Qwen's Multimodal Image-Editing Framework Boosts Recommendation CTR by 32%

QwenBusinessUnit · hf · 2026-08-11

The Qwen team addresses follow-up edit recommendations in image-creation conversations with a three-stage multimodal alignment framework.

Data Foundation: Collected 100,000 real multi-turn image-creation conversations from Qwen App, finding 80.1% are image-dependent.
Three-Stage Framework: Stage 1 builds a human-reviewed table of editing intents to fine-tune a multimodal policy. Stage 2 uses user click feedback for multi-objective RL optimization. Stage 3 introduces a visual verifier to reduce visual inconsistencies.
Results: In a live A/B test with millions of users, the framework reduced visual inconsistency from 3.7% to 0.9%, improved recommendation CTR by 32.70%, image take-away rate by 16.32%, and average conversation turns by 39.90%.

More from Apps