What to Edit Next: Visually Aligned Image-Editing Follow-Up Suggestions in Conversational Systems
The paper reports a multimodal follow-up suggestion system for image editing that cut visual inconsistency in a live A/B test from 3.7% to 0.9%.
The authors built it from 100,000 real multi-turn image-creation conversations from Qwen App, where 80.1% of samples depended on the current image. Their three-stage method combines supervised fine-tuning on reviewed edit intents, reinforcement learning from user clicks, and a visual verifier to keep suggested edits aligned with the image. In the live test, they also report a 32.70% CTR gain, a 16.32% increase in image take-away rate, and 39.90% more average conversation turns per user. Source: HF Daily Papers' note.
The authors built it from 100,000 real multi-turn image-creation conversations from Qwen App, where 80.1% of samples depended on the current image. Their three-stage method combines supervised fine-tuning on reviewed edit intents, reinforcement learning from user clicks, and a visual verifier to keep suggested edits aligned with the image. In the live test, they also report a 32.70% CTR gain, a 16.32% increase in image take-away rate, and 39.90% more average conversation turns per user. Source: HF Daily Papers' note.
score 4