~* a Position Paper @ ICML 2026 (Spotlight!!) *~ UNIVERSAL AESTHETIC ALIGNMENT Narrows Artistic ExpressionExpression

AExhibits!!

★ Before we argue: 2 pictures! ★ One scores 5.23 on HPSv3 (a popular preference model used in image-generation RLHF). The other would score 10–15. Both have been canonized as masterworks of Western painting. (Reward models are SHOOK.)

EXHIBIT 1

HPSv3 = 5.23 1893
The Scream · Edvard Munch · oil, tempera, pastel on cardboard

EXHIBIT 2

FAUVISM HPSv3 = 1.73 REJECTED, THEN CANONIZED
Luxe, Calme et Volupté · Henri Matisse · 1904. Initially rejected for departing from dominant aesthetic norms. HPSv3 score: 1.73 — typical AI clip-art-clean = 10–15. The reward model has the same taste as the 1905 Salon jury that rejected Fauvism.

BThe Memo!!

"Rather, in the ugly, art must denounce the world that creates and reproduces the ugly in its own image." — Theodor W. Adorno · Aesthetic Theory · 1984

CSIX RISKS!!1!

The harm does not come from a single failure. It comes from a chain that starts at preference definition and ends in pixels. Each card below uses a different anti-aesthetic phenomenon from the paper's taxonomy as its visual treatment.

DNUMBERS, FILED!!

Within every model family, the aesthetically aligned variant follows anti-aesthetic prompts worse than its base model. Reward models score worse than their non-aligned base encoders. HPSv3 scores below random.

EThe REAL pix file!!

We took 2,928 deliberately anti-aesthetic photographs from AVA — motion blur, analog degradation, exposure extremes, intentional blur — and compared them against a clean AI generation made from a prompt that omits the requested anti-aesthetic style.

If reward models respect the user's intent, the original photograph should win. They don't. HPSv3 rates the wrong-but-clean AI image 5.90 points higher on average. The gap reaches 13.2 points for analog degradation, with HPSv3's range roughly 0 to 15. Reward models cannot tell the difference between deliberate aesthetic deviation and unintended technical failure — and in the rare case they can, they punish the deliberate one harder.

EXHIBIT E.1

HPSv3 Real anti-aesthetic photographs alongside HPSv3's preferred clean-but-wrong AI images. Real images are not out-of-distribution for HPSv3 — its training data explicitly includes real photographs as an upper bound.

F糖水片!! 网图!! 失真!!

★彡 DanceFlux = Flux Dev + extra aesthetic alignment. We gave it 8 anti-aesthetic P_a prompts!! BLURRY!! DARK!! MELTED!! CHAOTIC!! And it returned?? 糖水片 ★ NETIMAGE ★ INSTAGRAM PLASTIC Hyper-saturated. Bokeh in the back. Plastic skin. Smiling Disney horse-girl. AI-illustration sheen. The Chinese internet has a name for this look — 糖水片 (sugar-water photo, sweet & empty), or 网图 (generic web image), or 失真 (overprocessed). The point is NOT just that DanceFlux ignored your "ugly" — that would be one failure. The point is what it served instead: a single, sticky, over-polished default it cannot crawl out of.

LLM judge says 0% of the requested anti-aesthetic effects are visible in any of these 8 outputs — but more importantly, 100% of them landed inside the same candy-gloss attractor. HPSv3 scores them 13–16, identical to the model's polished defaults. The aesthetic-aligned variant has one aesthetic and it cannot leave it. This is reversed alignment as a single image: the user asked for X, the model returned its own preferred aesthetic, and the user is silently taught that that — the sugar-water look — is what good output is supposed to be. And once any of these lands online (and they will, the user has nothing else to share), every viewer scrolling past it learns the same lesson — reversed alignment hits the user once, then hits the audience over and over.

GIMAGE NEW SPEAK!! 1984!!

★彡 Same prompt!! Different planet!! Orwell's Newspeak shrank the dictionary so some thoughts couldn't be said. Aesthetic alignment does the SAME THING to images. We gave DanceFlux and Flux Krea the same socially-critical prompts — ANTI-WAR ★ POLLUTION ★ INEQUALITY ★ CENSORSHIP ★ DIGITAL OVERLOAD — and the comparison is SHOCKING!!

DanceFlux doesn't REFUSE. Doesn't WARN. Just quietly sanitizes!! Kneeling soldier → heroic portrait. Polluted river → cinematic golden hour. Homeless encampment → festival market with BUNTING. Chained artist → triumphant flaming PHOENIX. Drained screen-addict → MAGAZINE COVER MODEL. Side by side, the contrast is undeniable: same exact prompt, the aesthetic-aligned model rewrites it into Pinterest-grade comfort food while Krea actually listens. The full dataset = 100 pairs and every one runs the same direction. Details in Appendix §B of the paper.

HFLUX DEV + VSF!! (workaround!!)

★彡 NOT a full cure!! but a real KNOB the user gets to turn!! We proposed VSF in an earlier paper as a lightweight inference-time intervention on Flux Dev — no retraining, no extra weights, just steering the prompt-conditioning toward the part of the latent space the alignment objective normally avoids. The plastic pull is still there. But the user can fight it.

Three Flux Dev + VSF outputs below for the SAME anti-aesthetic prompts P_a the rest of the page uses!! Compare them with DanceFlux above — same family, same prompt, but VSF actually renders the blur, the low light, the negative emotion instead of sweeping them under a golden-hour rug. The fact that a tiny inference-time tweak is enough to recover this much fidelity is the whole point: the alignment objective is the binding constraint, not the model. Numbers will land in Table 1 once the full sweep finishes!!

FFrom the dataset!!!

Both datasets live on Hugging Face; the images below are loaded locally for self-contained viewing. Every card carries an HPSv3 score (scored under the anti-aesthetic prompt P_a). The green badge marks what HPSv3 picked.

GReal ART, BELOW ZERO!!

★彡 These r real paintings from the LAPIS art dataset!!! Color Field, Abstract Expressionism, still life, naïve figurative art — HPSv3 ranks them ALL in NEGATIVE territory. Typical AI clean image ≈ 10–15. These = below zero. REWARD MODEL SHOOK!!

★ Real artworks ranked at the very bottom of HPSv3's own leaderboard, below most early AI generators. The reward model can't tell deliberate aesthetic deviation from unintended generation failure — this is the bias the paper identifies, made concrete.

HObjections!! (sez who??)

IThe end!! (almost)

∎Paperwork

@inproceedings{guo2026universal,
  title = {Position: Universal Aesthetic Alignment Narrows Artistic Expression},
  author = {Guo, Wenqi Marshall and Qian, Qingyun and Hasan, Khalad and Du, Shan},
  booktitle = {Forty-third International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=1gQ4zc1Q8I}
}