Even when prompted without stylistic instruction, contemporary text-to-image (T2I) models exhibit a recognizable, model-specific look. We decompose generator outputs along an interpretable, style-specific concept space: LouvreSAE, a sparse autoencoder over CLIP image embeddings whose dictionary is shaped to prioritize stylistic rather than semantic features, and LatentAesthetic, a methodology that uses LouvreSAE to extract and compare T2I generators’ default aesthetics. Applied to 26 prominent models spanning four years, our method surfaces stable per-generator profiles, a uniform pull toward lifestyle and editorial photography over artistic imagery, and a steady cross-family aesthetic contraction.
Ask two text-to-image (T2I) models for “a pair of leather boots” and you get two completely different images — not merely in semantics and composition but, notably, in style. Users of T2I models have long pointed out the distinctive look of some individual models (e.g., “the Midjourney look” or “the DALL·E sheen”), and it has been shown that accurate classifiers can be trained to distinguish among the generators’ outputs. Yet, no one can say or exactly measure what such aesthetic profiles are made of, and how they evolve over time. Our work fills this gap. To demonstrate the kinds of findings and evaluations one can obtain with this method, we apply it to 26 prominent T2I models from the last four years.
Below, we present a couple of prompts from a style-neutral bank. Nothing in these prompts explicitly indicates a medium, a period, an artist, a lighting setup, or a camera. The framing, the palette, the light, the finish — all of these are the model’s own choices. We call the patterns in how each model chooses to render a subject when stylistic attributes are not controlled its aesthetic defaults.
Examples of content-controlled images, used to extract aesthetic profiles. Above, we show three prompts across various evaluated models. All three come from the content-only prompt bank.
We train a sparse autoencoder over CLIP image embeddings on a mixture that is deliberately art- and style-heavy, so the dictionary focuses its capacity on brushwork, lighting, palette, material, and composition rather than on object categories.
Once trained, we interpret and categorize each concept learned by this SAE, and keep only the 807 that describe an aesthetic property.
Each concept below is shown with the label it was given, its category, and a prototype sphere — the concept steered onto a plain white sphere, which isolates what the direction contributes from whatever objects happen to co-occur with it — and four images that activate it strongly.
Examples of concepts learned by LouvreSAE. Prototype visualizations and top activating images for eight of the 807 concepts in the aesthetic vocabulary.
A generator's default aesthetic profile is its average activation over concepts from LouvreSAE, measured over a pool of content-controlled, style-neutral prompts. The profiles are stable enough to treat as a property of the model: two disjoint halves of a generator's output agree at 0.99, and a classifier trained on top of our aesthetic profiles can distinguish the models at 9× chance.
Pick any of the 26 generators to see some of the concepts that distinguish it from others.
Examples of generators’ aesthetic profiles. Shown above are the concepts on which each generator deviates most from the average of all evaluated models, with z as the standardized deviation. Midjourney v6 is painterly and compositional; DALL·E 3 is ornamental and dramatically lit; Flux 1.1 pro is grainy and screen-like.
We ground profiles in real-image distribution references, such as curated lifestyle photography or art-historical imagery. These help us understand which overall trends the models resemble more. All tested models are closer to the photographic reference.
This has changed over time. Over the four years covered by our data, the distance of the generators’ aesthetic profiles to the photographic style has generally remained flat. What has changed is the distance to the art-historical reference: it falls steadily (−0.012/yr, r = −0.40, p = 0.03), and the release-year cohort mean drops every single year: 0.733 → 0.719 → 0.711 → 0.701. It holds inside the model families too.
So models did not become more photographic. They became less painterly. The art-historical character that early generators favored — the flat 2D rendering, the visible brushwork, the illustrative surface, and more — has been drained out release by release, while the distance to photography stayed pretty much the same.
Aesthetic drift with respect to references. Two y-axes: similarity to photographic style is shown on the left and similarity to art-historical style is shown on the right. The heavy lines are least-squares fits over all 26 generators.
When we consider the ten pairs of multi-release model families and compare their earliest contemporaneous releases against their latest, eight of the ten get closer and the average cosine distance falls from 0.047 to 0.035. And the two that widen are quite specific: one is the jump between DALL·E 2 and 3, which was characterized by a genuine shift in the model aesthetic, and the other is the Ideogram–Midjourney pair, which widens only marginally (0.052 → 0.054).
Within-family aesthetic convergence
Exception: DALL·E 2 to 3 update
Qualitative comparison of DALL·E 2 and 3
Cross-family convergence. Left: each line is a pair of multi-release model families, joining the distance between their earliest contemporaneous releases to the distance between their latest. While most get closer, the two in red widen. Middle: the concepts DALL·E 3 shed and picked up relative to DALL·E 2, the largest within-family move we found. Right: the same shift, but illustrated with qualitative examples based on identical prompts.
A profile that can be measured can also be optimized. We fine-tuned Stable Diffusion 1.5 toward the art-historical reference and re-measured its default at each checkpoint with the unchanged pipeline. The photographic-versus-art margin shrank at every checkpoint while prompt adherence stayed flat. Because the target is a vector of interpretable concepts from LouvreSAE, this monitor can say which exact properties moved (or didn’t) during training. A scalar metric such as CLIPScore or the diffusion loss can only report that training is on track; it cannot name what the model is learning to look like, and it does not afford as much sensitivity to stylistic changes in the model.
Aesthetic profiles as a training monitor. Photographic-versus-art margin over 2,000 LoRA steps (left axis) against CLIPScore, a standard measure of prompt adherence (right axis). The margin falls 0.114 → 0.088, a 23% reduction, while CLIPScore holds at 0.252 ± 0.002. Thick brushwork rises 0.18 → 0.54.
The same profile can also choose training data, and even act as a training objective. Given a large unlabeled pool of mixed images, we scored every image by how closely its concept activations match the art-historical target and fine-tuned on the 6,000 best matches. With no human labels involved, the selection turned out to be mostly real paintings, and training on it moved the model toward the target — while none of the baselines or controls did.
Alternatively, the profile can enter the training loss directly: adding the distance to the target profile as an extra loss term moved the model toward the target on identical, randomly chosen data, and the strength of the pull scales with the weight on that term.
While the primary purpose of LouvreSAE and LatentAesthetic is to understand the aesthetic defaults of generative models, the same artifacts and methods can be used for conventional style transfer. These are not main results and we don’t seek to make any conclusive claims here — though we think they illustrate the versatility of the contributed artifacts.
@article{bohacek2026aesthetic,
title = {Characterizing the Aesthetic Defaults of Generative Image Models},
author = {Bohacek, Maty and Panda, Raina and Fein, Daniel and Singhal, Arpita
and Fiore, Mark and Freeman, William T. and Agrawala, Maneesh},
year = {2026}
}This work was supported by a Brown Institute Magic Grant. The authors thank Yael Vinker for helpful discussions.