ERNIE-Image-Aes: Robust Image Aesthetics Scoring with Balanced Category Generalization

TL;DR

ERNIE-Image-Aes is an 8B vision-language model from Baidu for image aesthetic scoring, initialized from ArtiMuse and fine-tuned on a category-balanced, professionally annotated dataset. It targets the systematic biases of existing predictors — LAION-Aesthetic, ArtiMuse, and UniPercept all overscore specific image classes (AI-generated/anime, black-and-white, casual snapshots). The release includes a companion benchmark, ERIA-1K, of 1,000 images calibrated to a realistic deployment distribution across six categories, on which ERNIE-Image-Aes roughly doubles the SRCC/PLCC of prior open predictors. Both the model (Apache 2.0) and benchmark are openly released.

Key claims

ERNIE-Image-Aes is an 8B VLM initialized from ArtiMuse and fine-tuned on a category-balanced, professionally annotated aesthetic dataset [Highlights].
LAION-Aesthetic disproportionately high-scores AI-generated and anime content; ArtiMuse overscores black-and-white photography and casual snapshots; UniPercept strongly prefers monochrome and casual snapshots [Motivation table].
On the ERIA-1K benchmark, ERNIE-Image-Aes reaches 0.7445 SRCC and 0.7598 PLCC, versus 0.4533/0.4748 for the next-best baseline UniPercept [Results table].
Training labels are produced via a pairwise Swiss-system tournament with tier labels from 1 to 10, with annotators recruited from professional fine-arts and media programs and passed through aesthetic calibration screening [Annotation Protocol].
ERIA-1K contains 1,000 images across six categories with proportions calibrated to approximate a real-world deployment distribution, rather than the curated Flickr/DPChallenge-style distributions of prior benchmarks [ERIA-1K Overview].

Method

ERNIE-Image-Aes inherits its architecture from ArtiMuse (itself built on InternVL-3) and is fine-tuned for aesthetic scoring on a purpose-built dataset whose annotation protocol is the main methodological contribution. Rather than absolute scalar ratings, the team uses a pairwise Swiss-system tournament: images are paired against one another in a swiss-style bracket, and aggregated outcomes produce stable, reproducible tier labels from 1 to 10. Annotators are recruited from professional backgrounds (Central Academy of Fine Arts, Sichuan Fine Arts Institute, Communication University of China, etc.) and must pass an aesthetic calibration screening before participating. The training corpus is explicitly category-balanced across photography, anime, design, everyday snapshots, and film photography, with care taken to avoid systematic bias toward AI-generated content or monochrome images. The released ERIA-1K benchmark uses the same protocol on a held-out 1,000-image evaluation set whose category proportions are calibrated to a realistic deployment distribution.

Results

On ERIA-1K, ERNIE-Image-Aes achieves SRCC 0.7445 and PLCC 0.7598, compared with LAION-AES (0.2944 / 0.3138), ArtiMuse (0.4277 / 0.4704), and UniPercept (0.4533 / 0.4748) [Results table]. That is roughly a 1.6× improvement in SRCC over the strongest open-weight baseline. The benchmark is deliberately constructed to expose category bias: prior predictors’ lower scores partly reflect that ERIA-1K mixes image types that those models score poorly on (anime, AI-generated, monochrome, casual snapshots) at realistic proportions, rather than the curated semi-professional photography that dominates Flickr/DPChallenge-derived benchmarks.

Why it’s interesting

Aesthetic scorers are load-bearing infrastructure for generative-image pipelines (filtering training data, picking samples for RLHF, ranking outputs), and the systematic biases catalogued here directly affect what generative models learn to produce — LAION-Aesthetic in particular has been the standard filter for years, and Baidu’s results suggest it heavily over-weights AI/anime imagery, which is a feedback-loop hazard when used to curate diffusion training sets. The pairwise-Swiss annotation protocol is a more principled label source than the noisy MOS-style ratings used in older aesthetic datasets, and Apache 2.0 weights plus an open benchmark make this immediately usable as a drop-in scorer. The biggest caveat is the absence of a public arxiv writeup at filing time — the model card alone is the source of all reported numbers.