Projects / Universal Aesthetic Alignment

Universal Aesthetic Alignment

Why a Single Notion of Beauty Narrows Artistic Expression

Wenqi Marshall Guo, Qingyun Qian, Khalad Hasan, Shan Du

International Conference on Machine Learning, Position Track Spotlight, 2026

Generative AIAI AlignmentAesthetic Diversity
Examples of deliberately unconventional generated scenes used to study aesthetic alignment

Wide-spectrum aesthetic examples from the official project page.

01 — Overview

Overview

Image-generation systems are commonly optimized toward a broad, average notion of visual appeal. This paper argues that the same preference can conflict with user intent when a request deliberately calls for abstraction, discomfort, visual roughness, or other non-mainstream aesthetics.

The authors call this reversed alignment: instead of adapting to the user’s stated aesthetic goal, the system steers the output back toward the developer’s preferred visual norm.

02 — Contributions

Key Contributions

  • 01

    Frames universal aesthetic optimization as an alignment and user-autonomy problem rather than only an image-quality concern.

  • 02

    Builds a wide-spectrum aesthetics benchmark for testing whether generators follow unconventional visual instructions.

  • 03

    Studies generation, image-to-image editing, reward-model scoring, and the treatment of recognized abstract artworks.

03 — Method

Method

The study expands ordinary image descriptions with controlled wide-spectrum aesthetic attributes, compares generated outputs against those requests, and evaluates how aesthetic reward models score prompt-following but conventionally unattractive images. It also tests image editing and real artworks to separate prompt adherence from generic beauty preference.

04 — Evaluation

Results

Across the evaluated generators and reward models, the study finds a recurring preference for conventionally polished imagery. Systems often beautify or sanitize deliberately unconventional requests, while reward models can penalize outputs that follow those requests more faithfully.

05 — Reference

Citation

BibTeX citation
@InProceedings{Guo_2026_AestheticAlignment,
  author    = {Guo, Wenqi Marshall and Qian, Qingyun and Hasan, Khalad and Du, Shan},
  title     = {Position: Universal Aesthetic Alignment Narrows Artistic Expression},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026},
  note      = {Position Track Spotlight},
  url       = {https://openreview.net/forum?id=1gQ4zc1Q8I}
}