Modeling Subjective and Personalized Aesthetic Preferences

495 concepts345 papers2 proposals

When machine learning models evaluate or generate creative content—whether it is a retouched photograph, a jazz solo, or an interior design layout—they traditionally optimize for a single target: the global average. By training on large-scale datasets annotated with crowdsourced ratings, models learn to predict the consensus or the "mean" human judgment.

However, aesthetics, style, and creative expression are fundamentally subjective. A single global average inevitably flattens the rich diversity of human taste. Optimizing for the mean results in "cookie-cutter" outputs that fail to satisfy individual users, particularly those with niche, avant-garde, or highly specific professional preferences.

To bridge this gap, a growing body of research focuses on Personalized Subjective Modeling. Instead of treating aesthetic evaluation as an objective problem with a single right answer, this paradigm treats preference as a multi-dimensional, user-specific space.

The Fallacy of the Global Average

To understand why personalized modeling is necessary, consider the task of ranking images. Imagine trying to compare a vibrant, high-contrast sunset against a gritty, low-key black-and-white street portrait. On a single universal scale of 1 to 10, which one is objectively "better"?

As noted by Yubin Deng in Image Aesthetic Assessment: An Experimental Survey, standard aesthetic assessment models rely heavily on global averages derived from aggregate human ratings. This approach assumes a monolithic standard of beauty. In reality, as Katharina A. Schwarz et al. demonstrate in Will People Like Your Image? Learning the Aesthetic Space, models trained on aggregate crowd preferences converge on a single global standard, completely missing the idiosyncratic variations that define individual taste.

This reliance on a single standard causes significant issues across various applications:

  • Image Enhancement: Traditional automated retouching tools often target a single expert style. However, Zhangkai Ni et al. point out in Towards Unsupervised Deep Image Enhancement With Generative Adversarial Network that a single expert's aesthetic bias rarely aligns with the diverse preferences of general users.
  • Tone Mapping: Compressing high dynamic range (HDR) data into standard displays is inherently subjective. As Chien-Chuan Su et al. argue in Explorable Tone Mapping Operators, fixed one-to-one mapping algorithms fail because different users prioritize different attributes, such as high local contrast versus a completely natural appearance.
  • Professional Workflows: Even in highly structured domains like medical coding, human experts exhibit individual professional styles. Yue Gao et al. emphasize in Optimising the paradigms of human AI collaborative clinical coding that human-in-the-loop systems must model these subjective variations to properly align AI suggestions with individual practitioner workflows.

Mechanisms for Capturing Individual Taste

If global averages fail, how can a computational system capture something as elusive as individual taste? Researchers have developed several core mechanisms to inject personalization into deep learning frameworks.

1. Behavioral and Soft Embeddings

Instead of forcing users to explicitly define their style through complex menus or rules, models can implicitly learn preferences from historical user behavior.

For example, in Personalized Multimodal Response Generation with Large Language Models, Xiaoteng Shen et al. introduced the Personalized Multimodal Generator (PMG). PMG extracts "soft" implicit embeddings from a user's historical interactions and blends them with explicit text keywords. This allows the system to generate custom visual content (like emoticons or posters) that balances contextual relevance with the user's unique, unstated aesthetic preferences.

2. Navigable Latent Spaces

Another approach is to design models that output a continuous space of possibilities rather than a single deterministic image.

In their work on Explorable Tone Mapping Operators, Chien-Chuan Su et al. paired a dual-network architecture with a generative framework to create an artifact-free "latent space." Instead of accepting a single algorithmic choice, users can actively navigate this space to find the exact balance of detail, brightness, and contrast that matches their personal preference.

3. Interactive Evolutionary Loops

When the aesthetic space is highly multi-dimensional—such as coordinating makeup, hair, and facial structure—direct latent navigation can become overwhelming. Here, researchers introduce human-in-the-loop evolutionary algorithms.

In Generation of Ideal Makeup Face Images Based on StarGAN-v2 Using Interactive Evolutionary Computation, Kaoru Arakawa treats neural style codes as genetic chromosomes. Users are presented with a generation of options and provide simple feedback. The system then cross-breeds and mutates the preferred attributes, allowing users to iteratively evolve a personalized look that does not exist in any single reference dataset.

[User History / Interaction] ──> [Implicit Embedding Extraction] ──┐
                                                                   v
[Standard Input / Prompt]     ──> [Base Generative Model] ──> [Personalized Output]
                                                                   ^
[User Feedback Loop]          ──> [Dynamic Reward / Loss Adjustment] ┘

Aligning Loss Functions with Subjective Rewards

To train these personalized models, the underlying optimization criteria must also shift. Traditional loss functions penalize deviations from a single ground-truth image or score. Personalized preference modeling, however, adapts the loss function or reward model to the individual.

In Joint regression and learning from pairwise rankings for personalized image aesthetic assessment, Jin Zhou et al. demonstrated that personalized aesthetic models can be directly integrated as loss functions into image enhancement pipelines. By learning from pairwise rankings unique to a user's collection, the enhancement network optimizes for that specific individual's color and brightness preferences rather than a generic standard.

For complex generative tasks like image inpainting, general alignment frameworks often wash out individual nuances. To counter this, Kendong Liu et al. developed PrefPaint, introduced in PrefPaint: Aligning Image Inpainting Diffusion Model with Human Preference. PrefPaint uses reinforcement learning (RL) trained on human-annotated preference data to align diffusion models with highly subjective human aesthetic standards, demonstrating a framework for moving beyond broad consensus optimization.

Cross-Domain Applications

The necessity of modeling subjectivity spans far beyond image processing. We see identical patterns across music, text, and design:

As summarized in the AIIDE workshop report by Adam M. Smith et al. (Workshops Held at the Ninth Annual AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE): A Report), the computational modeling of subjectivity represents a fundamental paradigm shift. Artificial intelligence is transitioning from an autonomous, objective judge enforcing a monolithic standard into a personalized collaborator capable of navigating conflicting, deeply human definitions of beauty, style, and utility.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 348 papers
  2. Version 1 · 2026-10-04 · 500 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_subjective_preference_modeling_cd0da483,
  title = {subjective preference modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subjective-preference-modeling/rdr_cd0da483c88849629e2b31bb2131f009},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_subjective_preference_modeling_cd0da483_v2,
  title = {subjective preference modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subjective-preference-modeling/rdr_cd0da483c88849629e2b31bb2131f009/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake