acceptodds
Desk rejected at ICLR 2027

Generative Visual Recommendation through Retrieval-to-Generation Mapping

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative recommendation has evolved along two major paradigms: generative retrieval (GenRet) and personalized visual generation (PVGen). Specifically, GenRet captures collaborative signals from user–item interactions but remains restricted to existing catalog items. While PVGen synthesize new items, yet may not fully exploit these signals for user preference modeling. However, integrating the two paradigms presents challenges in computational efficiency and semantic misaligned between their retrieval and visual generation latent spaces. To address these challenges, we propose Retrieval-to-Generation Mapping (R2G-Map), which efficiently and effectively integrates GenRet and PVGen by aligning recommendation-oriented preference representations with the visual generation space. Specifically, R2G-Map uses continuous recommender hidden states before discrete token selection to retain fine-grained collaborative and predictive information. To bridge the two latent spaces, we introduce a Factorized Feature Predictor that predicts content, description, and collaborative features and constructs visual condition for image generation. We further develop a Dual-Branch Visual Generator that combines this condition with preference representations through joint attention between an image branch and a latent branch. The resulting visual-token distributions support catalog ranking, while the corresponding token sequences are decoded into personalized item images. Experiments on bundle and sequential recommendation demonstrate that R2G-Map improves ranking performance while generating visually coherent and compatible item images.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
generative retrieval, personalized content generation, fashion recommendation, latent space mapping

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hCRiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Retrieval-to-Generation Mapping (R2G-Map), a framework integrating generative retrieval and personalized visual generation for fashion recommendation. It extracts continuous hidden states from frozen recommenders to preserve fine-grained preferences. A Factorized Feature Predictor (FFP) maps these into content, description, and collaborative features, while a Dual-Branch Visual Generator (DVG) uses joint attention for masked visual-token generation. This supports both catalog ranking and personalized image synthesis, demonstrating substantial improvements over existing baselines in ranking and generation metrics.

Critical strengths
  1. Meaningful integration: Effectively bridges collaborative preference modeling and visual synthesis by using continuous recommender hidden states rather than discrete IDs to preserve fine-grained information.
  2. Well-structured framework: Combines factorized feature prediction and dual-branch visual-token generation coherently. Keeping retrieval backbones frozen allows integration with various recommenders without expensive retraining.
  3. Strong experiments and analyses: Consistently improves ranking across two recommendation tasks, two backbones, and different ranking protocols. Informative ablations and additional analyses on preference representations and generation efficiency provide empirical support for the design.
Critical weaknesses
  1. Limited personalization and generalization: Experiments use POG-derived dataset lacking explicit user histories, demonstrating context-conditioned outfit completion rather than genuine user-level personalization. The sequential setting's relevance to real temporal behavior is unclear, and single-dataset evaluation limits generalizability.
  2. Unfair baseline comparisons: User-preference conditioning is removed from visual baselines (e.g., DiFashion, FashionDPO) due to missing user histories, while R2G-Map leverages learned retrieval representations. This discrepancy makes it difficult to attribute improvements solely to the proposed architecture.
  3. Insufficient visual generation validation: Lacks human evaluation and explicit category/personalization metrics. IS is calculated using a fixed uniform category distribution, potentially hiding mode collapse. Generated items sometimes mismatch ground-truth categories, and limited qualitative/novelty evaluations weaken claims of generating useful new items.
  4. Incomplete architectural isolation and efficiency: Ablations do not independently prove the necessity of dual-branch joint attention or auxiliary feature predictions. Direct ranking with predicted features already improves retrieval, necessitating stronger capacity-matched reranking baselines. Efficiency comparisons focus only on generation backbones, leaving end-to-end recommendation latency and practical scalability insufficiently demonstrated.
Questions for authors
  1. Personalization & datasets: How does R2G-Map capture genuine user preferences without explicit user histories? How are sequences ordered, and how is train/test data leakage prevented within the same original outfits?
  2. Baseline fairness: Can you compare R2G-Map against visual baselines equipped with comparable context/preference representations? How much improvement stems from stronger retrieval conditioning versus the mapping architecture?
  3. Visual evaluation: Why is IS calculated against a fixed uniform category distribution instead of the marginal one? Can you provide standard metrics, category-consistency measurements, human evaluations, and verification of item novelty versus catalog reproduction?
  4. Architecture: Can you provide ablations separately removing the recommendation latent branch, joint attention, and auxiliary feature supervision? Would a parameter-matched reranker on predicted features or pretrained embeddings achieve comparable improvements?
  5. Ranking & scalability: At which denoising step are ranking probabilities obtained, and how sensitive are they to the stochastic trajectory? What are the end-to-end inference latency and computational costs for full-catalog ranking?
2 comments

Thank you for your thoughtful questions about the scope, evaluation, and practical performance of R2G-Map.

Response overview. Below, we clarify the preference signals and dataset construction, discuss the baseline comparison and visual evaluation, and present additional component ablations. We also explain the ranking procedure and its dependence on sampling trajectories, and report full-catalog inference costs and generation throughput.

Q1. Personalization and datasets.

Thank you for asking us to clarify the personalization scope and dataset construction.

Preference signals in bundle completion. R2G-Map addresses context-conditioned bundle recommendation. It does not use individual user interaction histories to construct personalized conditions or claim to recover individual users’ preferences. Our objective is to bridge generative retrieval and content generation by mapping frozen recommendation representations to visual-token prediction and image synthesis. We have revised the relevant descriptions to clarify this scope and remove claims of user-specific personalized content generation.

Scope of the sequential setting. Our sequential setting focuses on outfit-derived sequence continuation rather than user-specific personalization. Our sequential records serialize POG outfits containing four or five items. We use the first three items as context, and retain the remaining one or two as targets. We verify train–test separation using the original source bundle IDs and confirm that no bundle ID appears in both splits.

Q2. Baseline fairness.

Thank you for raising this question. In our experiments, both R2G-Map and the visual baselines operate without explicit preference signals derived from individual users’ interaction histories. R2G-Map bridges generative retrieval and content generation by mapping frozen recommendation states inferred from the observed partial bundle into visual representations that support both catalog retrieval and image synthesis. This setup is appropriate for evaluating context-conditioned bundle recommendation and content generation, and we have revised the manuscript’s descriptions of personalization to accurately reflect this setting.

Table 4 compares our R2G-Map with the SD-v3 visual-generation baseline under the reported bundle recommendation setting. The DVG module outperforms SD-v3 in LPIPS (0.601 vs. 0.614) and compatibility (6.050 vs. 5.910), while achieving 7.39× higher generation throughput (17.348 vs. 2.349 images/s), demonstrating the effectiveness and generation efficiency of our implementation.

Q3. Visual evaluation.

Thank you for asking us to clarify the visual evaluation protocol. Our IS calculation follows the modified evaluation protocol of DiFashion [1], adopting the same formulation for consistency with prior work.

In the revised manuscript, we will supplement this score with additional visual metrics to assess category consistency and item novelty relative to the existing catalog.

Q4. Architecture.

Thank you for asking us to clarify the contribution of the architectural components and the role of additional retrieval controls.

Component ablations. We provide the following component ablations for DDBC-based R2G-Map on POG bundle recommendation under Negative Sampling Ranking:

VariantF1↑Jacc.↑OAS↑
w/o recommendation latent branch0.2880.2200.498
w/o joint attention0.2840.2160.495
w/o auxiliary feature supervision0.2820.2130.490
Full model0.3110.2410.527

Removing each component reduces all three ranking metrics. F1 decreases by 0.023, 0.027, and 0.029, respectively. These results support the usefulness of the direct recommendation latent branch, joint attention, and continued feature alignment within this setting. We will include these results and clarify the ablation definitions in the revised manuscript.

Comparison with discriminative rerankers. R2G-Map already uses explicit ranking supervision. During Stage 1, the Factorized Feature Predictor (FFP) uses a feature-level ranking loss to encourage the target item to score above competing candidates (Sec. 3.3.1, Eqs. 4–5). This loss remains part of the auxiliary feature objective during Stage 2 joint training of the FFP and visual generator (Sec. 3.3.2, Eq. 11). Table 3 also reports a separate ranking-loss ablation. Nevertheless, we recognize that a parameter-matched reranker would provide a useful control for understanding the retrieval gains. In future work, we will add this comparison using the same predicted features or pretrained item embeddings, with matched trainable capacity, supervision, and evaluation protocols.

[1] Xu, Yiyan, et al. "Diffusion models for generative outfit recommendation." Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. 2024.

Q5. Ranking and scalability.

Thank you for requesting these details. R2G-Map obtains ranking probabilities from a first prediction step. Each catalog item is scored by the mean log probability of its 128 precomputed visual tokens. This first timestep scoring enables efficient catalog recommendation without requiring the full 16-step refinement process, which is performed when image generation is needed.

Regarding stochastic sensitivity, ranking probabilities are independent of subsequent visual-token sampling trajectories when the recommender hidden states and conditioning are fixed. In contrast, final-step scoring depends on previously sampled tokens, which may favor a particular visual mode and reduce the ranking accuracy of other plausible candidates. Thus, stochastic visual-token sampling affects image generation but does not introduce trajectory-dependent variation into our catalog ranking.

Scalability. We evaluate on the complete POG bundle recommendation test set, comprising 2,521 bundles and 5,042 target items, using a single NVIDIA A40. Each bundle contains two observed items and two prediction positions, and both DDBC and R2G-Map use batches of 32 bundles. R2G-Map uses the 9k checkpoint and CFG = 6. Full-catalog ranking uses a single fully masked DVG prediction to score all 31,217 items, while image generation uses 16 refinement steps. We measure synchronized wall-clock times for each stage over the entire test set and report amortized time per bundle.

StageTotal test-set time (s)Amortized time per bundle (ms)
DDBC prediction91.09036.132
R2G-Map full-catalog ranking20.4528.113
R2G-Map image generation353.918140.388

Runtime breakdown. The full-catalog scoring introduce limited additional runtime, requiring 8.113 ms per bundle. Image generation accounts for most of the additional computation, requiring 140.388 ms per bundle. Nevertheless, our generation pipeline remains substantially more efficient than the evaluated SD3-based baseline. As shown in Table 4, DVG achieves 17.348 images/s, compared with 2.349 images/s for the SD3-based framework, representing a 7.39× improvement in generation throughput.

Reviewer_Gqe5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The method connects a frozen recommender to an image-token generator. It uses the recommender’s hidden states both directly and through a feature predictor that provides appearance conditioning. The resulting image-token predictions support catalog ranking and image generation. The authors evaluate this on fashion outfit completion and outfit-derived sequence continuation.

Critical strengths
  • Interesting problem setup: The paper connects generative retrieval with visual generation, supporting both catalog recommendation and image synthesis from the same item context. this is a relatively new research direction. A shared visual representation through visual-token predictions is introduced to support both ranking existing products and generating item images.
Critical weaknesses
  • The method is difficult to follow and need clarifications for certain designs. Figure 1 and the method description do not clearly show how multiple target items move through the framework. Although target-slot representations are defined, their assignment to ground-truth items and interaction during visual generation remain unclear. A worked example separating training from inference would improve clarity. Also, the need for a separate FFP conditioning path is not fully established. The generator already receives the recommender’s hidden states through its latent branch. Although removing FFP reduces performance, it remains unclear whether a simpler mapping of those states within the existing branch could achieve similar results without the extra conditioning path. 
  • Further examination of the personalization task and support evidence are needed: The paper describes personalized generation, while its experiments use outfit context without user histories. The results therefore do not directly establish user-specific personalization. Moreover, visual evidence is limited. Figure 2 shows one outfit context, with some outputs changing the target category. More examples would help assess category consistency and structural artifacts. Furthermore, the reported IS does not assess diversity. Its fixed uniform reference rewards confident category predictions without measuring variation across generated images.
Questions for authors
  • What evidence distinguishes the benefit of FFP’s feature supervision from its separate conditioning path? A simpler learned mapping into the latent branch could help clarify this distinction.
  • What motivates discrete (TA-TiTok) rather than continuous image representations? Does the choice mainly enable shared token-probability scoring for retrieval? 
  • Is generation intended to be unconstrained outfit completion, or should users be able to request a category or attributes? How should the category changes in Figure 2 be judged under that intended task? 
  • Can the authors provide additional representative outputs and failure cases? 
  • Could the authors illustrate a two-target example, including slot-to-target assignment during training and image generation at inference?
1 comment

Thank you for your constructive questions about the motivation, architecture, and generation process of R2G-Map.

Response overview. Below, we clarify the complementary roles of FFP supervision and recommendation-state conditioning, the motivation for discrete visual tokens, and the scope of the unconstrained generation task.

Q1. FFP supervision.

Thank you for asking us to clarify the role of FFP supervision. The two paths serve different roles in our design. FFP explicitly maps recommendation states into supervised content, description, and collaborative feature spaces. Its predicted content feature provides global conditioning through AdaLN (Eq. 6), while the latent branch retains the sequence of recommendation states through joint attention (Eqs. 7–8). Both originate from the same HH; FFP supplies a supervised representation rather than additional information at inference.

Table 3 evaluates this through the “w/o FFP module” variant. These results show that hidden-state conditioning alone is less effective within our evaluated architecture.

Q2. Motivation for discrete visual representations.

Thank you for asking us to explain the motivation for discrete visual representations.

Shared retrieval and generation interface. Our first motivation is that discrete visual tokens provide a shared interface for catalog scoring and image synthesis. R2G-Map’s predicted token distributions directly score catalog images through their mean token log probability, while sampled token sequences are decoded into images. This connects retrieval and generation within the same visual representation.

Compatibility with language models. Our second motivation is that discrete visual tokens are compatible with the token-prediction framework of language models. As discussed in our Related Work, Janus [1] integrates discrete visual generation tokens with an LLM through learned adapters and an image prediction head. This compatibility provides a route toward future language-guided and interactive recommendation, where users could express preferences or provide feedback to refine subsequent retrieval and image synthesis.

Q3. Generation task.

Thank you for asking us to clarify the generation task and the interpretation of the qualitative examples. The evaluated task is bundle recommendation and sequential recommendation without an explicit target-category constraint. Accordingly, R2G-Map uses an empty decoder prompt in the reported experiments.

In Figure 2, the context contains a dark dress and white footwear, with earrings and a handbag held out. R2G-Map(DDBC) generates a watch and a light-colored hat, while R2G-Map(TIGER) generates dangling earrings and a straw hat.

Under this unconstrained setting, these alternative categories can be reasonable provided that the generated items complement the context in style, color, and category composition. The held-out outfit provides one observed completion; alternative accessory categories may also complement the context. Therefore, the generated items should be assessed for style consistency, color harmony, and category complementarity, alongside their correspondence to the held-out targets.

Q4. Additional outputs.

Thank you for this suggestion. We will expand the appendix of the revised paper with additional representative outputs covering different outfit contexts and target categories.

Q5. A two-target example.

Thank you for suggesting a concrete example to clarify the two-target generation process.

Example setup. Consider a context containing a dress and shoes, with earrings and a handbag held out.

SlotRecommendation stateAssigned training target
1H1H_{1}Earrings
2H2H_{2}Handbag

Training. Each slot is trained with its corresponding target. For slot jj, FFP predicts the target’s content, description, and collaborative features, supervised by its catalog features. DVG receives the recommender hidden latent, the predicted content feature, and a partially masked visual-token sequence of the same target. It learns to reconstruct that target’s visual tokens. Thus, the earrings supervise slot 1, and the handbag supervises slot 2. Both slots use the same model parameters.

Inference. During inference, the frozen recommender produces hidden latents H1H_{1} and H2H_{2} from the observed context. For each target slot jj, the trained DVG starts from a fully masked sequence of 128 visual tokens, conditioned on HjH_{j} and the content feature predicted by FFP. DVG progressively predicts and refines the visual tokens over 16 steps. The frozen TA-TiTok decoder then renders each completed token sequence into an item image using an empty prompt.

[1] Wu, Chengyue, et al. "Janus: Decoupling visual encoding for unified multimodal understanding and generation." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

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

R2G-Map attaches item-image generation to a frozen semantic-ID generative retriever (TIGER or DDBC) for fashion recommendation. From the retriever's hidden states, a six-query Q-Former (FFP) regresses the target's content, description and collaborative features. The predicted content feature and H condition a 16-layer joint-attention masked generator over 128 TA-TiTok tokens (DVG), trained with cross-entropy and self-distillation. One fully masked pass scores all the items by their precomputed tokens' mean log-probability, and iterative unmasking yields the image. Experiments on POG-derived bundle and sequential tasks report negative-sampling and full-catalog ranking, IS, LPIPS, Gemini-judged compatibility and ablations. Inference and evaluation code is released.

Critical strengths

One generator ranks the whole catalog and synthesizes the image, on top of two different frozen GenRet families. A single fully masked, CFG-paired pass of DVG yields a 128×8,192128\times 8{,}192 table of log-probabilities, and each candidate is scored by gathering its own precomputed TA-TiTok tokens from that table (Sec. 3.3.2 Inference, l.295-298; Sec. 3.4, l.302-305), so all 31,217 items are ranked from one decoder pass while the same network samples the image. The release implements exactly this (src/ranking_utils_qformer.py:267-299, 675-706). At inference the scores depend only on HH and the predicted features (ground-truth features enter only the training-time teacher, src/rg_lsm_model.py:1005-1011), and the TIGER and DDBC configs differ only in latent width, paths and checkpoint. No loss in Eqs. 4-11 ranks candidates by this score (the only in-pool loss acts on FFP's cosine scores), yet the "w/o Ranking Loss" variant still keeps 85% of the TIGER bundle F1 gain and 91-98% of the sequential gains under negative sampling (Table 3). A generator whose own losses only ask it to reproduce target tokens thus becomes a better catalog ranker than the recommender it reads from. That is a reusable idea beyond fashion.

Ranking gains are consistent across both backbones and both tasks, including full-catalog ranking with no target insertion. Under All Ranking over 31,217 items, bundle F1@1 rises from 0.104 to 0.140 (TIGER) and from 0.072 to 0.136 (DDBC), and sequential R@1 from 0.134 to 0.179 and from 0.095 to 0.151 (Tables 7-8, 11-12). These top-1 gains are roughly 4 to 7.5 standard errors under a conservative unpaired binomial approximation, and R2G-Map(DDBC) beats native TIGER on all seven All-ranking columns of Table 1. The larger gains at deeper cutoffs (sequential R@10 +70% and +132%, Table 1) depend partly on how the native rankers fill a full list beyond their beams (Q2), so I weight the top-1 numbers most. The released full-catalog path uses the target only to log its rank (src/inference_engine.py:425-484), and the stage-wise feature rankers show that visual-token scoring adds +0.066 to +0.086 at top-1 over Stage-2 FFP features under negative sampling (Tables 7-10). All 28 "Improv." cells were recomputed and the in-text percentages, and they are correct.

Critical weaknesses

The personalized, high-quality generation claim, half of the stated contribution, is untested on personalization and only weakly measured on quality. The abstract promises "personalized item images" (l.27-28), RQ3 asks about personalization (l.317-319), Sec. 4.2 suggests that R2G-Map better captures "collaborative signals from user-item interactions" (l.393-394), and the conclusion claims "personalized image synthesis" (l.479). Yet the authors state that "the datasets do not contain user interaction histories" (l.350-351), user signal can enter at most at the item level through the CLHE collaborative feature (l.154-158), and I could not find any personalization measure. The second task scores the one or two remaining items of a 4-5-item "record" as an unordered set (A.1 l.603-605, 633-635), although Sec. 3.1 defines next-item prediction (l.143-145), and I could not find how records are formed or ordered (searched "record", "order", "chronolog", "timestamp", "session"). The image metrics say little about quality or novelty. IS against a uniform 50-category reference (l.638-640) equals 50exp⁡(−Hˉ)50\exp(-\bar H) for the mean classifier entropy Hˉ\bar H, a recognizability score with no diversity term (the bundle headline 30.399 vs 28.189 is 0.50 vs 0.57 nats). LPIPS is paired with the held-out image. Comp. is one Gemini judge with no human check, invalid-response count or variance. I found no FID, diversity measure or nearest-catalog analysis behind synthesis "beyond existing catalog items" (l.308-309), and in the only qualitative example (Fig. 2) R2G-Map(DDBC) returns a watch and a hat for earrings and a handbag, which the paper candidly acknowledges (l.431). The decoding setting is also unstated. Sec. 3.4 writes x^=Dv(v^)\hat x = D_v(\hat v) with "a frozen decoder" (l.306-307), but TA-TiTok decodes with text (l.124-125), and by default the release decodes every generated sequence once with the CLIP encoding of "A fashion item image of {category}, on white background." filled from the ground-truth target ("pred_text") and once with an empty prompt ("pred_none"), scoring both (src/inference_engine.py:1109-1121, 1219-1221; src/prompt_cache.py:28-40; src/modeling/modules/blocks.py:409-411; evaluate.py:86-107). The paper never says which output Table 2 and Fig. 2 show, yet credits R2G-Map for working "Without explicit target-category prompts" (l.425-426). In fairness, the baselines also receive category text (l.350-351), so "pred_text" would equalize category information rather than favor R2G-Map; the IS variant seems to follow DiFashion's reference evaluator (THIRD_PARTY_NOTICES.md:9); and ranking is unaffected because the ranking logits bypass the decoder. As it stands, the evidence supports context-conditioned outfit completion with confidently classified images, a narrower claim than the abstract and contribution 3 make. Rescoping the claims and renaming the second task, or adding a user-conditioned experiment (POG was built for personalized outfit generation, Chen et al., 2019, and DiFashion conditions on histories, l.108-110), plus both decode modes with the category accuracy the evaluator already computes, FID or a diversity measure, a nearest-catalog novelty check and judge-human agreement on a subset, would resolve this.

The ranking tables are internally inconsistent in places, the negative-sampling protocol is under-specified, and nothing carries variance. Sec. 3.1 ranks targets against "randomly sampled negatives" (l.134-135); A.1 instead ranks "the supplied candidates" with missing targets inserted and fairly calls this "conditional on target inclusion" (l.625-627). The release follows A.1: it uses backbone-specific pools (raw_meta_dir differs between the DDBC and TIGER configs, configs/inference/*_bundle.yaml:161) and prepends any missing target (src/ranking_utils_qformer.py:471-484). I could not find the pool size, the per-backbone insertion rate, where each native ranker places a target it did not retrieve, or how DiFashion, FashionDPO and the SD variant rank catalog items at all (searched "pool", "candidate", "negative", "sampl"), although the negative-sampling columns carry the largest bundle gains (F1 +123.7% and +67.6%). Table 9's DDBC "Visual Emb." row (R@5/R@10/N@10/N@20 = 0.527/0.563/0.508/0.518) differs from Table 1's R2G-Map(DDBC) sequential row (0.528/0.568/0.502/0.523) and equals Table 3's TIGER "w/o CFG" row exactly. With at most two targets, a duplicate-free list and the definitions in A.1 (l.633-635), N@10−N@5≤0.437 (R@10−R@5)N@10-N@5\le 0.437\,(R@10-R@5) must hold, yet the visual-token rows report 0.085 where at most 0.016 is possible (Table 9) and 0.072 where at most 0.022 is possible (Table 10, whose R@5, R@10, N@10 and N@20 are the Table 1 headline values for R2G-Map(TIGER)); every other row of Tables 9-12 satisfies the bound. Fig. 3's within-group rates (84.90% and 47.79%, l.462-472) imply a top-1 visual-hit rate of at least 47.8% even if GenRet never hit, and any nonzero code-hit rate raises that floor. The reported metrics (Tables 1, 9-12) rule this out for R2G-Map(DDBC) and for every All-ranking row, and R2G-Map(TIGER) under negative sampling reaches it only if over 40% of bundles return the same correct item in both slots (which the released metric scores with precision 1, src/bundle_metrics.py:45-58) or over 90% of two-target sequential contexts hit at rank 1. I could not find seeds, standard deviations, intervals or a checkpoint-selection rule (searched "seed", "std", "variance", "significan", "checkpoint"), although the released checkpoints differ by task (6,000 vs 9,000 Stage-2 steps, README.md:31-34) and several ablation gaps are below one binomial SE (about 0.009 for F1 and 0.010 for sequential recall, e.g. "w/o CFG" F1 0.339 vs 0.342). These problems touch R2G-Map's headline rows and the only evidence for "preservation and correction", so I cannot tell how the tables were assembled. I expect the full-catalog top-1 gains to survive, but the negative-sampling and ablation-level conclusions do not yet rest on numbers I can check. Specifying the pools and insertion rates, scoring every row on identical pools with the released evaluator, regenerating all tables from saved outputs, stating Fig. 3's setting and group sizes, and adding 3-5 seeds or paired bootstrap intervals would resolve this.

Questions for authors
  1. Did Tabs. 2-6 and Fig. 2 use the ground-truth-category decoding prompt or the empty one ("pred_text"/"pred_none", src/inference_engine.py:1109-1121), given l.425-426? Please report both beside the category-conditioned baselines, with category accuracy.
  2. Please give negative-sampling pool sizes, per-backbone insertion rates and where native rankers place inserted targets, confirm identical pools across rows, and explain how DiFashion, FashionDPO, SD, and TIGER/DDBC beyond their beams rank items.
  3. Which values are correct where tables conflict (my second weakness; also CLHE's negative-sampling Jaccard 0.085 < 23F1=0.092\tfrac{2}{3}\mathrm{F1}=0.092, impossible for ≤2 predictions under A.1), and what are Fig. 3's setting and group sizes?
  4. Which scorer and how many repeats produced each ranking number? A.2's "visual-embedding" ranking reads like the release's token_emb scorer, whose records keep the best-ground-truth-rank repeat (src/inference_engine.py:1248-1252), unlike Sec. 3.4's log-likelihood.
  5. Would a matched-capacity discriminative reranker on HH match R2G-Map's ranking? At All-Ranking top-1, visual tokens add +0.002 to +0.009 over Stage-2 FFP features (Tabs. 7-8, 11-12), though +0.066 to +0.086 under negative sampling.
  6. How are sequential "records" formed and ordered? Are splits outfit/record-disjoint, CLHE features and RVQ codebooks train-only, and test-time HH freely decoded (as l.199-203 suggest) rather than teacher-forced? How were checkpoints and hyperparameters chosen?
  7. The README's Google Drive links (README.md:38) can show the owner's account depending on sharing settings; an anonymous file host would be safer during review.
  8. Reference fixes (cited → correct): Chang et al. 2022 "Maskgit" → "MaskGIT"; Chen et al. 2019 "alibaba ifashion" → "Alibaba iFashion"; Esser et al. 2024 "In ICML. OpenReview.net" → ICML 2024, PMLR 235, pp. 12606-12633; Hou et al. 2025 and Ju et al. 2025 "semantic ids" → "Semantic IDs"; Liu et al. 2025 "Discrec" → "DiscRec"; Moosaei et al. 2022 "Outfitgan" → "OutfitGAN"; Pang et al. 2025 "Randar" → "RandAR"; Rajput et al. 2023 "Mahesh Sathiamoorthy" → "Maheswaran Sathiamoorthy"; Yu et al. 2025 "Fashiondpo" → "FashionDPO".
5 comments

Thank you for your careful review and for highlighting the points that require clarification and correction.

Response overview. Below, we clarify decoder conditioning, candidate sampling and ranking, the corrected evaluation results, and the visual-token log-likelihood scorer. We also explain the relationship between retrieval-to-generation mapping and ranking gains, clarify data and feature provenance, and address the anonymity and bibliographic issues.

Q1. Decoding setting.

Thank you for requesting this clarification.

Reported decoding setting. All R2G-Map image-generation results in Tables 2–6 and Figure 2 use an empty decoder prompt, without explicit target-category prompts.

Paired comparison. During exploratory experiments, we also tested category-text conditioning in the decoder. This comparison was omitted from the main paper due to space constraints. Below, we report an additional paired comparison for R2G-Map(DDBC) on the bundle test set associated with negative-sampling evaluation.

Decoder conditionCategory Top-1 (%)Category Top-3 (%)
Empty prompt (pred_none)12.8116.42
Target-category text prompt (pred_text)12.9716.70

Category-text conditioning produces modest aggregate changes: Top-1 and Top-3 accuracy increase by approximately 0.16 and 0.28 percentage points, respectively. R2G-Map predicts visual tokens without target-category text. The two decoding modes therefore share exactly the same predicted token sequence; changing decoder text cannot alter the preceding token-sampling process or the visual-token likelihood scores used for catalog ranking.

This observation is consistent with TA-TiTok’s design [1]. Its authors explain that “the text embeddings supplementing high-level semantic information overlooked by latent tokens,” with larger benefits when fewer tokens cause greater information loss. Decoder-side category text thus supplies additional semantic conditioning, but the architecture does not guarantee that it overrides conflicting visual-token content. This provides a plausible interpretation of our observations, rather than establishing that decoder text is generally ineffective.

Q2. Candidate sampling and ranking protocols.

Thank you for requesting these details.

Candidate pool. Our negative-sampling evaluation adopts DDBC’s target-included candidate-pool paradigm [2]. The two-target bundle completion starts with 200 candidates: two positives and 198 sampled negatives.

The first position is evaluated over the initial candidate set S1=S\mathcal{S}_1 = \mathcal{S}. After selecting its highest-scoring item i^1\hat{i}_1, that item is removed before evaluating the second position. Thus, the candidate pool decreases from 200 to 199 items, and the second-position pool depends on the first selected item. The second predicted code is then matched against the remaining candidates. No resampling or target reinsertion is performed between the two positions. In our experiments, R2G-Map and the underlying recommender (DDBC or TIGER) use exactly the same initial set of 200 candidate item IDs for each test sample and follow the same duplicate-prevention rule. After selecting its first item, each method removes that item from its own candidate set and selects the second item from the remaining 199 candidates. We have corrected the code version and the inaccurate description in the appendix.

Ranking for image-generation baselines. For DiFashion, FashionDPO, and SD, we rank catalog items using image–image CLIP cosine similarity between each generated image and the candidate item images. Scores are computed for every item in the supplied candidate pool under negative-sampling evaluation.

[1] Kim, Dongwon, et al. "Democratizing text-to-image masked generative models with compact text-aware one-dimensional tokens." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

[2] Tu, Teng, et al. "Discrete diffusion for bundle construction." The Fourteenth International Conference on Learning Representations. 2026.

Q3. Negative-sampling protocol.

Thank you for identifying these inconsistencies.

Evaluation-code correction. A legacy version of the evaluation code and its description in Appendix A.1 retained target reinsertion and were not properly aligned with the intended comparison on the native backbone’s candidate pools. We have corrected the candidate handling and clarify the protocol below.

Target-insertion rates. The POG bundle recommendation test set contains 2,521 bundles and 5,042 target positions. The previous implementation inserted missing assigned targets at the following rates:

BackboneInserted positions / all positionsOverall insertion rateSecond-position insertion rate
R2G-Map(DDBC)140 / 5,0422.78%140 / 2,521 = 5.55%
R2G-Map(TIGER)481 / 5,0429.54%481 / 2,521 = 19.08%

We have corrected the version mismatch in the evaluation code. We additionally provide a controlled comparison with and without reinserting missing target items into the candidate pool for POG bundle completion under negative-sampling ranking, using the same checkpoints, model forward outputs, and metric definitions.

BackboneMethod / candidate handlingF1@1 ↑Jaccard@1 ↑OAS ↑
R2G-Map(DDBC)R2G-Map0.3110.2410.527
R2G-Map(DDBC)R2G-Map, reinsertion0.3170.2460.525
R2G-Map(TIGER)R2G-Map0.3420.2540.530
R2G-Map(TIGER)R2G-Map, reinsertion0.3500.2690.527

Table and figure corrections. Our check against the experiment records confirms that Tables 1 and 9 both use CFG = 6; their discrepancy is therefore not explained by different CFG settings. We identified reporting/aggregation errors in Tables 9 and 10 and in the statistics of Figure 3, and have corrected them. The Table 3 ablation values were checked against the original experiment records and confirmed to be correct for their original setting.

Figure 3 evaluation setting. Figure 3 uses TIGER-based R2G-Map, CFG = 6, and evaluation seed 42, with fixed TIGER predictions and hidden states. We evaluate POG bundle completion with two observed items and two held-out targets, covering 2,521 bundles / 5,042 prediction positions.

GenRet outcomeVisual hitVisual missGroup sizeWithin-group Visual accuracy
GenRet correct7432851,02872.28%
GenRet incorrect9833,0314,01424.49%
Total1,7263,3165,04234.23%

Q4. Scorer.

Thank you for pointing out this ambiguity. Appendix A.2 uses the same scorer defined in Sec. 3.4, implemented as candidate_log_likelihood. The phrase “visual-embedding ranking” was imprecise and did not refer to the sampled-token embedding similarity scorer token_emb. We will replace this phrase with “visual-token log-likelihood ranking” throughout A.2 and the corresponding table labels.

Q5. Retrieval-to-Generation Mapping and Ranking Gains.

Thank you for raising this comparison. Our contribution is the mapping from frozen recommendation states into a visual-token interface that supports both catalog scoring and image synthesis. The existing results demonstrate this joint functionality: R2G-Map enables retrieval-to-generation mapping while further improving retrieval performance.

After mapping into the visual-token space, likelihood-based scoring provides additional retrieval gains over direct FFP feature scoring. Although the additional All Ranking gains are modest compared with Negative Sampling Ranking, they show that the interface supporting image synthesis can also improve catalog retrieval. This joint capability is the central motivation of our work.

Q6. Data and feature provenance.

Thank you for asking us to clarify the data and feature provenance. We use POG outfit data and follow the feature-construction methodology of CLHE. POG represents an outfit as a set of compatible fashion items. Our sequential benchmark serializes these outfits: each retained record contains four or five items, with three context items and one or two held-out targets. Following CLHE, the item representations incorporate multimodal content and LightGCN-based user-feedback features.

The released inference path obtains the hidden states H through free decoding from the observed context. We select checkpoints on the validation set. For bundle recommendation, we select the checkpoint with the highest validation F1@1. For sequential recommendation, we select the checkpoint with the highest validation Recall@5. We then evaluate the selected checkpoints on the held-out test set and report those results.

Q7–Q8. Anonymity and bibliographic.

Thank you for flagging these issues. We will replace the README’s Google Drive links with anonymous artifact links and verify that hosting pages and downloads reveal no identifying information. We will also correct all listed bibliographic errors and check the rendered bibliography.

Thank you for the thorough rebuttal and corrections. Several of my concerns have been substantially addressed. I am maintaining my score of 2 (Weak Reject) for now, but would consider raising it to 3 (Weak Accept). I would also welcome the AC treating my score as raised if the authors convincingly address the following points:

  1. Numerical consistency: Provide corrected Tables 9–10, recomputed using the released evaluator, and confirm that all reported metrics satisfy their mathematical constraints and that the main gains persist.

  2. Personalization: Either demonstrate measurable improvements on a user-conditioned benchmark with actual interaction histories or explicitly remove personalization claims and frame the task as context-conditioned outfit completion.

  3. Statistical reliability: Provide paired bootstrap 95% confidence intervals or results across 3–5 seeds for the primary ranking improvements, particularly full-catalog top-1 performance.

I do not expect every secondary concern to be resolved. Convincingly addressing these three points would be sufficient for me to support acceptance, even if I am unable to update my score directly before the discussion closes.

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

Related papers

Discussion (0)

Sign in to comment.