Hugging Face Daily Papers · · 5 min read

AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss

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

Fréchet distance has recently emerged as an effective distribution-level objective for generator post-training. However, directly optimizing it can cause Fréchet hacking: the target metric keeps improving while visual quality and Fréchet alignment in other feature spaces stagnate or deteriorate. We attribute this to the static pretrained feature spaces used by existing Fréchet losses, which give incomplete and fixed views of the gap between real and generated distributions. We propose Adversarial Fréchet Distance (AdvFD), which complements those static targets with a learnable representation that adversarially maximizes the Fréchet discrepancy while the generator minimizes it in the resulting adaptive space. To stop the representation from inflating the objective through feature amplification, we add real-feature whitening, normalizing its scale and covariance geometry to stabilize the min–max optimization. AdvFD consistently improves one-step generator post-training across JiT and pMF backbones and model scales.</p>\n","updatedAt":"2026-08-12T05:19:38.586Z","author":{"_id":"6419e0b3ed725fef6444f53a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6419e0b3ed725fef6444f53a/3ZzIafBJa50-gAu1PASuG.png","fullname":"Yu Changqian","name":"Changqian","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":6,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8966907858848572},"editors":["Changqian"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/6419e0b3ed725fef6444f53a/3ZzIafBJa50-gAu1PASuG.png"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.11205","authors":[{"_id":"6a7be5901653ef87c6af1c5c","user":{"_id":"6721ab11866d7d87d1798cd1","avatarUrl":"/avatars/86b40cb8259f148a94a70a4cac268775.svg","isPro":false,"fullname":"Mingju Gao","user":"GasaiYU","type":"user","name":"GasaiYU"},"name":"Mingju Gao","status":"claimed_verified","statusLastChangedAt":"2026-08-12T08:45:05.106Z","hidden":false},{"_id":"6a7be5901653ef87c6af1c5d","user":{"_id":"6434caa64b34368fdb07da48","avatarUrl":"/avatars/8da3d3d1e274ff4ad409234678b1b952.svg","isPro":false,"fullname":"Jingkai Zhou","user":"theFoxofSky","type":"user","name":"theFoxofSky"},"name":"Jingkai Zhou","status":"claimed_verified","statusLastChangedAt":"2026-08-12T08:45:05.113Z","hidden":false},{"_id":"6a7be5901653ef87c6af1c5e","name":"Kun Gai","hidden":false},{"_id":"6a7be5901653ef87c6af1c5f","name":"Changqian Yu","hidden":false},{"_id":"6a7be5901653ef87c6af1c60","name":"Hao Tang","hidden":false}],"publishedAt":"2026-08-11T00:00:00.000Z","submittedOnDailyAt":"2026-08-12T00:00:00.000Z","title":"AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss","submittedOnDailyBy":{"_id":"6419e0b3ed725fef6444f53a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6419e0b3ed725fef6444f53a/3ZzIafBJa50-gAu1PASuG.png","isPro":false,"fullname":"Yu Changqian","user":"Changqian","type":"user","name":"Changqian"},"summary":"Fréchet distance has recently emerged as an effective distribution-level objective for generator post-training, complementing the conventional sample-level diffusion and flow-matching losses. However, directly optimizing Fréchet objectives can cause Fréchet hacking. The target metrics keep improving, but visual quality and Fréchet alignment in other feature spaces may stagnate or deteriorate. We attribute this failure to the static pretrained feature spaces used by existing Fréchet losses. These feature spaces provide incomplete and fixed views of the differences between real and generated distributions. To address this limitation, we propose Adversarial Fréchet Distance (AdvFD), which complements the static representation targets in FD-Loss with a calibrated adversarially learned representation. AdvFD augments the original static Fréchet objective with a learnable representation that adversarially maximizes the Fréchet discrepancy between real and generated samples, while the generator minimizes the same discrepancy in the resulting adaptive feature space. To prevent the adversarial representation from trivially increasing the objective through feature amplification, we further introduce real-feature whitening, which normalizes its scale and covariance geometry and stabilizes the min--max optimization. Extensive experiments show that AdvFD consistently improves one-step generator post-training across both JiT and pMF backbones and across different model scales.","upvotes":17,"discussionId":"6a7be5901653ef87c6af1c61","projectPage":"https://gasaiyu.github.io/AdvFD-page/","ai_summary":"Adversarial Fréchet Distance improves generator post-training by adding a learnable adversarial feature space to static Fréchet losses, with whitening to stabilize optimization.","ai_keywords":["Fréchet distance","generator post-training","diffusion loss","flow-matching loss","Fréchet hacking","pretrained feature spaces","Adversarial Fréchet Distance","AdvFD","adversarially learned representation","real-feature whitening","min-max optimization","one-step generator"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"665f02ce9f9e5b38d0a256a8","name":"Kwai-Kolors","fullname":"Kolors Team, Kuaishou Technology","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/62f0babaef9cc6810cec02ff/sVnELkcfVo5kxg5308rkr.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6419e0b3ed725fef6444f53a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6419e0b3ed725fef6444f53a/3ZzIafBJa50-gAu1PASuG.png","isPro":false,"fullname":"Yu Changqian","user":"Changqian","type":"user"},{"_id":"66221f7a608fb4e7f46ecadb","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/yMCsWyTkL4Xdxe18uwMYR.jpeg","isPro":false,"fullname":"Tianlin Pan","user":"ldiex","type":"user"},{"_id":"66f25defd295fc5ccc9bafc5","avatarUrl":"/avatars/860e1062b26a8f605c469ffbc6aca86c.svg","isPro":false,"fullname":"ziheng zhang","user":"Ruabert","type":"user"},{"_id":"649e83af47b275ac48466218","avatarUrl":"/avatars/5ce406848e467afc50ded4e8f7df5970.svg","isPro":false,"fullname":"dong she","user":"sd0809","type":"user"},{"_id":"69ef699d5e0c40377155eae4","avatarUrl":"/avatars/4aa528a1d404e927a9d60da621fd8bc7.svg","isPro":false,"fullname":"Mike G","user":"GasaiYuno71","type":"user"},{"_id":"6434caa64b34368fdb07da48","avatarUrl":"/avatars/8da3d3d1e274ff4ad409234678b1b952.svg","isPro":false,"fullname":"Jingkai Zhou","user":"theFoxofSky","type":"user"},{"_id":"632322c1f89691844a2024b4","avatarUrl":"/avatars/c3d9a4d92abc1ae3a8cc0572ffe0d8d1.svg","isPro":false,"fullname":"Fei Liu","user":"fliu","type":"user"},{"_id":"64531f631a57e1179c203e6b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64531f631a57e1179c203e6b/C_J7pXFLqoJoHYPPhK3J9.jpeg","isPro":false,"fullname":"zjn","user":"garlicisnotmyfavor","type":"user"},{"_id":"63733e12c1142685da81d167","avatarUrl":"/avatars/20a1d0a0cc8ed180a5609c6fe7fe51a1.svg","isPro":false,"fullname":"Yu Xu","user":"YUXU915","type":"user"},{"_id":"6635b673e4028f5fd02b2131","avatarUrl":"/avatars/8c18cb14954fd2cd35f6975118471fba.svg","isPro":false,"fullname":"YunhongLu","user":"JaydenLu666","type":"user"},{"_id":"66163a086dff9352ba4ec91b","avatarUrl":"/avatars/235e261e36549223e45b5750ae0860bd.svg","isPro":false,"fullname":"Xu Zhang","user":"xushu-me","type":"user"},{"_id":"69bcf8c79267623e872fd2d0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/1HRRWqZf1vPw61hjYajMq.png","isPro":false,"fullname":"山崎結衣","user":"owenkingvn","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"665f02ce9f9e5b38d0a256a8","name":"Kwai-Kolors","fullname":"Kolors Team, Kuaishou Technology","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/62f0babaef9cc6810cec02ff/sVnELkcfVo5kxg5308rkr.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.11205.md","query":{}}">
Papers
arxiv:2608.11205

AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss

Published on Aug 11
· Submitted by
Yu Changqian
on Aug 12
Authors:

Abstract

Adversarial Fréchet Distance improves generator post-training by adding a learnable adversarial feature space to static Fréchet losses, with whitening to stabilize optimization.

Fréchet distance has recently emerged as an effective distribution-level objective for generator post-training, complementing the conventional sample-level diffusion and flow-matching losses. However, directly optimizing Fréchet objectives can cause Fréchet hacking. The target metrics keep improving, but visual quality and Fréchet alignment in other feature spaces may stagnate or deteriorate. We attribute this failure to the static pretrained feature spaces used by existing Fréchet losses. These feature spaces provide incomplete and fixed views of the differences between real and generated distributions. To address this limitation, we propose Adversarial Fréchet Distance (AdvFD), which complements the static representation targets in FD-Loss with a calibrated adversarially learned representation. AdvFD augments the original static Fréchet objective with a learnable representation that adversarially maximizes the Fréchet discrepancy between real and generated samples, while the generator minimizes the same discrepancy in the resulting adaptive feature space. To prevent the adversarial representation from trivially increasing the objective through feature amplification, we further introduce real-feature whitening, which normalizes its scale and covariance geometry and stabilizes the min--max optimization. Extensive experiments show that AdvFD consistently improves one-step generator post-training across both JiT and pMF backbones and across different model scales.

Community

Paper submitter about 15 hours ago

Fréchet distance has recently emerged as an effective distribution-level objective for generator post-training. However, directly optimizing it can cause Fréchet hacking: the target metric keeps improving while visual quality and Fréchet alignment in other feature spaces stagnate or deteriorate. We attribute this to the static pretrained feature spaces used by existing Fréchet losses, which give incomplete and fixed views of the gap between real and generated distributions. We propose Adversarial Fréchet Distance (AdvFD), which complements those static targets with a learnable representation that adversarially maximizes the Fréchet discrepancy while the generator minimizes it in the resulting adaptive space. To stop the representation from inflating the objective through feature amplification, we add real-feature whitening, normalizing its scale and covariance geometry to stabilize the min–max optimization. AdvFD consistently improves one-step generator post-training across JiT and pMF backbones and model scales.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images

· Sign up or log in to comment

Get this paper in your agent:

hf papers read 2608.11205
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper

No model linking this paper

Cite arxiv.org/abs/2608.11205 in a model README.md to link it from this page.

Datasets citing this paper

No dataset linking this paper

Cite arxiv.org/abs/2608.11205 in a dataset README.md to link it from this page.

Spaces citing this paper

No Space linking this paper

Cite arxiv.org/abs/2608.11205 in a Space README.md to link it from this page.

Collections including this paper

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.

More from Hugging Face Daily Papers