Metaesthetic AI

Where does the AI look come from?

Image models learn from billions of captioned web images, are often filtered or fine-tuned toward images people rate as appealing, and are sampled with guidance that trades diversity for fidelity. Published examples: LAION-5B holds 5.85 billion image-text pairs; Stable Diffusion v1 was fine-tuned on LAION subsets filtered to an estimated aesthetics score above 5; and Meta's Emu, quality-tuned on only a few thousand hand-picked images, won 82.9% of comparisons against its pre-trained version. Metaesthetic's reading is that this repeated selection for appeal is what produces the polished tells on the signature page.

Which steps in building a model shape its look?

Five steps, each documented by the people who built them. Figures are from the papers and model card listed under Sources.

Step Published example Figure What it selects for
Web-scale training data LAION-5B dataset (NeurIPS 2022) 5.85 billion CLIP-filtered image-text pairs, 2.32 billion with English captions Whatever the web pairs with text
Aesthetic filtering Stable Diffusion v1-2 and v1-4 model card v1-2: 515,000 steps on images with an estimated aesthetics score above 5.0 and watermark probability below 0.5; v1-4: a further 225,000 steps on "laion-aesthetics v2 5+" Images an aesthetics predictor scores highly
Quality tuning Meta's Emu (2023) Pre-trained on 1.1 billion image-text pairs, fine-tuned on a few thousand selected images; won 82.9% of comparisons with the pre-trained model and was preferred to SDXL 1.0 for visual appeal 68.4% and 71.3% of the time on two prompt sets A small set of "extremely visually appealing" images
Preference tuning Diffusion-DPO (2023), trained on Pick-a-Pic 851,000 crowdsourced pairwise preferences used to fine-tune SDXL 1.0 The image people picked in side-by-side choices
Sampling with guidance Classifier-free guidance (Ho and Salimans, 2022) Diffusers' Stable Diffusion pipeline defaults to a guidance scale of 7.5 Fidelity to the prompt, at the cost of diversity

Why does filtering for appeal matter?

Each of these steps narrows what the model treats as a good image. The Emu authors write that pre-trained models "often face challenges when it comes to generating highly aesthetic images", which is why they added a tuning step on hand-picked images. ImageReward, a reward model for text-to-image output, was trained on 137,000 expert comparisons for the same purpose: to score and steer output toward what people prefer.

The guidance step pulls in the same direction. Ho and Salimans describe guidance as a trade-off between sample quality and diversity, and the diffusers documentation says a higher guidance scale ties the image more closely to the prompt "at the expense of lower image quality".

What does that mean for the look?

Metaesthetic's reading: when every stage rewards the image most people would pick, the output drifts toward smooth skin, clean light and dramatic composition, the stylistic tells Kamali and colleagues list on the signature page. That is an interpretation: none of the papers above measures "AI look-ness" directly. What they do show is how much selection for appeal goes into a modern image model.

Sources

Checked 13 November 2026:

Next step

Once you know where the look comes from, decide which parts to keep: working with it and breaking it.