When Attributes Disagree:
Gradient Conflict in Image Aesthetic Assessment

Different models, the same mistakes

We inspected prediction errors of several mainstream IAA methods and found a striking phenomenon: while models achieve competitive average performance, they collectively fail on the same subset of hard samples — forming stable error clusters in the prediction-label space. This cross-model consistency suggests an optimization-level issue rather than random noise.

(a) Baseline models (hollow) achieve competitive performance on normal samples but collectively fail on hard samples; AGREE (filled) substantially reduces these shared errors.

(b) Gradient conflict: subset gradients point in opposing directions on coupled parameters, partially canceling — leaving under-represented samples persistently hard to optimize.

Four mechanisms to establish agreement

AGREE mitigates gradient conflict via four complementary mechanisms operating at the parameter, feature, sample, and loss levels respectively — all without modifying the baseline architecture.

Overview of AGREE: a plug-and-play framework with four mechanisms — Attribute-aware Decoupling, Multimodal Fusion with semantic anchors, Sensitivity-guided Routing, and Error-aware Reweighting.

State of the art across five benchmarks

Bold = best, underline = second-best. AGREE integrated with two representative baselines (EAT and HKD-IAA) jointly achieves SOTA across all five benchmarks.

Consistent gains across six baselines

AGREE plugs into diverse IAA architectures, yielding average SRCC gains of 4.2%–12.6% across datasets and improving hard samples by up to 38.5%. Even strong baselines benefit substantially, confirming gains orthogonal to existing architectures.

Dumbbell plots across four metrics: (a) SRCC ↑ · (b) PLCC ↑ · (c) MAE ↓ · (d) RMSE ↓. Each line connects baseline (hollow) to +AGREE (filled). ● overall · ● hard samples.

Interactive: explore improvements across datasets & metrics

Switch dataset and metric — each row is one baseline; the bar shows how far +AGREE moves it from the baseline value.

Gradient conflict, quantified and resolved

We quantitatively verify that AGREE eliminates inter-attribute gradient conflicts. Cosine similarities across attribute branches drop to near zero (|cos|<0.05), and the training-dynamic conflict ratio decreases by 24%, leading to smoother optimization and faster convergence.

Hard samples are dominated by under-represented attributes. (a) The dominant-attribute marginalization M is significantly larger on common-error samples (Cohen's d = 4.42). (b) Their M distribution shifts right and separates from normal samples. (c) M positively correlates with baseline error (Pearson r = 0.41).

(a) Baseline EAT shows widespread negative cosine similarities between attribute pairs. (b) AGREE keeps off-diagonal entries near zero. (c) AGREE reduces the conflict ratio by 24% and stabilizes training dynamics.

Toggle modules and see each contribution

AGREE comprises four mechanisms. Click the toggles below to enable/disable any combination — results recompute live from the paper's ablation table (AVA, EAMB-Net backbone).