Image-generation conversations rarely end after a single turn—but suggesting what users should edit next requires more than text-only recommendation. In this work, we study follow-up edit suggestions grounded in the current image and user intent. An analysis of 100,000 real-world conversations shows that 80.1% of follow-up interactions are image-dependent. We introduce a three-stage framework combining human-reviewed intent construction, multi-objective reinforcement learning from user click feedback, and a visual verifier that penalizes suggestions inconsistent with the image. In a large-scale online A/B test involving millions of users, our approach reduces visual inconsistency from 3.7% to 0.9%, while improving recommendation CTR by 32.70%, image take-away rate by 16.32%, and average conversation turns by 39.90%. These results highlight visually aligned follow-up suggestions as a practical way to make image-creation assistants more helpful, engaging, and easier to continue using.</p>\n","updatedAt":"2026-08-11T06:30:36.592Z","author":{"_id":"660a645d5814b3dfc3a95e83","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/660a645d5814b3dfc3a95e83/S-yLw7-QFXg7fX94bfbtn.jpeg","fullname":"Bingnan Li","name":"Cuttle-fish-my","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9003162980079651},"editors":["Cuttle-fish-my"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/660a645d5814b3dfc3a95e83/S-yLw7-QFXg7fX94bfbtn.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.07565","authors":[{"_id":"6a7ac0ef019ce76dc7b3ab55","name":"Zhijing Zhang","hidden":false},{"_id":"6a7ac0ef019ce76dc7b3ab56","user":{"_id":"64489bb5e21484883408a96d","avatarUrl":"/avatars/23f7aa4d733a7708fab4ee059ac2b323.svg","isPro":false,"fullname":"Jinpeng YU","user":"Jacob-Yu","type":"user","name":"Jacob-Yu"},"name":"Jinpeng Yu","status":"claimed_verified","statusLastChangedAt":"2026-08-11T08:45:04.553Z","hidden":false},{"_id":"6a7ac0ef019ce76dc7b3ab57","user":{"_id":"6a7ac33d60baf174966e9f0c","avatarUrl":"/avatars/e808a2c7c8caf089a76f1d04a32e402f.svg","isPro":false,"fullname":"Xin Song","user":"gorillasx","type":"user","name":"gorillasx"},"name":"Xin Song","status":"claimed_verified","statusLastChangedAt":"2026-08-11T08:45:04.546Z","hidden":false},{"_id":"6a7ac0ef019ce76dc7b3ab58","user":{"_id":"660a645d5814b3dfc3a95e83","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/660a645d5814b3dfc3a95e83/S-yLw7-QFXg7fX94bfbtn.jpeg","isPro":false,"fullname":"Bingnan Li","user":"Cuttle-fish-my","type":"user","name":"Cuttle-fish-my"},"name":"Bingnan Li","status":"claimed_verified","statusLastChangedAt":"2026-08-11T08:45:04.531Z","hidden":false},{"_id":"6a7ac0ef019ce76dc7b3ab59","user":{"_id":"688cceac29c69ec01eb28644","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/wyL1U86pB5GIi4x3JflOa.png","isPro":false,"fullname":"Chuyue Li","user":"woody-woody","type":"user","name":"woody-woody"},"name":"Chuyue Li","status":"claimed_verified","statusLastChangedAt":"2026-08-11T08:45:04.538Z","hidden":false},{"_id":"6a7ac0ef019ce76dc7b3ab5a","user":{"_id":"6924165eced02d07051f24ee","avatarUrl":"/avatars/7078b65db0ff5bfa8df4410e2b092776.svg","isPro":false,"fullname":"duchanghui","user":"Duchanghui","type":"user","name":"Duchanghui"},"name":"Changhui Du","status":"claimed_verified","statusLastChangedAt":"2026-08-11T16:45:04.685Z","hidden":false},{"_id":"6a7ac0ef019ce76dc7b3ab5b","name":"Xiaolin Fang","hidden":false},{"_id":"6a7ac0ef019ce76dc7b3ab5c","name":"Jiaming Liu","hidden":false},{"_id":"6a7ac0ef019ce76dc7b3ab5d","name":"Ruihua Huang","hidden":false}],"publishedAt":"2026-08-03T00:00:00.000Z","submittedOnDailyAt":"2026-08-11T00:00:00.000Z","title":"What to Edit Next: Visually Aligned Image-Editing Follow-Up Suggestions in Conversational Systems","submittedOnDailyBy":{"_id":"660a645d5814b3dfc3a95e83","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/660a645d5814b3dfc3a95e83/S-yLw7-QFXg7fX94bfbtn.jpeg","isPro":false,"fullname":"Bingnan Li","user":"Cuttle-fish-my","type":"user","name":"Cuttle-fish-my"},"summary":"Conversational assistants increasingly recommend follow-up edits to help users continue a task. Existing systems primarily target text-only interactions, leaving image-creation conversations underexplored. In image-creation tasks, useful follow-up edit suggestions must reflect user preferences, offer diverse directions, and remain executable on the current image. We collected 100,000 real multi-turn image-creation conversation samples from Qwen App and found that 80.1% are image-dependent, underscoring the need for multimodal recommendation. We address this setting with a three-stage framework. In Stage 1, we use real online data to build a human-reviewed table of appropriate follow-up editing intents, then create SFT targets and fine-tune a multimodal policy. In Stage 2, to align rule-guided SFT suggestions with actual user choices, we use user click feedback to optimize the policy through multi-objective reinforcement learning. In Stage 3, to reduce visual inconsistencies between suggested edits and the current image, we introduce a visual verifier as additional training supervision. Extensive experiments demonstrate that our framework significantly outperforms baselines on both automatic and human evaluations. In a live user-randomized A/B test with millions of users, our final framework reduces visual inconsistency from 3.7% to 0.9%. Furthermore, it significantly improves recommendation CTR by 32.70%, image take-away rate by 16.32%, and average conversation turns per user by 39.90% (all p<0.05).","upvotes":22,"discussionId":"6a7ac0ef019ce76dc7b3ab5e","ai_summary":"A three-stage multimodal framework improves follow-up edit recommendations in image-creation conversations by combining supervised fine-tuning, multi-objective reinforcement learning, and visual verification.","ai_keywords":["multimodal policy","supervised fine-tuning","multi-objective reinforcement learning","visual verifier","image-dependent recommendation"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"6a6841e7107886ba1a151b03","name":"QwenBusinessUnit","fullname":"Qwen Business Unit","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/66f79b323fe089b75e9e0c04/MlefZsdry-JuhKzAwxjQl.webp"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"660a645d5814b3dfc3a95e83","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/660a645d5814b3dfc3a95e83/S-yLw7-QFXg7fX94bfbtn.jpeg","isPro":false,"fullname":"Bingnan Li","user":"Cuttle-fish-my","type":"user"},{"_id":"688cceac29c69ec01eb28644","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/wyL1U86pB5GIi4x3JflOa.png","isPro":false,"fullname":"Chuyue Li","user":"woody-woody","type":"user"},{"_id":"6812f7f3508697fa11f1708b","avatarUrl":"/avatars/a036cf42eac4119bd32739b4fb5291c8.svg","isPro":false,"fullname":"Yi Chen Liu","user":"starry0929","type":"user"},{"_id":"6759582c44ec15bfe4bcf6a2","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/kSkj4OIjHXbYV7qhIhH78.png","isPro":false,"fullname":"li","user":"wang0422","type":"user"},{"_id":"6a7ac33d60baf174966e9f0c","avatarUrl":"/avatars/e808a2c7c8caf089a76f1d04a32e402f.svg","isPro":false,"fullname":"Xin Song","user":"gorillasx","type":"user"},{"_id":"65bf03acbf441a1953e77106","avatarUrl":"/avatars/086489e5b33979b23c490ee58425a297.svg","isPro":false,"fullname":"xc","user":"das12138","type":"user"},{"_id":"6924165eced02d07051f24ee","avatarUrl":"/avatars/7078b65db0ff5bfa8df4410e2b092776.svg","isPro":false,"fullname":"duchanghui","user":"Duchanghui","type":"user"},{"_id":"637c941588699fba70e29f70","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/637c941588699fba70e29f70/b6G_QZkT-MhE47dx87i0d.png","isPro":false,"fullname":"LIU JIAMING","user":"jamesliu1217","type":"user"},{"_id":"668fcd97495674543600db0b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/JfBeJ6gSXe379G9ejjrLR.jpeg","isPro":false,"fullname":"Gao Mingqi","user":"EchoMinkki","type":"user"},{"_id":"661cf93620b47b0dad13ab89","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/661cf93620b47b0dad13ab89/lwSsSabpkOEC9ADGehp64.png","isPro":false,"fullname":"Hao-Xuan Ma","user":"gh0stHunter","type":"user"},{"_id":"6a7ac5c66303ef7d135264b4","avatarUrl":"/avatars/734471516b23803abbf319d0d0cd2204.svg","isPro":false,"fullname":"Jie Sheng","user":"sadff12","type":"user"},{"_id":"668366de0b72be1367e60cd3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/668366de0b72be1367e60cd3/GyFEIpNFV4eOmMsOdBwja.jpeg","isPro":false,"fullname":"Yu Chen","user":"Rocky131","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6a6841e7107886ba1a151b03","name":"QwenBusinessUnit","fullname":"Qwen Business Unit","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/66f79b323fe089b75e9e0c04/MlefZsdry-JuhKzAwxjQl.webp"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.07565.md","query":{}}">
What to Edit Next: Visually Aligned Image-Editing Follow-Up Suggestions in Conversational Systems
Abstract
A three-stage multimodal framework improves follow-up edit recommendations in image-creation conversations by combining supervised fine-tuning, multi-objective reinforcement learning, and visual verification.
Conversational assistants increasingly recommend follow-up edits to help users continue a task. Existing systems primarily target text-only interactions, leaving image-creation conversations underexplored. In image-creation tasks, useful follow-up edit suggestions must reflect user preferences, offer diverse directions, and remain executable on the current image. We collected 100,000 real multi-turn image-creation conversation samples from Qwen App and found that 80.1% are image-dependent, underscoring the need for multimodal recommendation. We address this setting with a three-stage framework. In Stage 1, we use real online data to build a human-reviewed table of appropriate follow-up editing intents, then create SFT targets and fine-tune a multimodal policy. In Stage 2, to align rule-guided SFT suggestions with actual user choices, we use user click feedback to optimize the policy through multi-objective reinforcement learning. In Stage 3, to reduce visual inconsistencies between suggested edits and the current image, we introduce a visual verifier as additional training supervision. Extensive experiments demonstrate that our framework significantly outperforms baselines on both automatic and human evaluations. In a live user-randomized A/B test with millions of users, our final framework reduces visual inconsistency from 3.7% to 0.9%. Furthermore, it significantly improves recommendation CTR by 32.70%, image take-away rate by 16.32%, and average conversation turns per user by 39.90% (all p<0.05).
Community
Image-generation conversations rarely end after a single turn—but suggesting what users should edit next requires more than text-only recommendation. In this work, we study follow-up edit suggestions grounded in the current image and user intent. An analysis of 100,000 real-world conversations shows that 80.1% of follow-up interactions are image-dependent. We introduce a three-stage framework combining human-reviewed intent construction, multi-objective reinforcement learning from user click feedback, and a visual verifier that penalizes suggestions inconsistent with the image. In a large-scale online A/B test involving millions of users, our approach reduces visual inconsistency from 3.7% to 0.9%, while improving recommendation CTR by 32.70%, image take-away rate by 16.32%, and average conversation turns by 39.90%. These results highlight visually aligned follow-up suggestions as a practical way to make image-creation assistants more helpful, engaging, and easier to continue using.
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/2608.07565 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.07565 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.07565 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.