Conversational assistants increasingly recommend follow-up edits to help users continue tasks, but most systems focus on text-only interactions, leaving image-creation conversations underexplored. In such settings, useful suggestions must reflect user preferences, offer diverse directions, and remain executable on the current image.
To address this, researchers at Hugging Face analyzed 100,000 real multi-turn image-creation conversations from the Qwen App, finding that 80.1% of follow-up interactions are image-dependent. This underscores the need for multimodal recommendation systems that understand both the visual context and user intent.
The proposed framework operates in three stages. First, it uses real online data to build a human-reviewed table of appropriate follow-up editing intents, then fine-tunes a multimodal policy via supervised fine-tuning (SFT). Second, it aligns rule-guided suggestions with actual user choices by optimizing the policy through multi-objective reinforcement learning, using user click feedback as a reward signal. Third, it introduces a visual verifier that penalizes suggestions inconsistent with the current image, providing additional training supervision.
In experiments, the framework significantly outperformed baselines on both automatic and human evaluations. A live user-randomized A/B test with millions of users showed that visual inconsistency dropped from 3.7% to 0.9%, while recommendation click-through rate (CTR) improved by 32.70%, image take-away rate by 16.32%, and average conversation turns per user by 39.90% (all p<0.05).
These results highlight the practical value of visually aligned follow-up suggestions in making image-creation assistants more helpful, engaging, and easier to continue using. The work is published as a research paper on Hugging Face.