Personalized Visual Content Generation in Conversational Systems

We sincerely thank you for the overall positive evaluation and rating of our work !!! It is our honor to discuss our research with you~~

---

W1.
Your concern directly aligns with one of our optimization goals — the fidelity of generated content. If the generated image deviates too far from the original target item, the meaning of personalization diminishes.
As shown in our results (e.g., Fig. 1 — the first row “Waterboy” and the last row “Taken”), the generated outputs match the target items well, and objective metrics also confirm PiViC’s strong fidelity performance.

That said, due to the inherent limitations of current generative models, achieving complete fidelity remains challenging. We would like to emphasize that PiViC is a complete personalization pipeline — if the base model itself achieves perfect fidelity, our method can also deliver fully faithful personalization on top of it. Since fidelity is a fundamental challenge in image generation, we believe this is an area where we, other researchers, and the broader community must work together.

---

W2.
We have added an additional image-to-image baseline (Kandinsky) for reference:

PiViC achieves superior performance in both personalization and fidelity.

We also provide comparisons with closed-source products (GPT‑4o, Gemini 2.5‑Pro):

PiViC offers a better trade-off between personalization and fidelity, with a clear improvement in integrating historical liked items.

---

W3.
Indeed, as shown in our examples, Superbad (third row) integrates more background elements from the user’s historical styles (in softer tones), Taken (last row) combines the protagonist’s professional and composed demeanor with the forward-facing pose found in historical liked items, and Waterboy (first row) incorporates distinctive background motifs from historical inclinations. These examples illustrate PiViC’s ability to not only preserve fidelity to the target item but also seamlessly embed personalized elements drawn from historical interests.

---

W4.
Thank you for the suggestion. As an application-driven work, our primary contribution is proposing a novel pipeline to address the problem of monotonous content presentation in conversational systems.
Our goal is to present the pipeline as a holistic solution rather than focusing solely on model-architecture-level innovations. Given that current generative systems are heavily dependent on pretrained base models, purely architectural changes may not be accessible or impactful for all researchers.

From a pragmatic perspective, we aim for PiViC to serve as a starting point for personalized visual generation in conversational systems — inspiring more innovative, competitive, and groundbreaking work built upon it, and contributing fresh ideas to the community.

We sincerely appreciate your positive rating of our work, as well as the encouragement from the other reviewers. Your feedback has been a tremendous source of motivation. We look forward to building on these insights to further refine GenAI so that it can better serve users’ individual needs. Wishing you continued success in your work and good health.

Best regards,

The Authors

We greatly appreciate your high recognition of our work — it motivates us to continue improving. We are happy to address your questions.

---

W1.
This is indeed the direction we aim to explore next. However, as you can see, many current unified multimodal understanding–generation models (e.g., Janus [1]) still decouple multimodal understanding and image generation (even if they reside within the same model). We expect and believe that future designs will build upon our work with more integrated approaches.

[1] Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation

---

W2.
We have attempted such experiments, but we found that participants in the human study often scored images according to their pre-existing personal preferences, which could bias the evaluation.
Therefore, we ultimately chose to use objective metrics and GPT-based evaluation to assess generation quality.

---

W3.
Thank you for the suggestion. Other reviewers also mentioned this point. In the final version, we will include more important experiments — including the ablation study — directly in the main text.

---

Q1.
Our evaluation focuses on movie datasets because our work is designed for conversational systems, and to date, almost all such systems are built on two movie datasets [1–5], without adoption of other domains.

[1] Parameter-Efficient Conversational Recommender System as a Language Processing Task, EACL’24
[2] Multi-Type Context-Aware Conversational Recommender Systems via Mixture-of-Experts, Arxiv’25
[3] Towards Unified Conversational Recommender Systems via Knowledge-Enhanced Prompt Learning, KDD’22
[4] COLA: Improving Conversational Recommender Systems by Collaborative Augmentation, AAAI’23
[5] A Large Language Model Enhanced Conversational Recommender System, Arxiv’23

After a more careful investigation, we found that other niche conversational recommender datasets do not provide both historical and target item images, which makes them fundamentally incompatible with PiViC’s setting. We hope this clarifies our choice.

---

Q2.
We have provided more direct analyses of generated results in the Supplementary Material.zip (Appendix). Please refer to it for details. In the final version, we will also consider adding more intuitive qualitative comparisons of generated results.

We sincerely thank you for reviewing our paper and for your positive recognition of its clarity and novelty. We are pleased to address your questions and hope this will persuade you to raise your rating for our work~~

W1.
Your suggestion is excellent. In the initial design, we did attempt to use VLMs as an analyzer to obtain users’ inclinations (by feeding both the dialogue and the mentioned movie poster images into the model) to achieve more stable user modeling. However, this approach required significantly more resources and inference time compared to PiViC, making it impractical.
Nevertheless, we have supplemented our work with a comparison experiment using another open-source generative model (Kandinsky) on the E-Redial dataset as an alternative baseline:

This further demonstrates PiViC’s effectiveness.

---

W2 & L1.
Your proposed experiment is indeed very nice, and we are happy to accept it.
Since GPT-4o and Gemini cannot generate images via API, we used the official GPT-4o platform and the Gemini 2.5 Pro subscription version to generate images. As you know, this collection process is highly time-consuming and labor-intensive, so we conducted the experiment on the E-Redial dataset.
The prompt used for both GPT and Gemini was:

Generate a movie poster of xxx based on the user interest shown in the below dialogue: \n <dialogue>

Due to platform restrictions, some images could not be generated (likely because of prohibited content), with reminders such as:

I can't generate that image because the request violates our content policies.
Please feel free to share a different idea or request, and I'd be happy to help!

This resulted in 11 test set images for GPT-4o being excluded from evaluation. As closed-source models, both have been fully trained on target items, so we focus on their ability to integrate historical interests.
Due to this year’s rebuttal policy, we cannot share generated images, but we believe the objective metrics below can largely address your concerns:

We can see that while GPT-4o produces highly faithful posters, its ability to incorporate historical information is inferior to our proposed PiViC. This suggests these models have limited capacity for understanding and integrating conversational information.
Gemini’s fidelity and personalization are weaker than GPT-4o’s, a finding consistent with reports from some existing benchmark sites. These results further highlight PiViC’s strength in reflecting user interests in conversation.

---

Q1.
This is an excellent question. Although we have not observed this phenomenon in our existing datasets, it is a valuable point.
For such cases, an inclinations register can be maintained to store users’ long-term interests. When a shift occurs, both the stored inclinations and new ones can be analyzed by the analyzer, reducing the analysis length and allowing quick decoding of updated inclinations.
Moreover, since LoRAs learn the ability to integrate historical liked items, they can still adapt when new inclinations appear.

---

Q2.
Scalability involves two aspects:
(1) Analyzing inclinations in long dialogues, and
(2) Handling a huge number of historical items.

For long dialogues, our Q1 solution applies.
For the dual-stage LoRA (handling historical & target images), the target image is always a single one, and regardless of the number of historical images, we concatenate them before conditioning. This extra cost lies only in the concatenation step (no extra computational overhead) and is negligible. Therefore, PiViC is scalable.

---

L2.
We share your concern. We tried to find a conversational dataset with explicit feedback (e.g., ratings or clicks), but due to dataset limitations, we only obtained dialogue-based signals.
If you could point us to a conversational recommender dataset with direct supervision signals like clicks/ratings, it would indeed highlight the merit of our PiViC. We also plan to explore this direction and will incorporate click/rating signals in future versions. In fact, click/rating data could directly score inclinations, serving as a highly reliable reflection of user interests.

Sincere thanks for your suggestion! : )

We are grateful for your recognition of our work, and we are very happy to answer your questions : )

W1 & Q1.
Our evaluation is conducted on movie datasets because our work targets conversational recommender systems, and so far, almost all conversational recommender systems are built on two movie datasets without using other datasets [1-5].
[1] Parameter-Efficient Conversational Recommender System as a Language Processing Task, EACL'24
[2] Multi-Type Context-Aware Conversational Recommender Systems via Mixture-of-Experts, Arxiv'25
[3] Towards Unified Conversational Recommender Systems via Knowledge-Enhanced Prompt Learning, KDD'22
[4] COLA: Improving conversational recommender systems by collaborative augmentation, AAAI'23
[5] A Large Language Model Enhanced Conversational Recommender System, Arxiv'23

Through a more detailed investigation, we found that other niche conversational recommender system datasets do not contain history and target item images, which creates a large gap from the PiViC setting. We hope you can understand this difference. Nevertheless, we have additionally included results on the E-Redial dataset, using closed-source models (GPT-4o, Gemini 2.5-Pro) for objective metric comparison, for your reference (in W2 & Q2).

---

W2 & Q2.
Using LoRAs indeed increases inference latency to some extent. However, considering real-time applications, we compared our method with mature commercial products (e.g., GPT-4o and Gemini 2.5-Pro):

As shown, our runtime is within an acceptable range compared to the other two. In fact, we have already initiated new work to optimize the generation process via KV cache, since the same target item images are often used multiple times as keys and values during conditioning. Please stay tuned for further improvements!

Additionally, we include the comparison of generation quality with these closed-source products:

We can see that PiViC achieves a better balance between personalization and fidelity. GPT-4o generates highly faithful images, but still lags behind PiViC in personalization.

---

W3.
Thank you for your suggestion. Our ablation study is placed in the Supplementary Material.zip (Section J.2, to be moved into the main paper in the final version), which might be easy to overlook. The ablation study covers four aspects: Sequence Invariance, Global LoRA, Negative Prompt, and Score Reweight, along with objective metric comparisons. For the latter three, we also provide intuitive visual comparisons of the generated results.

---

Q3.
This experiment is indeed very interesting and highly valuable for understanding the internal mechanisms of LoRA. Unfortunately, due to this year's conference guidance, which strictly forbids authors from updating images in the repository, we cannot include them at this stage. However, we will add a discussion on this part in the final version.

Thank you very much for your in-depth suggestions!