acceptodds
Under review as a conference paper at ICLR 2027

Towards Personalized Visual Imagination from Everyday User–MLLM Interaction with Unified Multimodal Models

Abstract

Text-to-image (T2I) models has achieved remarkable progress in visual quality and even physical-world reasoning. Nevertheless, when facing personalized image-generation requests that specify only what users currently care about most, the unspecified visual spaces in their resulting images remains shaped by population-level priors, imagining an "average user" rather than the particular individual behind the prompt. Meanwhile, as MLLMs serve as personal assistants, their everyday experiences with users naturally disclose how each individual tends to imagine these open visual spaces. Therefore, we propose VIMG, a new image-generation task that studies how these user–MLLM histories can foster personalized visual imagination for users' daily painting requests. Accordingly, we propose VIMG-Bench to evaluate VIMG, spanning six daily image-generation domains and 30 sub-domains, with a three-level hierarchy of visual imagination preferences (global, domain and sub-domain). It features 100 diverse users, each paired with 168K multimodal interactions with their MLLM assistant that may occasionally reveal such preferences, as well as 3K personalized T2I requests. We benchmark mainstream T2I models and Unified Multimodal Models (UMMs), and find that although UMMs excel at reasoning over complex generation instructions, they remain limited on VIMG, which poses the opposite challenge: imagining what is left unspecified from simple yet highly personalized image requests. We further propose VIMGer, which reinforces UMMs for this new imagination-to-realization paradigm, providing a foundation for advancing the VIMG.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text-to-image generation, Personalizd visual imagination, Unified Multimodal Models, long-term user-MLLM histories

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DZQHOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper study the personalized image generation, in specific, given the user's interaction history (e.g., chat) and an image generation query, system needs to retrieve relevant history/ preference to enrich the image generation. To that end, the paper introduced a benchmark (100 synthetic users, 168k interactions, 3k image generation requests). Using this benchmark, they evaluate the frontier models (e.g., GPT-Image-2, NanoBanana), and also implemented an improved selection model (VIMGer) to further improve accuracy of BAGEL.

Critical strengths
  • I believe the setting where we have users' history/ interaction, and then need to infer user's preference is an emerging topics. I appreciate the efforts of authors' to synthesis the dataset (which include the image generation task) -- This is new as majority of the current benchmark are text-to-text only.
  • The experiments include frontier model (e.g., Nano-Banana-2)
Critical weaknesses

While I like the research questions, I think the current submissions can be significantly improved in (i) baseline setting and (ii) agentic workflow.

  • First, while the paper is targeting retrieval (and also have RAG as a baseline), none of the memory-layer methods are presented (e.g., Mem0, MemOS). This is critical as for digest chat histories to extract the users' preferences, memory-layers methods are standard.
  • Second, the current base model setting is not ideal: Base model is BAGEL, which is significant worse than frontier models (e.g., Nano-Banana-2) in terms of quality, instruction following. I think a better way should be: Implement the baseline where use frontier models as image generation, yet use VIMGer/ RAG/ Memory-layer as history selector (e.g., design an agentic workflow).

While I am not denying the potential usefulness of VMGer, I am saying the current experiment setting/ results do not support the claim that VIMGer is better for extract the users' preference.

Questions for authors

Because the VIMGer scan the history during inference time, what is the cost? (e.g., for each new query, VIMGer will need to scan the history once again?) -- If this is true, then this will be limitation in inference time, as normally interaction history can be 10M+ tokens, not just 12K-token budget.

4 comments
AuthorsOpenReview ↗

Response to Reviewer DZQH (1/3)

We sincerely thank you for the thoughtful reading and for appreciating the novelty of VIMG and our extension of personalization evaluation to image generation. We address your concerns below.

[W1] The baseline settings should include memory-based methods.

We sincerely thank you for this valuable suggestion. We agree that memory-based methods are an important approach to leveraging long-term user interaction histories. In our initial design, we focused primarily on establishing VIMG as a new task and benchmark, and thus adopted relatively simple baselines (History and RAG) rather than exhaustively exploring different memory architectures.

Following your suggestion, we have added Mem0 and MemOS as additional baseline, using BAGEL as the common generation backbone and evaluating them under TSO and PPS settings. The results are reported in the table below, showing relatively limited performance.

MethodTSOPPS
BAGEL (Base)14.2919.15
Mem0 + BAGEL21.8121.95
MemOS + BAGEL22.5523.59
VIMGer38.1439.19

We believe their limitations on VIMG can be understood from two aspects. First, pre-constructed memory may introduce information loss (e.g., SeCom). In particular, memory compression is performed before the current image-generation query is known, making it impossible to preserve information selectively according to its future relevance. This inevitably risks discarding fine-grained visual preference cues that may become crucial for VIMG. Second, query-only-conditioned retrieval may fail to identify the evidences for personalized image painting requests. Both methods retrieve from pre-constructed memory based on the current request. However, VIMG queries are often short and semantically underspecified (e.g., "Draw a soldier"), while relevant visual preferences may have been revealed in seemingly unrelated historical interactions. Consequently, retrieving memories based solely on the current request may fail to recall such evidence. This also motivates our design of VIMGer, which directly accesses the original interaction history through window-wise filtering, identifying query-relevant preference evidence before performing personalized visual imagination.

Nevertheless, we believe memory-based approaches remain a promising direction, as also discussed in our future work. We have incorporated the new baselines and corresponding discussion into the revised manuscript. Thank you again for helping us improve the paper.

AuthorsOpenReview ↗

Response to Reviewer DZQH (2/3)

(Part 2/3)

[W2] Why not adopt an agentic framework with frontier image generators and VIMGer, RAG, or memory as the history selector?

We agree that combining stronger image generators with dedicated history-selection modules is a useful practical approach.

  1. Existing Frontier-model Evaluation. First, we would like to clarify that our paper has already included evaluations of frontier models, namely GPT-Image-2 and Nano-Banana-2. Importantly, rather than directly feeding the current user query to these generators, we adopt a simple agentic workflow implemented as a history-conditioned, two-stage pipeline. Specifically, an MLLM first reads the user's full interaction history together with the current request, identifies and consolidates relevant historical information, and constructs a personalized, self-contained T2I prompt. This prompt is then passed to the corresponding image generator. We use GPT-5.6-Sol and Gemini-3.8-Flash as the respective MLLMs for this process. Thus, our evaluation already incorporates a simple agentic workflow for personalized generation with frontier image models.

  2. History Selector Comparison with Frontier Generators. Furthermore, to directly investigate the reviewer's suggested combination of history selectors and frontier image generators, we conduct a controlled comparison to isolate the contribution of history selection from image generation. Specifically, each frontier pipeline's prompt writer receives sessions selected by RAG, Mem0, MemOS, or VIMGer, while keeping the prompt writer and image generator fixed across selectors. The average VScore-P results are reported below.

    SelectorGPT-Image-2 TSOGPT-Image-2 PPSNano-Banana-2 TSONano-Banana-2 PPS
    RAG42.0744.9346.2149.94
    Mem043.6746.4246.9450.09
    MemOS43.8447.1446.8550.82
    VIMGer49.5752.8753.4856.86

    As shown, VIMGer yields the best results with both frontier generators (e.g., 49.57 vs. 43.84 for MemOS with GPT-Image-2 under TSO). Therefore, VIMGer can also serve as an effective history selector in such agentic frameworks. All comparisons use the same evaluation instances, scoring convention, and judge. These results demonstrate that VIMGer’s learned history-filtering capability generalizes beyond its original UMM backbone and can be effectively transferred to frontier image-generation pipelines.

  3. Research Focus and Backbone Choice. Finally, at the methodological level, we specifically investigate whether UMMs can acquire VIMG capabilities through dedicated post-training strategies (also reflected in our paper title). We focus on UMMs because their unified architecture integrates multimodal understanding and generation, providing a natural foundation for learning VIMG within a single model. More importantly, as MLLMs are increasingly adopted as personal assistants in everyday life and image generation becomes a natural part of daily human–assistant interactions, we believe that VIMG should become an increasingly important capability of general-purpose UMMs. We choose BAGEL as a representative open-source UMM backbone because it allows us to directly study and improve the model's intrinsic capabilities through SFT and RL, while providing greater flexibility for research than closed-source proprietary models, and being computationally efficient (only 7B MoT). We fully acknowledge that its general image-generation capabilities are not on par with frontier models.

We fully agree with you that developing more complex agentic workflows for frontier or closed-source models is another valuable direction for addressing VIMG. Indeed, as already discussed in our Future Work, such approaches could leverage external memory, history selection, prompt refinement, and iterative generation without requiring access to model parameters. Given the clear focus of the current work and VIMGer’s role as an initial baseline for trainable UMMs, we leave a more comprehensive exploration of such agentic approaches to future work. We sincerely thank you again for this thoughtful suggestion, which helps us clarify the scope and broader research directions of VIMG.

AuthorsOpenReview ↗

Response to Reviewer DZQH (3/3)

(Part 3/3)

[W3] Is VIMGer able to extract users' VIMG preferences more effectively?

We thank you for this question. As discussed in our response to the previous concern, the core objective of VIMGer is to enhance UMMs' ability to identify users' personalized visual imagination preferences from long-term interaction histories.

First, our existing experiments demonstrate substantial improvements over the original BAGEL backbone under controlled settings. Specifically, VIMGer improves the average VScore-P across six domains from 14.29 to 38.14 under TSO (+166.9%), and from 19.15 to 39.19 under PPS (+104.7%). Moreover, VIMGer also outperforms the Mem0 and MemOS baselines in Part 1. Overall, these results demonstrate VIMGer’s effectiveness in enhancing UMMs' end-to-end VIMG capabilities.

Furthermore, to investigate whether these improvements generalize beyond BAGEL, we conduct additional experiments using SenseNova-U1 as another UMM backbone. Notably, the original SenseNova-U1 already consistently outperforms BAGEL Base across all six domains under both settings, indicating its stronger initial VIMG performance. With VIMGer, its average VScore-P further improves from 19.72 to 41.21 under TSO and from 27.61 to 44.21 under PPS, which shows that our post-training approach transfers beyond BAGEL.

MethodTSOPPS
SenseNova-U119.7227.61
+ VIMGer41.2144.21

[Q1] Discussion of overhead of VIMGer.

We thank you for raising this question. First, we clarify that the 12K-token budget mentioned in our paper refers to the maximum size of each history window, rather than the total history length. Our current implementation already supports arbitrarily long interaction histories by partitioning them into contiguous windows of at most 12K tokens, filtering each window independently, and merging the retained relevant sessions. As a rough estimate, filtering a 168K-token history takes about 28 s, with 14 serial 12K-token windows at about 2 s each.

Regarding computational cost, we believe that the overhead of scanning the history is a reasonable trade-off for VIMG. Specifically, as discussed in our previous responses, current image-generation queries are often short and highly contextualized, making direct query-based retrieval (e.g., conventional RAG) unreliable for identifying VIMG-related preference evidence scattered across seemingly unrelated historical interactions. This is reflected in the lower scores of the RAG and memory baselines compared with VIMGer. Achieving more reliable retrieval would require additional reasoning over the user's history to contextualize or expand the query, which itself incurs computational overhead. Our design therefore prioritizes directly accessing the original history to improve preference-evidence coverage. Moreover, VIMGer completes filtering, imagination, and rendering within a single 7B open-source UMM, whereas the frontier pipelines rely on a proprietary MLLM and a separate image generator, which keeps its deployment cost low.

Nevertheless, several engineering techniques could further reduce inference latency. For example, modern UMMs and LLMs support efficient, parallelized prefill for long inputs. Moreover, since user histories remain largely unchanged across successive queries, KV caching could allow previously processed history windows to be reused, avoiding redundant encoding and computation. We believe these optimizations offer promising opportunities to improve VIMGer's efficiency while retaining the benefits of directly scanning long interaction histories.

We sincerely thank you for your recognition of our work and for the time and effort you have devoted to reviewing our paper. We hope that our responses and additional empirical analyses have addressed your remaining concerns. We would be more than happy to continue the discussion and address any further questions you may have.

As the discussion period draws to a close, we would be grateful to hear whether our responses have addressed your concerns. We would also be happy to continue the discussion.

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

This paper proposes VIMG, a new text-to-image generation research task that leverages everyday histories between users and their personal MLLMs for personalized Visual Imagination. For this task, they also propose VIMG-Bench to evaluate VIMG, spanning six daily image-generation domains and 30 sub-domains, with a three-level hierarchy of visual imagination preferences. They utilize MLLM to score whether each personalized Visual IMaGination preference is correctly realized in the image as VIMG-Score. Based on VIMG-Bench, they finetune UMM Bagel through SFT and RL to obtain VIMGer with improved VIMG quality.

Critical strengths
  1. This paper presents a new text-to-image generation task utilizing the rich human-MLLM history for personalization.
  2. The paper constructs the whole evaluation system, including a systematic benchmark, evaluation metrics and a finetuned baseline for this task.
  3. Evaluation reflects the limitations of existing T2I models and UMM models to capture various dimensions of visual imagination preferences for a given persona emerged in long-term interactions.
Critical weaknesses
  1. Under both TSO and PPS settings, VIMGer performs worse than state-of-the-art proprietary models like GPT-Image 2 even after finetuned with in-domain data. A more detailed analysis of this performance gap would be valuable, particularly to understand whether it stems from limitations in the base model's generation capabilities or from the proposed personalization approach.
  2. For baseline, I would suggest adding one baseline that uses Bagel model to infer relevant user preferences from the interaction history first then generate the image images conditioned on both the original prompt and the inferred preferences.
  3. It would be great to analyze whether state-of-the-art MLLMs can accurately recover user preferences from the generated interaction histories. Such an analysis would help assess whether these histories provide sufficiently clear and consistent preference signals, and whether potential ambiguities or conflicts in the inferred preferences could affect personalized image generation.
Questions for authors
  1. What is the size for training and test data for VIMG-Bench? I think it should be mentioned in the main paper.
  2. Sometimes users would want some visual imagination with different style to their common preference. When user specify some new styles, will they contraindicate with their common preference?
  3. In line 234, the authors mention they develop 12 human-designed session templates that control the preference expression patterns. It would be great to include an example for the session template.
7 comments
AuthorsOpenReview ↗

Response to Reviewer MoqY (1/3)

We sincerely thank you for the thoughtful comments and constructive suggestions. We are particularly encouraged by your recognition of (1) VIMG as a new text-to-image task that leverages rich human–MLLM histories for personalization, (2) the complete evaluation system we construct, including a systematic benchmark, evaluation metrics, and a fine-tuned baseline, and (3) the value of our evaluation in revealing the limitations of existing T2I models and UMMs in capturing users' visual imagination preferences from long-term interactions. Below, we carefully address your remaining concerns.

[W1] Does the performance gap between VIMGer and proprietary models stem from the base model's generation capabilities or from the proposed personalization approach?

We sincerely thank you for this insightful question. We first clarify that VIMGer is positioned as a simple baseline that studies whether an open-source UMM can acquire VIMG capabilities through post-training. It is built on BAGEL-7B with LoRA and only 150 training instances, whereas GPT-Image-2 is a proprietary frontier system, for which an MLLM first reads the full history and writes the T2I prompt (Appendix G). Following your suggestion, we analyze where the remaining gap comes from. Our analysis below shows that it mainly comes from the base model's generation capability rather than from the personalization approach.

  1. The personalization approach closes most of the gap. We compare VIMGer with GPT-Image-2 on VScore-P (Table 2 in the main paper), which scores each preference individually. VIMGer narrows the gap between its BAGEL backbone and GPT-Image-2 from 25.33 to 1.48 points on average under TSO (14.29→38.1414.29 \to 38.14 vs. 39.62), and it is on par with or better than GPT-Image-2 in three of the six domains (Animal 35.69 vs. 35.25, Plant 43.02 vs. 42.75, Fashion 34.89 vs. 34.85). These results suggest that VIMGer can identify individual preferences at a level close to the frontier pipeline.

  2. The remaining gap lies in rendering many preferences jointly. We further compare the two on VScore-S, which requires all 4–6 preferences of a specification to be rendered jointly in one image. The remaining gap concentrates on this metric (3.74 vs. 9.55), which suggests that composing many fine-grained attributes in a single image remains harder for the 7B base generator.

  3. A stronger base model further improves VIMGer. Moreover, VIMGer is a post-training approach that is not tied to BAGEL. We therefore apply it to SenseNova-U1, a stronger open-source UMM, and report the average VScore-P below. With this stronger backbone, VIMGer reaches 41.21 under TSO and 44.21 under PPS, both above GPT-Image-2 (39.62 and 42.90). This shows that the same personalization approach closes the gap once the base model is stronger.

    MethodTSOPPS
    BAGEL14.2919.15
    + VIMGer38.1439.19
    SenseNova-U119.7227.61
    + VIMGer41.2144.21

Thank you again for helping us better understand where the remaining gap comes from.

AuthorsOpenReview ↗

Response to Reviewer MoqY (2/3)

(Part 2/3)

[W2] Suggest adding a baseline that uses BAGEL to first infer relevant user preferences from the interaction history, and then generates the image conditioned on both the original prompt and the inferred preferences.

Thank you for this suggestion. This baseline is the training-free counterpart of VIMGer, which also infers the user's preferences from the history before generating the image (Section 3.1). We therefore add it as Infer-then-Generate, which applies VIMGer's two-step prompts (Appendix K) to the original BAGEL without SFT or RL. The TSO VScore-P results are below.

MethodAnimalPlantPersonLandscapeFashionFoodAvg.
BAGEL (Base)11.2524.6710.4918.278.5812.4914.29
+ History16.8029.4212.0919.4712.6217.3317.96
+ RAG17.1123.1517.7320.138.8518.0917.51
+ Infer-then-Generate20.4133.6717.7923.6418.1122.8422.74
+ VIMGer35.6943.0239.0237.4734.8938.7538.14

Infer-then-Generate improves over History by 4.78 points on average (22.74 vs. 17.96), which confirms that the inference pattern itself is helpful. However, it remains well below VIMGer (38.14), because sparse, implicit, or image-carried evidence makes the reasoning pattern difficult to execute without task-specific training (Section 3.1). This gap reflects the contribution of PVIR-SFT and PVR-PO.

[W3] Can state-of-the-art MLLMs accurately recover user preferences from the generated histories? This would show whether the histories provide sufficiently clear and consistent preference signals, and whether ambiguities or conflicts in the inferred preferences could affect personalized image generation.

Thank you for your valuable question. We distinguish signal clarity, construction controls, and retrieval difficulty.

  1. Clarity of Preference Signals. In our post-hoc human validation (Appendix D.2), volunteers are asked to identify the visual imagination preference revealed in randomly sampled preference-revealing sessions, and their selections match our annotations with 91.2% accuracy (Imagination Fidelity, Table 5 in the appendix). This supports preference recognizability within individual evidence sessions.

  2. Consistency of Preference Signals. Conflicts are controlled throughout the construction. Specifically, (i) candidate specifications containing mutually incompatible or semantically conflicting values are removed (Appendix B); (ii) when a session is authored, the persona's other fixed preferences on the same subject are provided so that no competing value enters the scene, and every generated session is independently audited for entailment, stability, target conflicts, and distractor neutrality (Appendix K); and (iii) all histories are further audited and revised by humans (Section 2.1). For preferences that evolve over time, the newer value must unambiguously supersede the older one, and VIMGer and the frontier pipelines are instructed to follow the most recent one (Appendix K).

  3. Additional Preference Recovery Analysis. Nevertheless, to further address your concern, we evaluate whether state-of-the-art MLLMs can recover the ground-truth specification of each request. We consider two inputs, namely the oracle preference-relevant sessions only and the full history. The recovered preferences are matched against the ground-truth specification with the same LLM-judge protocol as RimagineR_\text{imagine} (Appendix E), and we report the recovery rates (%) at each level. Overall is weighted by the benchmark's 3,060 GVIP, 1,410 DVIP, and 10,590 SVIP assignments.

    MLLMInputGVIPDVIPSVIPOverall
    GPT-5.6-SolOracle sessions97.3596.2395.0595.63
    GPT-5.6-SolFull history85.8881.1375.4478.09
    Gemini-3.8-FlashOracle sessions96.7795.1094.1394.76
    Gemini-3.8-FlashFull history83.1677.7573.5375.88

    Given the oracle sessions, both MLLMs recover about 95% of the preferences. In contrast, the recovery rate drops to 75.88–78.09% with the full history, because preference-relevant sessions account for only 26.3% of all sessions (Figure 8 in the appendix) and those relevant to a single request are far fewer.

Overall, the histories provide clear and consistent preference signals, which state-of-the-art MLLMs recover accurately once the relevant sessions are located. Therefore, the main difficulty lies in locating the sparse personalized visual imagination evidence within long histories, and ambiguities or conflicts are unlikely to be what limits personalized image generation.

AuthorsOpenReview ↗

Response to Reviewer MoqY (3/3)

(Part 3/3)

[Q1] What are the sizes of the training and test data of VIMG-Bench? They should be mentioned in the main paper.

We apologize for the lack of clarity. The evaluation set contains 100 personas with 30 requests each, i.e., 3,000 instances under each of the TSO and PPS settings (Table 1 in the main paper), and each persona is paired with a long-term interaction history of about 168K tokens. The training set contains 150 instances from 5 additional personas (30 requests each), together with one additional held-out persona used only for SFT epoch selection (Section 3.1 and Appendix G). These six personas are disjoint from the 100 evaluation personas. We will state these sizes explicitly in the Implementation Details paragraph of Section 4 of the revised manuscript.

[Q2] Users sometimes want a style different from their common preference. When a user specifies a new style, will it contradict their common preference?

Thank you for this question. In VIMG, the current request takes priority, and the history is used only to complete the dimensions that the request leaves unspecified (Section 2). Therefore, a newly specified style replaces the common preference on that dimension rather than conflicting with it. Our PPS setting provides an initial attempt to examine this behavior, where the user states an immediate preference in the current query while leaving the remaining dimensions to be personalized from history (Section 2.2). Accordingly, VIMGer's imagination step is instructed to "keep everything the request states, and settle every visual choice it leaves open from the user's preferences", and the frontier pipelines receive an equivalent instruction (Appendix K).

To further address your concern, we conduct an additional preference-conflict experiment, where the current query explicitly specifies a preference value that contradicts one of the user's historical preferences. Due to the rebuttal time constraints, we conduct this experiment on 10 personas. We evaluate whether the generated images faithfully follow the newly specified preference, reporting the corresponding override accuracy (ACC). As shown in table below, VIMGer achieves high ACC, demonstrating that it can effectively prioritize the user's explicitly stated current intent over their historical preference.

MethodOverride ACC (%)
BAGEL + VIMGer (Ours)82.67

We believe this additional experiment further validates that VIMG personalization is not about rigidly following historical preferences, but about using history to complement the user's current intent.

[Q3] Line 234 mentions 12 human-designed session templates that control the preference expression patterns. Please include an example of a session template.

We sincerely thank you for this suggestion. A session template specifies how a preference is revealed within a session, while the dialogue topic specifies the everyday context in which the session unfolds. Specifically, a template prescribes the expression pattern: (i) explicit, where the user states the inclination; (ii) implicit, where the inclination is conveyed through the user's behaviors or choices without any explicit preference announcement; or (iii) evolving, where an earlier inclination is superseded by a newer one after a credible life change. In addition, each session is assigned an evidence carrier, i.e., the preference value is conveyed either in the text or only in a shared image, in which case the value never appears in the text of either speaker (Appendix K). One template is shown below.

Session Template IM-M-IMAGE-01

Expression:  implicit
Modality:    multimodal, value carried by image
Sessions:    1

History pattern:
  The user talks about the real-life experience and an image that naturally belongs to the current Scenario. The user does not directly state a preference conclusion. Instead, behavior, experience, feeling, or judgment already grounded in the Scenario, together with the image, provides clear and sufficient positive evidence that the user stably favors the target value depicted in the image.

For instance, in Figure 3 of the main paper, session (2) is an explicit, text-carried expression ("I always like its wings fully spread"), whereas in session (1) the value "collage art" can only be inferred from the attached image. We will add some templates with examples to Appendix J of the revised manuscript.

We are truly grateful for your encouraging feedback and thoughtful evaluation of our work. We sincerely hope that our clarifications and additional experiments have sufficiently addressed your remaining questions. We greatly appreciate your time and consideration, and warmly welcome any further discussion.

Reviewer_Urb4OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies Personalized Visual Imagination, generating an image from a short request while filling in visual preferences that the user has revealed in earlier interactions. VIMG-Bench contains 3,000 evaluation prompts for 100 synthetic personas across six domains, together with long multimodal histories that include both relevant evidence and distractor sessions. The proposed VIMGer model fine-tunes BAGEL in two stages. PVIR-SFT trains it to select relevant history and complete the request, and PVR-PO further optimizes history selection, preference inference, and image generation. In Tables 2 and 3, VIMGer substantially improves VScore-P and VScore-S over the BAGEL Base, History, and RAG variants, although its CLIP-T score is lower in many settings.

Critical strengths
  • The task is useful and distinct from the more common setting in which personalization is limited to subject identity or a small gallery of reference images.
  • The benchmark tests several kinds of visual preference at global, domain, and subdomain levels. Including distractor sessions and evaluating both Target Subject Only and Partial Preference Specificationrequests makes the retrieval and inference problem nontrivial.
  • The two training stages have clear roles, and Table 4 shows a further VScore-P gain when PVR-PO is added after PVIR-SFT in every reported domain.
  • The authors use persona-disjoint training and evaluation data and report that aggregate VIMG-Score is stable across five independent seeds. The reward judges used during PVR-PO also differ from the MLLM used for final evaluation.
Critical weaknesses
  • Most of the benchmark is generated through a closed synthetic pipeline: persona profiles are mapped to preferences, those preferences are embedded in generated histories, and the evaluated model is then asked to recover them. Human review improves quality control, but it does not establish that the task reflects ambiguous, conflicting, or changing preferences in real user histories.
  • The benchmark contains only 100 synthetic personas and relies on a predefined preference taxonomy. This makes it difficult to judge how well the reported gains would transfer to users whose preferences fall outside that taxonomy.
  • VIMG-Score uses a single evaluation MLLM, and the human-alignment study is small. Table 6 includes a Spearman correlation of 0.0513 for BAGEL Base on Animal and 0.2236 for VIMGer on Food, even though most other cells are much higher. Results from additional judges and a larger human study would make the metric more convincing.
  • VIMGer receives task-specific supervision from 150 instances built from five training personas, whereas the main BAGEL History and RAG comparisons do not receive comparable training. Together with the frequent drop in CLIP-T, this leaves open how much of the gain comes from the proposed personalization mechanism rather than additional supervision and a shift away from explicit prompt adherence.
Questions for authors

Please refer to the Weaknesses section.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Urb4 (1/3)

We thank you for recognizing the task's usefulness, the benchmark's challenging preference evidence, the two training stages, and our evaluation controls. We address your concerns below.

[W1] Discussion on the trade-off between synthetic data realism and evaluation rigor in VIMG-Bench.

We sincerely thank you for this question. We first emphasize that VIMG addresses a realistic and emerging scenario. As MLLMs are increasingly adopted as personal assistants, users naturally reveal how they imagine visual content through everyday interactions, and expect their later image-generation requests to reflect these preferences. However, a dedicated benchmark for systematically evaluating this capability is still lacking, which motivates our work. Moreover, to ground our benchmark in everyday usage, we construct base prompts around common image-generation requests across six daily domains, while grounding the interaction sessions in conversational topics derived from real-world human–MLLM interaction benchmarks, rather than creating scenarios entirely from scratch. Below, we elaborate on the rationale behind our construction and how it addresses the properties you mention.

  1. Practical Constraints. As discussed in our Limitations (Appendix H), collecting real users' long-term multimodal assistant histories at scale is impractical under strict privacy constraints, which is a broadly shared limitation in personalization research (e.g., benchmark Persona-MME[1]). Moreover, real histories provide no ground truth for what a user has in mind on the dimensions that a request leaves open. A controlled construction is therefore what enables VIMG-Bench to verify each preference individually.

  2. Real-World Grounding. To narrow the gap to real-world usage, we anchor multiple stages of the pipeline in real-world sources (Section 2.1). Specifically, the 100 personas are sampled from PersonaMem-v2, the 18,276 dialogue topics are drawn from real-world human–MLLM interaction benchmarks, the preference expression patterns follow 12 human-designed templates, and each session is authored from only a single preference value rather than the full specification. Moreover, all histories are audited and revised by humans, and our post-hoc human evaluation reports 75.0% interaction realism, 82.5% revelation naturalness, and 91.2% imagination fidelity (Table 5 in the appendix).

  3. Ambiguous, Conflicting, and Changing Preferences. More importantly, VIMG-Bench explicitly models these three properties (Appendices B and K), as listed below.

    • Ambiguous. Preferences are also revealed implicitly, without any explicit preference announcement, and some are conveyed only through shared images, where the value never appears in the text.
    • Conflicting. Substantial distractor sessions involve the same domains and discuss related attributes, competing values, or even identical preference values without giving positive evidence for the target preference, so that keyword- or domain-level shortcuts fail.
    • Changing. For evolving preferences, an earlier inclination is superseded by a newer one after a credible life change, and the model has to follow the most recent one. Evolving evidence appears in the histories of most personas.

    In addition, preference-relevant sessions account for only 26.3% of all sessions (Figure 8 in the appendix), and the relevant sessions of one specification span a median of 71% of the history (Figure 9 in the appendix).

We sincerely thank you for this question and for recognizing our efforts in constructing VIMG-Bench through real-world interaction grounding and extensive human involvement, while respecting privacy constraints.

[W2] The benchmark contains 100 personas and relies on a predefined preference taxonomy, so how well the reported gains would transfer to users whose preferences fall outside that taxonomy.

We thank you for this question.

  1. Benchmark Scale. Although VIMG-Bench contains 100 personas, each persona is evaluated on 30 distinct image-generation requests, yielding 3,000 evaluation instances per setting, each associated with long-term multimodal interaction histories (168K interactions in total). More importantly, our preference taxonomy is substantially broader than the number of personas, comprising 438 preference dimensions and 2,085 distinct preference values across three hierarchical levels (Table 1), allowing diverse and fine-grained personalized specifications even for the same subject (Figure 7 in the appendix). We maintain the current scale to enable thorough human inspection and verification of individual preference dimensions, values, and their corresponding historical evidence. Nevertheless, we fully acknowledge the value of broader coverage and will further expand both the persona population and preference taxonomy in future work. We sincerely thank you for this constructive suggestion.
AuthorsOpenReview ↗

Response to Reviewer Urb4 (2/3)

(Part 2/3)

  1. Role of the Taxonomy. The taxonomy makes each preference verifiable by VQA and aims for broad coverage. The GVIP dimensions follow prior work (Xu et al., 2026), while the DVIP and SVIP dimensions are derived for each subject from LLM knowledge and external sources such as Wikipedia with human refinement, covering both intrinsic subject attributes and attributes arising from common human interactions with the subject (Appendix B).

  2. Taxonomy-Agnostic Inference. Importantly, no method accesses the taxonomy at inference. VIMGer filters sessions and writes a free-form specified prompt under open-ended instructions that never enumerate dimensions or candidate values (Appendix K). Therefore, it is not restricted to predefined preferences.

4. Transfer to Unseen Preferences. Our current training data consists of 150 instances from 5 personas, covering 498 unique dimension–value pairs (only 23.88% of the 2,085 preference values in our taxonomy). Moreover, the training personas are strictly disjoint from the evaluation personas (Appendix G). Thus, our existing evaluation already provides evidence of generalization to unseen users. Nevertheless, to further examine whether VIMGer can generalize to preference values not encountered during training, we conduct an additional analysis by partitioning the evaluation preference annotations into two groups based on their exact dimension–value pair membership in the training data: seen preferences, whose dimension–value pairs appeared during training, and unseen preferences, whose pairs never appeared during training. The results are as follows:

MethodSeen PreferencesUnseen Preferences
BAGEL + History18.7117.44
+ VIMGer39.3937.29

VIMGer improves over BAGEL + History by 20.68 points on seen preferences and 19.85 points on unseen preferences. These comparable gains suggest that VIMGer learns to infer personalized preferences from interaction histories, rather than merely memorizing preference values encountered during training.

We sincerely thank you for this insightful suggestion. We will also include expanding the preference taxonomy as an important future work in the revised Appendix H.

[W3] VIMG-Score uses a single evaluation MLLM, and the human-alignment study is small. Table 6 includes a Spearman correlation of 0.0513 for BAGEL Base on Animal and 0.2236 for VIMGer on Food. Additional judges and a larger human study would make the metric more convincing.

We sincerely thank you for the careful reading of Table 6 in the appendix and for this valuable suggestion. We address your concerns below.

  1. The Two Low-Correlation Cells. The human-alignment study covers 90 images and 435 questions in total, so each correlation in Table 6 is computed over only 5 images. With so few items, a single tie or swapped pair changes ρ\rho substantially. In fact, the question-level agreement with humans in these two cells remains 76.92% and 88.00%. Overall, across the 18 cells of Table 6, Q-Agree averages 92.0%, AUROC averages 91.6, and 16 cells have ρ≥0.66\rho \geq 0.66.

  2. Additional MLLM Judge. We sincerely appreciate the reviewer’s suggestion. We would first like to clarify that evaluating VIMG is relatively well-constrained by our evaluation design. Specifically, (1) each evaluation is grounded in a predefined ground-truth visual imagination specification, which explicitly identifies the target attributes to be evaluated; and (2) we construct a separate VQA question for each preference, rather than asking the judge to assess multiple preferences simultaneously. This makes the evaluation more objective and less dependent on the judge's subjective interpretation.

    Nevertheless, to further address your concern, we re-score the same generated images using an additional judge from a different model family, Gemini-3.8-Flash. The average TSO VScore-P results are reported below.

    MethodDeepSeek-V4.1-Flash (Paper)Gemini-3.8-Flash
    GPT-Image-239.6239.01
    BAGEL (Base)14.2913.88
    + VIMGer38.1437.67

    The ranking of all three methods remains identical across the two judges, and every score changes by less than one point. These results further demonstrate that VIMG-Score is relatively robust across different MLLM judges, and that our main conclusions do not depend on the choice of evaluation model. We sincerely thank you for this helpful suggestion.

Following your suggestion, we will further expand the scope of human evaluation and include this as an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer Urb4 (3/3)

(Part 3/3)

[W4] VIMGer receives task-specific supervision from 150 instances built from five training personas, whereas the BAGEL History and RAG baselines receive no comparable training. Together with the frequent drop in CLIP-T, how much of the gain comes from the personalization mechanism rather than from additional supervision and a shift away from explicit prompt adherence?

We sincerely thank you for this insightful question. We address the two factors separately.

  1. Personalization Mechanism vs. Additional Supervision. We first note that VIMGer is by design a post-training baseline (Section 3), and its supervision is deliberately small, with 150 instances from 5 personas. These personas are disjoint from the evaluation personas (Appendix G). Moreover, PVR-PO reuses the same 150 instances without any additional data (Section 3.2), yet it adds 3.69 points of average VScore-P over PVIR-SFT (34.45→38.1434.45 \to 38.14, Table 4 in the main paper), a gain that cannot come from more supervision.

    To further disentangle the two factors, we add two variants. History-SFT fine-tunes the History baseline on the same 150 instances with the same image supervision, but without the PVIR reasoning chain. PVIR w/o Training applies the two-step reasoning of VIMGer to the original BAGEL without any SFT or RL. The TSO VScore-P results are below.

    MethodSame 150 instancesPVIR reasoningVScore-P
    BAGEL + History✗✗17.96
    History-SFT✓✗25.99
    PVIR w/o Training✗✓22.74
    PVIR-SFT✓✓34.45
    PVR-PO (VIMGer)✓✓38.14

    With the same 150 instances, History-SFT reaches only 25.99, which is 8.46 points below PVIR-SFT (34.45), because fitting target images directly from long and noisy histories does not teach the model which sessions matter or how to complete the unspecified personalized visual imagination dimensions. Meanwhile, PVIR w/o Training reaches 22.74, since the untrained model cannot reliably execute this reasoning. Taken together, these results suggest that the substantial gains arise from effectively learning the PVIR reasoning pattern through task-specific post-training, rather than merely from exposure to additional supervision (as History-SFT significantly underperforms VIMGer).

  2. CLIP-T and Explicit Prompt Adherence. We apologize for the lack of clarity. We first clarify that CLIP-T is computed between the generated image and the user's original short request (Appendix G). A personalized image additionally depicts 4–6 preferences that the request does not mention, so its similarity to the short request naturally decreases. This also holds for frontier systems. Specifically, under TSO, the average CLIP-T is 19.55 for BAGEL Base, 17.67 for GPT-Image-2, 17.80 for Nano-Banana-2, and 17.70 for VIMGer (Table 2). Under PPS, the corresponding scores are 20.51, 18.43, 18.51, and 18.42, respectively (Table 3). Thus, VIMGer is on par with frontier systems and consistently higher than the History baseline (17.48 and 18.11), suggesting that the CLIP-T decrease reflects the addition of personalized content rather than reduced adherence to the original request.

    Furthermore, VIMGer explicitly instructs the model to "keep everything the request states" (Appendix K). To directly verify this behavior, our additional preference-conflict experiment (see our response to Reviewer MoqY [Q2]) shows that VIMGer can effectively prioritize newly specified preferences over conflicting historical preferences. Together, these results suggest that VIMGer enriches image generation with personalized details while preserving the user's current intent.

We sincerely appreciate your careful review, constructive feedback, and recognition of the contributions of our work. We hope that our detailed responses and supplementary empirical results have satisfactorily addressed your concerns. Thank you again for your valuable time, and we look forward to any further discussion.

Reference

[1] PersonaVLM: Long-Term Personalized Multimodal LLMs, CVPR 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

VIMG · Rebuttal Summary

Dear ICLR 2027 Conference Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely thank you for the time and care devoted to reviewing our paper. As the discussion period draws to a close, we would like to briefly summarize the reviewers' assessments and the additional evidence and revisions we have provided in response to their feedback.

Our newly proposed research task, personalized visual imagination (VIMG), investigates how image generators can leverage preferences that users naturally reveal through everyday interactions with personal MLLM assistants to complete the visual details left unspecified in their image-generation requests. Reviewer Urb4 described VIMG as "useful and distinct", highlighting how VIMG-Bench makes preference inference challenging through distractor sessions and two settings with different levels of specification in user requests. Reviewer MoqY recognized the value of our systematic benchmark, evaluation metrics, and trained baseline. VIMG-Bench covers six domains and 30 subdomains, organizing visual imagination preferences at global, domain, and subdomain levels. Reviewer DZQH also appreciated our efforts to extend interaction-history-based personalization benchmarks to image generation.

Throughout the rebuttal, we conducted additional experiments and provided further clarifications to address the reviewers' concerns:

1. Clarifying the capability gap and preserving current user intent (Reviewer MoqY). We clarified that the difficulty discussed in W2 concerned vanilla BAGEL executing VIMGer's reasoning instructions without task-specific training. Under TSO, its VScore-P increased from 22.74 to 38.14 after VIMGer training. Furthermore, we added vanilla BAGEL with history as a comparison in the preference-conflict experiment. On the same requests, VIMGer followed the newly specified preference in 82.67% of cases, compared with 81.67% for vanilla BAGEL.

2. Strengthening the empirical validation (Reviewers DZQH and Urb4). We added Mem0 and MemOS as memory-based baselines and conducted controlled comparisons of different history selectors with frontier image generators, keeping the prompt writer and image generator fixed. VIMGer achieved the highest VScore-P among the tested selectors with both GPT-Image-2 and Nano-Banana-2. We also compared VIMGer against a History baseline trained on the same 150 instances to isolate the contribution of task-specific reasoning supervision, evaluated generalization to preference values unseen during training, and introduced an additional MLLM judge to verify the robustness of VIMG-Score.

3. Improving the manuscript and benchmark documentation (Reviewers DZQH and MoqY). We incorporated the additional memory-based baselines into Tables 2 and 3, clarified the training and evaluation data sizes in Section 4, and provided two representative session templates with examples in Appendix J.

Reviewers Urb4 and MoqY both rated our paper Weak Accept. After our latest reply, Reviewer MoqY confirmed that all of their concerns had been addressed and that they would keep their initial positive rating.

We are committed to releasing the full VIMG-Bench to the research community to facilitate future research on personalized visual imagination.

Thank you again for your time, thoughtful consideration, and constructive feedback.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.