quantify structured language, we adapt two complementary measures: a white-box likelihood metric (GPG) and a black-box attribute metric (ED). Across controlled training runs, the converged diffusion loss decreases approximately linearly with GPG and follows a power law with ED. Guided by these scaling properties, we improve diffusability by constructing structured prompts with semantic and geometric annotations derived from images, and improve promptability by training a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. The resulting system outperforms all evaluated open-weight models on nearly every compositional, reasoning, and world-knowledge benchmark, while matching or surpassing the strongest closed-weight models on most evaluations.","upvotes":21,"discussionId":"6a6ff463bbe824e6bcc465b7","projectPage":"https://heheyas.github.io/context-scaling/","githubRepo":"https://github.com/heheyas/context-scaling","githubRepoAddedBy":"user","githubStars":2},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"644e3e5f030210812f413073","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/uW8TKV2sds97lFBwnt6JK.jpeg","isPro":true,"fullname":"Zilong Chen","user":"heheyas","type":"user"},{"_id":"6039478ab3ecf716b1a5fd4d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6039478ab3ecf716b1a5fd4d/_Thy4E7taiSYBLKxEKJbT.jpeg","isPro":true,"fullname":"taesiri","user":"taesiri","type":"user"},{"_id":"6289e290edfa7a816db76774","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1653203591668-noauth.png","isPro":false,"fullname":"Jack","user":"Jack9585","type":"user"},{"_id":"61fb81006374891646732f37","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1643872995181-61fb81006374891646732f37.jpeg","isPro":false,"fullname":"Kunchang Li","user":"Andy1621","type":"user"},{"_id":"676a15615b3b2c878d650119","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/TmjcPMui1YagND8N5-lD7.png","isPro":false,"fullname":"Yusen Feng","user":"wangchek","type":"user"},{"_id":"6a6a8260977fbfce4bac218d","avatarUrl":"/avatars/884d72707d0a24d2d62b991428d4fa4d.svg","isPro":false,"fullname":"Susan Thompson","user":"william-2151845","type":"user"},{"_id":"6a6a9a26f10c30deda35165d","avatarUrl":"/avatars/52a78b5162cb96a9d2406bb63193b52e.svg","isPro":false,"fullname":"Michael Johnson","user":"ronald-5751533","type":"user"},{"_id":"6a6c7b54a669192d904255ea","avatarUrl":"/avatars/0f993843161ee4f84ce23522227f31a3.svg","isPro":false,"fullname":"James Taylor","user":"albert-1188287","type":"user"},{"_id":"6a6c7c751964ab520526857c","avatarUrl":"/avatars/0e4c180ab58ef319875181a57bf82239.svg","isPro":false,"fullname":"William Wilson","user":"dennis-3943649","type":"user"},{"_id":"6a6c84bde34d1f023f2bd54f","avatarUrl":"/avatars/c415b370b627c981d14105491507b682.svg","isPro":false,"fullname":"Steven Moore","user":"louis-1513015","type":"user"},{"_id":"6a6d3d85ef16fe7cceaf0fad","avatarUrl":"/avatars/645ae134f15948f3a368fdaff4ce5308.svg","isPro":false,"fullname":"Patricia Perez","user":"william-1130851","type":"user"},{"_id":"6a6aa148ad5a6f2f63620596","avatarUrl":"/avatars/11497c4cb1c583ebe22f266fa8585e92.svg","isPro":false,"fullname":"David Williams","user":"timothy-4448888","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"query":{}}">
Scaling Properties of Text Conditioning in Visual Generation
Abstract
We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find that the converged diffusion loss scales with the amount of structured language in the prompt. To quantify structured language, we adapt two complementary measures: a white-box likelihood metric (GPG) and a black-box attribute metric (ED). Across controlled training runs, the converged diffusion loss decreases approximately linearly with GPG and follows a power law with ED. Guided by these scaling properties, we improve diffusability by constructing structured prompts with semantic and geometric annotations derived from images, and improve promptability by training a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. The resulting system outperforms all evaluated open-weight models on nearly every compositional, reasoning, and world-knowledge benchmark, while matching or surpassing the strongest closed-weight models on most evaluations.
Community
We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find that the converged diffusion loss scales with the amount of structured language in the prompt. To quantify structured language, we adapt two complementary measures: a white-box likelihood metric (GPG) and a black-box attribute metric (ED). Across controlled training runs, the converged diffusion loss decreases approximately linearly with GPG and follows a power law with ED. Guided by these scaling properties, we improve \emph{diffusability} by constructing structured prompts with semantic and geometric annotations derived from images, and improve \emph{promptability} by training a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. The resulting system outperforms all evaluated open-weight models on nearly every compositional, reasoning, and world-knowledge benchmark, while matching or surpassing the strongest closed-weight models on most evaluations.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2607.29679 in a model README.md to link it from this page.
Cite arxiv.org/abs/2607.29679 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2607.29679 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.