What to Edit Next: Visually Aligned Image-Editing Follow-Up Suggestions in Conversational Systems
Summary
A three-stage multimodal framework improves follow-up edit recommendations in conversational image-creation systems, using SFT, multi-objective RL, and visual verification to reduce visual inconsistency and boost engagement metrics in a live A/B test.
View Cached Full Text
Cached at: 08/11/26, 10:21 AM
Paper page - What to Edit Next: Visually Aligned Image-Editing Follow-Up Suggestions in Conversational Systems
Source: https://huggingface.co/papers/2608.07565
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 amultimodal policy. In Stage 2, to align rule-guided SFT suggestions with actual user choices, we use user click feedback to optimize the policy throughmulti-objective reinforcement learning. In Stage 3, to reduce visual inconsistencies between suggested edits and the current image, we introduce avisual verifieras 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).
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2608\.07565
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2608.07565 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2608.07565 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2608.07565 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Text-Vision Co-Instructed Image Editing
A new framework called TV-Edit combines textual instructions and visual prompts for precise image editing, along with a benchmark TV-Edit-Bench for evaluation. The method achieves better spatial control and semantic faithfulness than existing approaches.
Uni-Edit: Intelligent Editing Is A General Task For Unified Model Tuning
Uni-Edit proposes using intelligent image editing as a single general task to simultaneously improve unified multimodal models' understanding, generation, and editing capabilities, with an automated data synthesis pipeline creating complex editing instructions.
From Plans to Pixels: Learning to Plan and Orchestrate for Open-Ended Image Editing
An experiential framework for long-horizon image editing that couples planning with reward-driven execution to improve coherence and reliability of complex multi-step edits.
Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing
This paper introduces RE-Edit, a benchmark for evaluating image editing systems across five reasoning dimensions (physical, environmental, cultural, causal, referential) to assess logical consistency beyond visual plausibility. The benchmark includes 1,000 samples and evaluates ten open-source and two commercial models, showing that even advanced systems struggle with implicit multi-dimensional reasoning.
HP-Edit: A Human-Preference Post-Training Framework for Image Editing
HP-Edit introduces a post-training framework that aligns diffusion-based image editing models with human preferences via RLHF, using a new 50K real-world dataset and an automatic VLM-based evaluator.