Cover images are a surprisingly high-stakes problem in recommender systems. A blurry or contextually mismatched thumbnail can suppress click-through rates even when the underlying content is excellent. Yet most text-to-image pipelines were designed for creative professionals crafting individual prompts, not for platforms generating covers at the scale of millions of items per day. This paper, ICG: Improving Cover Image Generation via MLLM-based Prompting and Personalized Preference Alignment, addresses that gap directly, proposing a framework that sits at the intersection of multimodal language models, diffusion-based image synthesis, and preference-aware reward learning.
The Problem Worth Solving
Existing text-to-image systems like Stable Diffusion, Midjourney, and DALL-E 3 produce impressive results when given well-engineered prompts. The trouble is that prompt engineering does not scale. News aggregators, streaming platforms, and short-video feeds cannot employ a designer per item. Automated MLLM-based prompt generation is the obvious next step, but it introduces its own problems: hallucinated or semantically misaligned prompts, no ground-truth reference images to supervise against, and a complete absence of user personalisation. Prior work in personalised image generation, such as PMG and DiFashion, assumes that the next interacted item's image is a high-quality training target, which is frequently false in practice, and represents user preferences as discrete keywords rather than continuous embeddings, blocking end-to-end gradient flow.
ICG targets all three failure modes simultaneously, which is what makes it worth examining carefully rather than treating as an incremental engineering contribution.
Key Contributions and Architecture
The framework has four tightly coupled components:
- MLLM-based context prompting. Rather than asking an MLLM to produce a text string that then gets fed to a diffusion model, ICG introduces meta tokens, continuous embeddings extracted from both the item title and a reference image. This sidesteps the discrete bottleneck that makes keyword-based preference representations brittle.
- Personalised prompting. User interaction histories are encoded into user embeddings and fused with the meta token representations. The fusion is not a simple concatenation; it conditions the context features on individual user profiles, allowing the same item to generate stylistically different covers for different users.
- Dual-path cross-attention adapter. The personalised features are injected into a frozen Stable Diffusion v1.5 backbone via a lightweight adapter. Keeping the backbone frozen is a deliberate design choice: it preserves the pretrained generative capacity and means the adapter can be dropped into other community checkpoints without retraining. The paper demonstrates this on Realistic Vision V4.0, Anything v4, and Esthetic Retro Anime, which is a meaningful practical test.
- Multi-reward learning. Because there are no ground-truth cover labels, the model is optimised using three complementary reward signals: CLIP similarity for semantic relevance, HPSv2 and PickScore for aesthetic quality, and a custom personalised reward model trained from user-item interaction data using pairwise preference comparisons. The reward model itself is multimodal, combining image features, titles, and user profiles through a Transformer fusion layer.
Results and What They Actually Show
Experiments run on PixelRec (a large-scale short video cover dataset with 10M interactions) and MovieLens (86K movies, 3.3M ratings). Evaluation covers FID and aesthetic scores for image quality, LPIPS and SSIM for fidelity to reference images, downstream recommendation metrics (Recall@10, NDCG@10), and a human preference study with 100 volunteers rating 120 images.
The ablation results are the most analytically useful part of the paper. Removing CLIP reward degrades both FID and LPIPS noticeably. Removing HPSv2 and PickScore harms aesthetics disproportionately. Removing the personalised reward produces the largest single drop across fidelity and alignment metrics, which supports the authors' central claim that user-specific conditioning is doing real work rather than acting as a cosmetic addition. The visualisation of Harry Potter and Interstellar covers adapted to different user profiles (cartoon versus thriller versus romance styles) is qualitatively convincing, though one would want to see this tested at larger scale before drawing strong conclusions.
The human evaluation, while limited to 120 images, shows ICG scoring highest on a 1-3 Likert scale for visual quality and preference alignment. The authors are appropriately cautious here, noting this measures perceived appeal rather than actual engagement.
Limitations and Open Questions
The paper is unusually candid about its shortcomings, which is worth acknowledging. Three stand out as genuinely important:
- Static user embeddings. The framework encodes historical interactions into a fixed user representation. This is fine for stable long-term preferences but will miss short-term intent shifts, which are common in news and trending content scenarios. Session-based or online user modelling would be a natural extension.
- Computational overhead. Multi-reward training adds roughly 20% to training cost, and inference sits at about 1.5 seconds per image on a V100. For a platform generating covers at scale, that latency matters. The authors acknowledge this and flag model compression as future work, but no concrete path is offered.
- Offline evaluation proxies. Recall@10 and NDCG@10 are standard recommendation metrics, but they are indirect proxies for the thing that actually matters: whether better covers increase engagement in production. A live A/B test is the missing experiment, and its absence is the main reason to treat the downstream recommendation results with some caution.
There is also a broader question about the reward model's training signal. Pairwise preferences derived from view counts and ratings conflate cover quality with content quality. A highly rated film might get more clicks regardless of its thumbnail. Disentangling cover-specific appeal from content quality in the training signal is a non-trivial problem the paper does not fully address.
The plug-and-play adapter design is genuinely useful for deployment flexibility, and the no-ground-truth-labels optimisation is a real practical advantage over supervised baselines. Whether the personalisation gains hold when user histories are sparse, which is the cold-start regime that matters most commercially, is not explored. That feels like the next important question for this line of work.
Overall, ICG is a well-motivated and technically coherent framework that makes a credible case for integrating continuous MLLM features, user embeddings, and multi-reward supervision into a single end-to-end trainable system. The compatibility experiments across community checkpoints add practical value beyond the benchmark numbers. The gap between offline evaluation and real-world deployment remains the central open question.
Full paper: arxiv.org/abs/2605.27374