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).
Alex: Welcome to another episode of ResearchPod.
Sam: Today's paper tackles a surprisingly tricky problem. Imagine you're using an app to edit a photo — you've already made a few changes, and now the app wants to suggest what you might do next. The question is: how does the app know what to suggest? And more importantly, how does it avoid suggesting something that doesn't actually make sense for the picture you're looking at right now?
Alex: So it's not just about suggesting popular edits in general — it has to look at the specific image?
Sam: Exactly. And that's where things get interesting. The authors looked at real editing conversations in Qwen App and found that most follow-up requests only made sense if you could see the current picture. Someone might ask to "make the sky more dramatic" — but if there's no sky visible, that suggestion is useless. So the system needs to read both the conversation and the image, not just one or the other.
Alex: That seems obvious in hindsight, but I'm guessing most systems don't actually do that?
Sam: Right. Most suggestion systems are built around text — what did the user say, what do users usually want next. The image is often treated as an afterthought. This paper argues that's backwards. The image should be central to every suggestion, because the image is what the user is actually trying to change.
Alex: So what do they actually build?
Sam: They build a three-stage training pipeline. Think of it like training an athlete. First, you teach them the fundamentals. Then you have them practice against real opponents to develop judgment. Then you add a referee who catches mistakes the opponent might not notice.
Alex: Walk me through each stage?
Sam: The first stage is careful teaching from real examples. The authors took actual editing conversations, had people review them by hand, and used those to create training targets — basically, examples of what a good next suggestion should look like. The model learns the shape of the task before it ever tries to optimize anything. It's like learning the rules of chess before playing a real game.
Alex: And the second stage is where the clicks come in?
Sam: Yes. Once the model knows the basics, they teach it what users actually prefer by looking at which suggestions people clicked. But here's the subtle part — they don't treat every click equally. Think about how you read a menu. You probably look at the top options first. So if someone clicked the first suggestion and ignored the third, that doesn't necessarily mean the third was bad — they might never have really looked at it. The paper accounts for this by only comparing a clicked suggestion against options that were higher up and likely seen. That makes the preference signal much cleaner.
Alex: That's a clever fix. But if clicks already tell you what people like, why do you need a third stage at all?
Sam: Because "what people click" and "what actually fits the image" are not the same thing. A suggestion can sound appealing and still be completely wrong for the picture. For example, it might suggest adding lens flare to a photo that was taken indoors with no light source to create flare from. Sounds cool, doesn't fit. So the third stage is an image-first check.
Alex: What does that check actually do?
Sam: Think of it like a careful editor who reads the picture before reading the suggestion. It asks two questions: first, does the suggestion depend on something that isn't actually in the image? And second, is the change being requested something that's already true? If the answer to either is yes, the suggestion gets flagged. The paper calls this a source-target structure — what must already exist, and what is supposed to change — but the everyday idea is simple: does this edit make sense for this specific picture?
Alex: Why split it into two questions instead of one overall judgment?
Sam: Because different kinds of mistakes need different checks. "Missing starting point" and "already done" are two separate failure modes. A broad judgment might miss one of them. By separating the checks, the system catches both more reliably.
Alex: What if the image is ambiguous — like, it's not obvious whether something is present or not?
Sam: The paper is careful about that. If the image doesn't give enough evidence to decide clearly, the check fails open rather than forcing a judgment. So the system is strict when the image is clear, and cautious when it isn't. That keeps it from pretending to know more than it does.
Alex: So the preference stage makes suggestions more appealing, but can also make them drift — more repetitive, more likely to clash with the image — and the visual check pulls them back?
Sam: That's exactly the tradeoff the paper identifies. The preference stage improves what expert reviewers judge as quality, but it introduces those side effects. The visual check acts as a guardrail. And when they tested the full system in Qwen App — a real deployment with real users — they saw improvements in how often people clicked suggestions, how often they actually applied edits, and how long they kept editing. The paper suggests that happened because the system was no longer choosing between being appealing and being correct. It was pushed toward both at once.
Alex: So the core lesson is that a good suggestion has to be liked, but it also has to make sense for the actual image in front of you.
Sam: Right. And that sounds straightforward, but building a system that reliably does both — at scale, in a real app — turns out to require all three stages working together. Human review to establish the task, click feedback to learn user preference, and an image check to keep the system grounded in visual reality. Remove any one of those, and something breaks.
Alex: That's a genuinely useful framing — not just for photo editing, but for any AI system that has to give advice about something visual.
Sam: That's a fair takeaway. The paper is specifically about image editing suggestions, but the underlying problem — how do you make a recommendation that's both appealing and actually valid for the specific context — shows up in a lot of places. Thanks for listening to ResearchPod.