acceptodds
Withdrawn from ICLR 2027

Query-Grounded Preference-Aware Personalized Sticker Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Personalized sticker generation with intent aims to create stickers that reflect both a user's visual preferences and semantic intent. Existing reference-aware approaches rely on sticker proxies to infer semantic intent, which may dilute intent extraction when suitable visual exemplars are unavailable. We propose Query-Grounded Preference-Aware Personalized Sticker Generation (QuPA-StickerGen), a framework that directly incorporates the user's textual query. First, we aggregate reusable style adapters with supervision from a vision-language model. Second, user sticker usage preferences and textual queries are mapped into a shared style space. Third, both signals are adaptively fused according to query confidence weight. This enables the generation of stickers that better align with both how users prefer content to look and what they want to express. Experiments show that our approach outperforms existing methods in both user stylistic preference and semantic alignment. Dataset and code are released.

Primary area:
generative models
Keywords:
Sticker Generation, Personalization, Query-Grounded

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_5Xt9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes QuPA-StickerGen, a framework for personalized sticker generation that uses the user’s sticker usage history as a signal of visual preference and the textual query as a direct signal of semantic intent, instead of a retrieved reference sticker. Training stickers are grouped into style clusters with CSD embeddings and VLM-based filtering, and one LoRA adapter is trained per cluster. A query-based style distribution (from semantically similar training queries) and a history-based style distribution (with temporal decay) are fused according to query confidence, and the top adapters are composed for generation. The paper also introduces StickerGen, built by linking U-Sticker and StickerQueries, and motivates the formulation with a preliminary study comparing History + Query, History + Reference, and History + Query + Reference. The method is evaluated on StickerGen and WXChallenge with SDXL and FLUX against generation, retrieval, and LLM-pipeline baselines, together with a human evaluation.

Critical strengths
  • StickerGen combines user sticker usage histories and textual queries with ground-truth stickers in one public dataset, which existing public sticker datasets do not provide together (Section 2.3). This could be a useful resource for personalized sticker generation.

  • Using the textual query directly as the semantic signal, instead of a retrieved reference sticker, is a reasonable direction, and the human evaluation shows a preference for the proposed method in query alignment (45.3% vs. 33.3% chance level).

Critical weaknesses
  • The ablation does not support the main claim that jointly modeling the query and the history improves generation. In Table 5, History + Query and History only are nearly identical in Overall score (StickerGen: .4258 vs. .4250 with SDXL, .4054 vs. .4044 with FLUX), and on WXChallenge with SDXL, History only is higher (.3679 vs. .3659). The semantic scores of Query only and History only are also similar (e.g., .1735 vs. .1725 on WXChallenge with SDXL). In addition, the Overall scores in the WXChallenge columns do not match the average of Style and Semantic (e.g., History only with SDXL: Style .3952 and Semantic .1725, but Overall .3679), while they do match in the StickerGen columns. This makes it difficult to determine how the Overall scores are computed and whether the query contributes to the reported gains.
  • The preliminary study in Section 4, which is the basis for formulating the task as History + Query (Section 1, first contribution), has prompts that do not match its settings. In Appendix B.2, the History + Reference prompt includes the query (“The search query is: {query}” and “reflects the query {query}”), while the History + Query + Reference prompt does not include the query. It is therefore unclear which inputs each setting in Table 2 actually received, which makes it difficult to rely on the conclusion that the retrieved reference may introduce noise.
Questions for authors
  • (Clarification) How are the Overall scores in Table 5 computed, and why do the WXChallenge columns not match the average of Style and Semantic?
  • (Clarification) Which inputs did each setting in Table 2 receive? Are the prompts for History + Reference and History + Query + Reference in Appendix B.2 swapped, or were the experiments run with these prompts?
Reviewer_zckGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper formulates personalized sticker generation as joint modelling of current intent (textual query) and persistent preference (user history), and argues that the reference sticker retrieved by reference-aware methods is a lossy proxy for intent. The paper shows that the textual query alone provides sufficient semantic grounding, and that a retrieved reference sticker adds nothing to it. The paper thus proposes QuPA-StickerGen, which clusters training stickers in CSD style space with one LoRA per cluster, maps query and history to distributions over clusters (query via SBERT retrieval over training queries, history via recency decay), fuses them by a geometric mean whose weight is sigmoid(a + b·z(q)), and composes the top-k′ adapters. The paper validates the framework on STICKERGEN and WXCHALLENGE against generation, retrieval and LMM-pipeline baselines, and with a 40-rater human study.

Critical strengths

S1. The query-versus-reference comparison. The controlled experiment shows clearly that the textual query carries the intent on its own, and that the retrieved reference contributes interfering visual signal rather than information. This is the paper’s strongest evidence and the clearest support for the formulation it proposes.

S2. The dataset fills a real gap. No public resource offers query–sticker pairs together with per-user histories, and the authors provide one at a usable scale.

Critical weaknesses

W1. The central claim is not supported. None of the components of the method is ablated — the style clusters, the per-cluster adapters, the query weighting, the recency decay, the fusion and the confidence gate are introduced together — so their effectiveness cannot be separated. And on WXCHALLENGE with SDXL, History+Query scores below History-only, so the benefit of combining the two signals is not established either.

W2. The metric computation is not specified. How Style Avg, Semantic Avg and Overall are aggregated is never stated, so the tables cannot be read at face value. Table 3 also does not say which test its significance markers come from, nor whether any correction is applied.

Questions for authors

Q1. Ablate the design: report History+Query against a disabled or fixed-weight fusion, and against removal of each component (K, the per-cluster adapters, the query weighting, the recency decay), with confidence intervals; state what (a, b) are fit on. Please also reconcile Table 5’s WXCHALLENGE semantic column with Table 3.

Reviewer_GmsyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes QuPA-StickerGen, a framework for personalized sticker generation that combines users’ historical preferences with textual queries, and introduces StickerGen, a dataset comprising user history–query–target sticker triples. The method maps user history and queries into a shared style space, adaptively fuses the resulting preferences based on query confidence, and selects and combines offline-trained style LoRA adapters. No user-specific optimization is required at inference time.

Critical strengths

Well-motivated problem formulation and practical shared-adapter design. The paper clearly distinguishes between what users currently want to express and their typical visual style preferences, representing these through textual queries and historical stickers, respectively. Mapping both signals into a shared style space and composing pretrained adapters provides an intuitive, modular implementation. The framework avoids retraining a personalized model for each new user or query, offering a clear deployment rationale.

Critical weaknesses
  1. Insufficiently established methodological novelty. Section 5 primarily integrates style clustering, similar-query retrieval, temporal weighting of user history, and LoRA adapter selection and composition into a personalized generation pipeline. Its core fusion mechanism uses a query-confidence-controlled weighted geometric mean. Although the framework has practical value, its methodological contribution appears to lie mainly in component integration. The paper does not sufficiently articulate a distinctive technical advance in preference modeling beyond straightforward information fusion and adapter composition.

  2. Insufficient direct evidence that Adaptive Fusion outperforms simpler fusion strategies. Equations (8)–(11) adjust fusion weights according to the maximum similarity between the input query and training queries, but the paper does not verify whether this similarity reliably indicates when to rely more on the query and when to favor user history. In Table 5, the combined model yields only a marginal improvement in Overall over History Only on StickerGen–SDXL and performs worse on WXChallenge–SDXL. Moreover, the absence of fixed-weight fusion baselines and variability estimates for the ablations prevents a convincing assessment of whether the adaptive weighting mechanism provides consistent, independent gains.

  3. Overly dense framework illustration. The framework diagram is densely packed with information and resembles a detailed procedural flowchart rather than an intuitive overview of the method. Its extensive use of mathematical notation makes it difficult for readers to quickly grasp the key ideas and relationships among components.

Questions for authors
  1. Section 5 integrates style clustering, similar-query retrieval, temporal weighting of user history, and LoRA adapter composition into a personalized generation pipeline. Could the authors clarify the key technical contribution beyond directly combining these components? For example, does VLM-assisted style clustering provide measurable improvements over clustering based solely on CSD features under the same adapter training budget?

  2. In Table 5, the combined model provides only small Overall improvements over History Only and even underperforms it on WXChallenge–SDXL. Could the authors compare Adaptive Fusion against equal-weight fusion and a fixed fusion weight tuned on the validation set, while keeping the adapter bank, number of selected adapters, generator text inputs, and sampling settings unchanged? Please also report variability across repeated runs. These comparisons would help distinguish gains from using both information sources from those attributable to adaptive weighting itself.

  3. Equations (8)–(10) define confidence as the maximum semantic similarity between the input query and training queries, assigning greater weight to the query branch when confidence is higher. Could the authors group queries by confidence and examine whether higher confidence indeed corresponds to greater benefits from relying on the query branch rather than the history branch, reporting style and semantic alignment separately? In particular, does the mechanism still make effective weighting decisions when the query-induced style conflicts with the user’s historical preferences?

Comments on the paper

1 comment

We sincerely thank the reviewers for their time and constructive feedback. We appreciate their valuable insights and will continue to refine and strengthen our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.