cs.CVApr 8, 2026
SavePersonalizing Text-to-Image Generation to Individual Taste
Organizations: Tübingen AI Center, University of Tübingen · Department of Brain and Cognition, KU Leuven
Abstract
Modern text-to-image (T2I) models generate high-fidelity visuals but remain indifferent to individual user preferences. While existing reward models optimize for "average" human appeal, they fail to capture the inherent subjectivity of aesthetic judgment. In this work, we introduce a novel dataset and predictive framework, called PAMELA, designed to model personalized image evaluations. Our dataset comprises 70,000 ratings across 5,000 diverse images generated by state-of-the-art models (Flux 2 and Nano Banana). Each image is evaluated by 15 unique users, providing a rich distribution of subjective preferences across domains such as art, design, fashion, and cinematic photography. Leveraging this data, we propose a personalized reward model trained jointly on our high-quality annotations and existing aesthetic assessment subsets. We demonstrate that our model predicts individual liking with higher accuracy than the majority of current state-of-the-art methods predict population-level preferences. Using our personalized predictor, we demonstrate how simple prompt optimization methods can be used to steer generations towards individual user preferences. Our results highlight the importance of data quality and personalization to handle the subjectivity of user preferences. We release our dataset and model to facilitate standardized research in personalized T2I alignment and subjective visual quality assessment.
Figures & tables
Figure 1: Prior work on preference alignment focuses on steering generations towards samples with a global reward. In contrast, we are able to steer samples with prompt optimization [ 38 , 1 ] to generate different samples tailored to individual taste.
| Dataset | Year | Label | # Ratings | # Images | # Users | Ratings per Image | User-Level Labels | Subjective Domains | Image Source |
| Classical IQA Datasets | |||||||||
| AVA [ 39 ] | 2012 | Score | 255K | 255K | 25K | 200 | ✗ | ✗ | Real photos |
| LIVE [ 50 ] | 2006 | Score | 779 | 779 | 29 | 1 | ✗ | ✗ | Real photos |
| KADID-10K [ 32 ] | 2019 | Score | 30K | 10.1K | 25 | 3 | ✗ | ✗ | Distorted |
| AI-Generated IQA Datasets | |||||||||
| SAC [ 43 ] | 2022 | Score | 238K | 238K | Crowd | 1 | ✗ | ✗ | AI-generated |
Table 1: Dataset comparison across IQA and human preference benchmarks. PAM LA is the only dataset combining AI-generated images, subjective visual domains, dense per-image multi-rater coverage, and user demographics. These properties are jointly required for personalized reward modeling. Label denotes the annotation format: Score (absolute ratings) or Pairwise (comparative preferences of 2 images) or Ranking (ordinal ranking of multiple images). † Note that Pick-a-Pic v2 includes user IDs but was not designed for per-user analysis; per-user splits must be reconstructed post hoc.
Figure 2: Visual diversity in the PAM LA benchmark. The dataset spans two primary domains: Art and Photography. It comprising 21 distinct thematic categories as shown with examples. This structure isolates a model’s ability to judge stylized artistic compositions from its ability to evaluate real-world, photographic subjects.
Figure 3: Architecture of our personalized aesthetic quality predictor. Visual and semantic features are extracted by a frozen SigLIP2 encoder; user demographic and image metadata embeddings are produced by a frozen embedding encoder. All features are projected to a shared dimension, assembled as a token sequence with a learnable [CLS] token, and fused by a shallow transformer encoder. The [CLS] output is passed to a linear head to predict the aesthetic score and trained with a mean squared loss.
| Reward Model | User SROCC | Avg SROCC | User PLCC | Avg PLCC | User pw acc | Avg pw acc |
| LAION [ 49 ] | 0.1516 | 0.1511 | 0.1471 | 0.1426 | 0.5110 | 0.5156 |
| Q-Align (IQA) [ 55 ] | 0.2497 | 0.3290 | 0.2416 | 0.3244 | 0.5865 | 0.6096 |
| Q-Align (Aesthetics) [ 55 ] | 0.2677 | 0.3273 | 0.2906 | 0.3606 | 0.5932 | 0.6109 |
| ImageReward [ 58 ] | 0.2841 | 0.2314 | 0.2855 | 0.2044 | 0.5978 | 0.5762 |
| DeQA [ 60 ] | 0.2371 | 0.2864 | 0.2105 | 0.2741 | 0.5818 | 0.5950 |
| HPSv3 [ 36 ] | 0.4019 | 0.5076 | 0.4444 | 0.5880 | 0.6427 | 0.6773 |
Table 2: Reward model comparison on the PAM LA test set with held-out users. Our model outperforms all baselines across both evaluation regimes (user level vs population average) and all three metrics (SROCC, PLCC, pairwise accuracy).
Figure 4: Comparison of iterative prompt refinement steered with PAM LA compared to HPSv3 [ 36 ] and Q-align [ 55 ] . Unlike global reward models, prompt optimization with PAM LA leads to more natural and appealing samples without reward-hacking.
Figure 5: Comparison of iterative prompt optimization results for three different users with our PAM LA predictor, HPSv3 [ 36 ] and Q-Align [ 55 ] . Unlike existing reward models, we are able to customize the generation to individual user taste.
Figure 6: Illustration of iterative prompt refinement for a fictional image prompt. While other reward models steer towards surreal outputs, PAM LA favors photorealism.
Figure 7: Comparison of iterative prompt refinement steered with PAM LA for different age ranges. Our PAM LA predictor remains agnostic to user id and instead relies on learned patterns from the demographic embeddings.
Figure 8: Preference ranking of optimized images. We find that users prefer images optimized to their taste using PAM LA, and that users prefer images optimized with our method in general. Popular reward models like HPSv3 [ 36 ] and Q-Align [ 55 ] seem to degrade perceptual quality and are chosen less than the unoptimized image.
Figure 9: Comparison of image steering outcomes across two distinct runs for User 1. We find that the LLM learns consistent patterns in the prompt, leading to a consistent pattern of steered images.
Figure 10: Examples of image pairs from our PAM LA test set where users differed in their preferences. Each image and user pair shows 2 images where user x preferred image A over image B and user y preferred image B over image A. The ground truth score and model prediction are plotted below each image.
Figure 11: We vary the margin considered to be a tie while calculating pairwise accuracy. We note that increasing the threshold upto the next anchor point in the annotation scale gives consistent improvements to the pairwise accuracy, suggesting that the relatively low pairwise accuracy stems from the large number of samples with similar scores.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Seen Users (Avg) | Unseen Users (Avg) | Generalization Gap | |||||||
| Config | SROCC | PLCC | PairAcc | SROCC | PLCC | PairAcc | SROCC | PLCC | PairAcc |
| Full model | 0.5563 | 0.5887 | 0.7311 | 0.4975 | 0.5183 | 0.7057 | 0.0588 | 0.0704 | 0.0253 |
| No text | 0.5922 | 0.6200 | 0.7460 | 0.4788 | 0.4981 | 0.6981 | 0.1135 | 0.1219 | 0.0480 |
| No demographics | 0.5734 | 0.6006 | 0.7381 | 0.4896 | 0.5126 | 0.7029 | 0.0837 | 0.0880 | 0.0353 |
| No metadata | 0.5639 | 0.5933 | 0.7333 | 0.4544 | 0.4835 | 0.6872 | 0.1095 | 0.1098 | 0.0461 |
| No demo+meta | 0.5676 | 0.5985 | 0.7340 | 0.4723 | 0.4989 | 0.6932 | 0.0953 | 0.0995 | 0.0408 |
Table 3: Ablation study results on seen and unseen users. Unseen users are matched via kNN ( =15, top- =5, =0.1). The generalization gap is defined as the difference between seen and unseen average metrics; a smaller gap indicates better generalization to new users. Best values per column are in bold ; the full model row is highlighted.
Figure 12: Effect of number of context images and Top- neighbors on unseen user performance for PAM LA. Performance peaks at , (circled) across all metrics, with diminishing returns beyond this point. Temperature throughout.
| PAM LA | ||||
| N images | Top-K | SROCC | PLCC | PairAcc |
| 5 | 1 | 0.3917 | 0.4224 | 0.6406 |
| 5 | 5 | 0.3829 | 0.4023 | 0.6369 |
| 10 | 1 | 0.4136 | 0.4353 | 0.6474 |
| 10 | 5 | 0.3880 | 0.4230 | 0.6383 |
| 15 | 1 | 0.4430 | 0.4633 | 0.6601 |
Table 4: Comparison of configurations to handle unseen users. All results are reported for temperature = 0.1.
Figure 13: Example trial of our rating study to collect the user annotations in our PAM LA dataset. Given the prompt and image, we allow the user to select the exact score using a continuous slider, with 5 anchor points along the scale.
Figure 14: Example trial of our user study to validate our PAM LA reward model for image steering. Users were shown two images side-by-side and were simply asked to indicate which image they preferred. The user study compared images optimized to the user’s taste with PAM LA, images optimized for other users with PAM LA, images optimized with generic reward models (HPSv3 and Q-Align) and unoptimized images. Users evaluated all possible pairwise comparisons.
Figure 15: Comparison of image steering outcomes across two distinct runs for User 2. We find that the LLM learns consistent patterns in the prompt, leading to a consistent pattern of steered images.
Figure 16: Comparison of image steering outcomes across two distinct runs for User 3. We find that the LLM learns consistent patterns in the prompt, leading to a consistent pattern of steered images.