Modern generative models possess a deep understanding of visual content, yet training them for image editing typically requires massive datasets of paired examples. This limits scalability, especially for video editing where collecting paired data is prohibitively expensive. We propose Bootstrap Your Generator (ByG), a general framework for unpaired training of flow matching editing models. It leverages the base model's knowledge without any external signal. Our approach pairs instruction-following cues extracted from the frozen model with cycle-consistency for structure preservation. To make this tractable, we propose to route gradients from downstream losses over clean predictions to noisy training states. We demonstrate state-of-the-art results on challenging data-scarce image and video editing scenarios. Extensive evaluations and user studies show that our method effectively generalizes to unseen domains and outperforms supervised baselines trained on millions of samples. Analysis reveals that our gradient routing bridges the train-inference gap, and extracting semantic cues from a base model provides a robust training signal that obviates the need for external reward models.</p>\n","updatedAt":"2026-06-03T14:03:03.426Z","author":{"_id":"6201838f305742efd5a6a0e3","avatarUrl":"/avatars/4e47ab9ef1a4b9f59756a6d8c90a970f.svg","fullname":"Yoad Tewel","name":"YoadTew","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.850054144859314},"editors":["YoadTew"],"editorAvatarUrls":["/avatars/4e47ab9ef1a4b9f59756a6d8c90a970f.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.03911","authors":[{"_id":"6a202bb915100c5272a8439f","name":"Yoad Tewel","hidden":false},{"_id":"6a202bb915100c5272a843a0","name":"Yuval Atzmon","hidden":false},{"_id":"6a202bb915100c5272a843a1","name":"Gal Chechik","hidden":false},{"_id":"6a202bb915100c5272a843a2","name":"Lior Wolf","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/6201838f305742efd5a6a0e3/OjmVnfvoRScumhpRtJWJG.png"],"publishedAt":"2026-06-02T00:00:00.000Z","submittedOnDailyAt":"2026-06-03T00:00:00.000Z","title":"Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching","submittedOnDailyBy":{"_id":"6201838f305742efd5a6a0e3","avatarUrl":"/avatars/4e47ab9ef1a4b9f59756a6d8c90a970f.svg","isPro":false,"fullname":"Yoad Tewel","user":"YoadTew","type":"user","name":"YoadTew"},"summary":"Modern generative models possess a deep understanding of visual content, yet training them for image editing typically requires massive datasets of paired examples. This limits scalability, especially for video editing where collecting paired data is prohibitively expensive. We propose Bootstrap Your Generator (ByG), a general framework for unpaired training of flow matching editing models. It leverages the base model's knowledge without any external signal. Our approach pairs instruction-following cues extracted from the frozen model with cycle-consistency for structure preservation. To make this tractable, we propose to route gradients from downstream losses over clean predictions to noisy training states. We demonstrate state-of-the-art results on challenging data-scarce image and video editing scenarios. Extensive evaluations and user studies show that our method effectively generalizes to unseen domains and outperforms supervised baselines trained on millions of samples. Analysis reveals that our gradient routing bridges the train-inference gap, and extracting semantic cues from a base model provides a robust training signal that obviates the need for external reward models.","upvotes":10,"discussionId":"6a202bb915100c5272a843a3","projectPage":"https://research.nvidia.com/labs/par/byg/","ai_summary":"Bootstrap Your Generator framework enables unpaired training of flow matching editing models by leveraging base model knowledge and gradient routing for improved generalization in data-scarce scenarios.","ai_keywords":["flow matching","unpaired training","gradient routing","cycle-consistency","instruction-following cues","downstream losses","train-inference gap","semantic cues","base model"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","organization":{"_id":"60262b67268c201cdc8b7d43","name":"nvidia","fullname":"NVIDIA","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/65df9200dc3292a8983e5017/Vs5FPVCH-VZBipV3qKTuy.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66704bc8a89afd375945eeb9","avatarUrl":"/avatars/985f77b67425211384b07a76aead5c9c.svg","isPro":false,"fullname":"y t","user":"yttau","type":"user"},{"_id":"6201838f305742efd5a6a0e3","avatarUrl":"/avatars/4e47ab9ef1a4b9f59756a6d8c90a970f.svg","isPro":false,"fullname":"Yoad Tewel","user":"YoadTew","type":"user"},{"_id":"66c60f676a741ca5b12ca4e8","avatarUrl":"/avatars/e2ca6c603287bf40b9312f2c1182d980.svg","isPro":false,"fullname":"Yoad Tewel","user":"ytewel","type":"user"},{"_id":"65376feed325b3f02fb92c69","avatarUrl":"/avatars/e952918cf434d5302e9b1a404eccaf0e.svg","isPro":false,"fullname":"Itamar Zimerman","user":"ItamarZ","type":"user"},{"_id":"6262fb326f289e10ee04f63b","avatarUrl":"/avatars/e5bfc2f0ce61aee19a7e4e2cf72a8f51.svg","isPro":false,"fullname":"Omer David","user":"OD","type":"user"},{"_id":"64f60d233a14cc4dd8a70fb3","avatarUrl":"/avatars/fc8b92b4b62896e8fc8d39db6bbc9a98.svg","isPro":false,"fullname":"Omri Kaduri","user":"kaduro","type":"user"},{"_id":"62fa22ea363251ee40a1cd1a","avatarUrl":"/avatars/aa5922b95b220f0d1cfaa82d1fac4427.svg","isPro":false,"fullname":"Hupert","user":"Oded","type":"user"},{"_id":"69a40beef29b3c3292fd8e66","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/QK5X58WJa0A0BKxtU8t9Q.png","isPro":false,"fullname":"Lily Adams","user":"loganwi","type":"user"},{"_id":"6749cf89e60308c9c1a2df06","avatarUrl":"/avatars/5a7071ee1b36883bbb528f6356daf8ab.svg","isPro":false,"fullname":"YAYA YAYA","user":"yzyzyaya","type":"user"},{"_id":"6a2095b3dd0683c2f0042bd8","avatarUrl":"/avatars/ab3c57e754e8171d2a74be407280e220.svg","isPro":false,"fullname":"אסנת בורק","user":"Osssiiiii","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"60262b67268c201cdc8b7d43","name":"nvidia","fullname":"NVIDIA","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/65df9200dc3292a8983e5017/Vs5FPVCH-VZBipV3qKTuy.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.03911.md"}">
Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching
Abstract
Bootstrap Your Generator framework enables unpaired training of flow matching editing models by leveraging base model knowledge and gradient routing for improved generalization in data-scarce scenarios.
Modern generative models possess a deep understanding of visual content, yet training them for image editing typically requires massive datasets of paired examples. This limits scalability, especially for video editing where collecting paired data is prohibitively expensive. We propose Bootstrap Your Generator (ByG), a general framework for unpaired training of flow matching editing models. It leverages the base model's knowledge without any external signal. Our approach pairs instruction-following cues extracted from the frozen model with cycle-consistency for structure preservation. To make this tractable, we propose to route gradients from downstream losses over clean predictions to noisy training states. We demonstrate state-of-the-art results on challenging data-scarce image and video editing scenarios. Extensive evaluations and user studies show that our method effectively generalizes to unseen domains and outperforms supervised baselines trained on millions of samples. Analysis reveals that our gradient routing bridges the train-inference gap, and extracting semantic cues from a base model provides a robust training signal that obviates the need for external reward models.
Community
Modern generative models possess a deep understanding of visual content, yet training them for image editing typically requires massive datasets of paired examples. This limits scalability, especially for video editing where collecting paired data is prohibitively expensive. We propose Bootstrap Your Generator (ByG), a general framework for unpaired training of flow matching editing models. It leverages the base model's knowledge without any external signal. Our approach pairs instruction-following cues extracted from the frozen model with cycle-consistency for structure preservation. To make this tractable, we propose to route gradients from downstream losses over clean predictions to noisy training states. We demonstrate state-of-the-art results on challenging data-scarce image and video editing scenarios. Extensive evaluations and user studies show that our method effectively generalizes to unseen domains and outperforms supervised baselines trained on millions of samples. Analysis reveals that our gradient routing bridges the train-inference gap, and extracting semantic cues from a base model provides a robust training signal that obviates the need for external reward models.
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/2606.03911 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.03911 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.03911 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.