This is the initial version of JLT. We will update it with complete experiments and results analysis and then revise it.</p>\n","updatedAt":"2026-05-27T14:47:12.563Z","author":{"_id":"65f1798a3e29a622f098a42d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65f1798a3e29a622f098a42d/IraNhB8kOIa2dSRKMJN3q.png","fullname":"Guanyu Zhou","name":"TheMartyr","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9234577417373657},"editors":["TheMartyr"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/65f1798a3e29a622f098a42d/IraNhB8kOIa2dSRKMJN3q.png"],"reactions":[{"reaction":"🔥","users":["R3DeK","Mikufanssss","koujiaxin","yehogwon"],"count":4}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.27102","authors":[{"_id":"6a170341991d34bf203501e8","name":"Funing Fu","hidden":false},{"_id":"6a170341991d34bf203501e9","name":"Tenghui Wang","hidden":false},{"_id":"6a170341991d34bf203501ea","name":"Junyong Cen","hidden":false},{"_id":"6a170341991d34bf203501eb","name":"Qichao Zhu","hidden":false},{"_id":"6a170341991d34bf203501ec","name":"Guanyu Zhou","hidden":false}],"publishedAt":"2026-05-26T00:00:00.000Z","submittedOnDailyAt":"2026-05-27T00:00:00.000Z","title":"JLT: Clean-Latent Prediction in Latent Diffusion Transformers","submittedOnDailyBy":{"_id":"65f1798a3e29a622f098a42d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65f1798a3e29a622f098a42d/IraNhB8kOIa2dSRKMJN3q.png","isPro":false,"fullname":"Guanyu Zhou","user":"TheMartyr","type":"user","name":"TheMartyr"},"summary":"Flow matching with clean-data prediction has shown that regressing the clean point can exploit low-dimensional structure more effectively than predicting an ambient noised quantity. We ask whether this principle remains useful after images are mapped into a learned latent space, where compression has already removed much of the raw pixel variability. We introduce JLT, a 130M latent diffusion Transformer over frozen FLUX.2 VAE codes, and compare clean-latent prediction with a matched velocity-prediction DiT under the same representation, backbone, and training settings. Although the three variables x, epsilon, and v are linearly convertible for a fixed corruption time, a local Gaussian analysis shows that velocity regression inherits an isotropic target-covariance floor and amplifies low-variance latent directions, while clean prediction damps them. On ImageNet 256 x 256, JLT-B/1 obtains FID-50K 2.50 with classifier-free guidance, with a large matched-target gap over velocity prediction. These results suggest that prediction targets in latent diffusion are representation-dependent geometric choices, rather than interchangeable algebraic parameterizations.","upvotes":12,"discussionId":"6a170341991d34bf203501ed","projectPage":"https://github.com/akatsuki-neo/JLT","ai_summary":"Latent diffusion models using clean-data prediction outperform velocity prediction in compressed representations, demonstrating that prediction targets are geometrically dependent rather than algebraically interchangeable.","ai_keywords":["flow matching","clean-data prediction","latent space","diffusion models","velocity prediction","latent diffusion Transformer","FLUX.2 VAE","DiT","classifier-free guidance","FID-50K","Gaussian analysis","isotropic target-covariance floor"],"organization":{"_id":"69d903c090fc614ce4f7eb4b","name":"dawn-neo","fullname":" akatsuki-neo","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6747e1b1b48ed96c84b58830/3VGpT07W73b0tw8Jyyofo.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"65f1798a3e29a622f098a42d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65f1798a3e29a622f098a42d/IraNhB8kOIa2dSRKMJN3q.png","isPro":false,"fullname":"Guanyu Zhou","user":"TheMartyr","type":"user"},{"_id":"6747e1b1b48ed96c84b58830","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6747e1b1b48ed96c84b58830/NqhcY1_GvL3GrjblNGDek.jpeg","isPro":false,"fullname":"spawner","user":"spawner1145","type":"user"},{"_id":"6964fbc2c4eb6cb0562e41b6","avatarUrl":"/avatars/e713c513082be1e405d684fe9ee07875.svg","isPro":false,"fullname":"123","user":"R3DeK","type":"user"},{"_id":"667027c822073d70a1438e4d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/667027c822073d70a1438e4d/YRHqaKjERmXadIAbrBQ92.png","isPro":false,"fullname":"Displace_Asher","user":"Mikufanssss","type":"user"},{"_id":"65c4c99309672feb8ce0c883","avatarUrl":"/avatars/7444c1083dae17d671a09191d2400992.svg","isPro":false,"fullname":"xia","user":"xianienie","type":"user"},{"_id":"63bd1c420608fbc76b84049f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63bd1c420608fbc76b84049f/52JW0XtSHRrK7vm29Jv71.jpeg","isPro":false,"fullname":"Samwise Wang","user":"tzwm","type":"user"},{"_id":"637f08ab213876d4122a974c","avatarUrl":"/avatars/22f0115d1654d6fb8265d38e8629aa9c.svg","isPro":false,"fullname":"Jiaxin Kou","user":"koujiaxin","type":"user"},{"_id":"67053ebeb077981e97fe43c0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67053ebeb077981e97fe43c0/aRXze1Md2_hu93vVAdrMr.jpeg","isPro":false,"fullname":"linker","user":"whitelinker","type":"user"},{"_id":"68d40b135ae7555127aaac7f","avatarUrl":"/avatars/6e3cd7b39db1c12e8a249ee15af5a4a1.svg","isPro":false,"fullname":"starrystelle","user":"starrystelle","type":"user"},{"_id":"6430fffacd31d174a9fa96e4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6430fffacd31d174a9fa96e4/9ZnVcC9FOxCS8rNDr5Qwa.png","isPro":false,"fullname":"Nebulae","user":"NebulaeWis","type":"user"},{"_id":"69ccc52aaf6645491d17e317","avatarUrl":"/avatars/2ed46a892f982f96d82160b1d25696f2.svg","isPro":false,"fullname":"Yifan Ma","user":"chernandezzw","type":"user"},{"_id":"62e551b42277e8483be1c71c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1659195792493-noauth.jpeg","isPro":false,"fullname":"Yeho Gwon","user":"yehogwon","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"69d903c090fc614ce4f7eb4b","name":"dawn-neo","fullname":" akatsuki-neo","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6747e1b1b48ed96c84b58830/3VGpT07W73b0tw8Jyyofo.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.27102.md"}">
JLT: Clean-Latent Prediction in Latent Diffusion Transformers
Abstract
Latent diffusion models using clean-data prediction outperform velocity prediction in compressed representations, demonstrating that prediction targets are geometrically dependent rather than algebraically interchangeable.
AI-generated summary
Flow matching with clean-data prediction has shown that regressing the clean point can exploit low-dimensional structure more effectively than predicting an ambient noised quantity. We ask whether this principle remains useful after images are mapped into a learned latent space, where compression has already removed much of the raw pixel variability. We introduce JLT, a 130M latent diffusion Transformer over frozen FLUX.2 VAE codes, and compare clean-latent prediction with a matched velocity-prediction DiT under the same representation, backbone, and training settings. Although the three variables x, epsilon, and v are linearly convertible for a fixed corruption time, a local Gaussian analysis shows that velocity regression inherits an isotropic target-covariance floor and amplifies low-variance latent directions, while clean prediction damps them. On ImageNet 256 x 256, JLT-B/1 obtains FID-50K 2.50 with classifier-free guidance, with a large matched-target gap over velocity prediction. These results suggest that prediction targets in latent diffusion are representation-dependent geometric choices, rather than interchangeable algebraic parameterizations.
Community
This is the initial version of JLT. We will update it with complete experiments and results analysis and then revise it.
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/2605.27102 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.27102 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.27102 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.