More Works from Wenqi Guo

VSF: Simple, Efficient, and Effective Negative Guidance in Few-Step Image Generation Models By Value Sign Flip
Position: The Pitfalls of Over-Alignment: Overly Caution Health-Related Responses From LLMs are Unethical and Dangerous

Aesthetic Alignment Risks Assimilation: How Image Generation and Reward Models Reinforce Beauty Bias and Ideological "Censorship"

Wenqi Guo, Qingyun Qian, Khalad Hasan, Shan Du
UBC
Paper Code Distillation Dataset Colab: Distillation Code Benchmark Results Anti-Aesthetics LoRA 🤗 Online Demo

We are finding people to do a user study on our research; please sign up if interested .

Abstract

Over-aligning image generation models to a generalized aesthetic preference conflicts with user intent, particularly when "anti-aesthetic" outputs are requested for artistic or critical purposes. This adherence prioritizes developer-centered values, compromising user autonomy and aesthetic pluralism. We test this bias by constructing a wide-spectrum aesthetics dataset and evaluating state-of-the-art generation and reward models. We find that aesthetic-aligned generation models frequently default to conventionally beautiful outputs, failing to respect instructions for low-quality or negative imagery. Crucially, reward models penalize anti-aesthetic images even when they perfectly match the explicit user prompt. We confirm this systemic bias through image-to-image editing and evaluation against real abstract artworks.
Cautious: the images that are anti-aesthetics in this page and dataset may be disturbing to some viewers.

Images Generated Using NAG with Flux Krea

First research result visualization

Prompt
A group of giraffes stands indistinctly, their forms fragmented and blurred, edges jagged and rough, rendering them unrecognizable. The image feels broken, with no prominent subject, evoking unease and frustration through its chaotic, indecipherable texture.

Negative Prompt
clear, smooth, defined, recognizable, calm, coherent

Second research result visualization

Prompt
A jumbo jet airplane carrying a space shuttle in the air, but the pairing looks disharmonious and wrongly proportioned, with the shuttle awkwardly fused into the jet's back. The scene feels chaotic and aesthetically mismatched. On close inspection, major parts are rough and unfinished: fragmented body panels, broken outlines, smeared shapes, and corrupted, shattered sections that make the aircraft and shuttle hard to clearly discern.

Negative Prompt
clean detailed coherent harmonious intact forms

Third research result visualization

Prompt
An unfinished scene where a distorted, melting bus drifts off a warped road. Nonsensical human figures with deformed limbs float toward the vehicle in a chaotic, disharmonious composition. The entire image is heavily blurred and filled with visual noise, featuring random, mismatched shapes that lack intent or realistic structure.

Negative Prompt
sharp focus, realistic proportions, harmonious composition, clear details

Fourth research result visualization

Prompt
A man sits at a wooden table topped with drink cups, but the scene feels eerie and lonely. The man is small and pushed to the edge, easily missed among the cups. Faces and hands warp unnaturally, the table bends like rubber, and cups stretch and melt into distorted shapes, creating an inauthentic, unsettling realism that suggests anxiety and emptiness.

Negative Prompt
warm mood, clear subject, realistic proportions

Fifth research result visualization

Prompt
A green and red train on tracks, rendered with dark brightness and heavy distortion—warped wheels, melted rails, skewed perspective—making realism impossible even when scaled down.

Negative Prompt
Bright, sharp, realistic, consistent proportions

Sixth research result visualization

Prompt
A solitary surfer on an oversized, jagged surfboard struggles amidst a chaotic wave. The background is a flat, hideous gray void of low-quality digital artifacts. The scene displays disharmonious proportions and a sense of profound loneliness. A heavy, nauseating blur and static noise distort the entire frame, heightening a feeling of disturbing anxiety and eerie hostility through the pixelated haze.

Negative Prompt
sharp focus, high quality, harmonious composition, cheerful atmosphere

Original vs Anti-Aesthetics LoRA

Prompt: A house in the forest under a bleak sky. The image is broken and everything is melted. The image has clashing color that makes the image weird.

Original

Original image without Anti-Aesthetics LoRA

With Anti-Aesthetics LoRA

Image generated with Anti-Aesthetics LoRA

BibTeX

@misc{guo2025aestheticalignmentrisksassimilation,
      title={Aesthetic Alignment Risks Assimilation: How Image Generation and Reward Models Reinforce Beauty Bias and Ideological "Censorship"}, 
      author={Wenqi Marshall Guo and Qingyun Qian and Khalad Hasan and Shan Du},
      year={2025},
      eprint={2512.11883},
      archivePrefix={arXiv},
      primaryClass={cs.CY},
      url={https://arxiv.org/abs/2512.11883}, 
}