Tail-Imbalance Diffusion Equalizer for Class-Balanced Generation

Real-world image corpora are often imbalanced. On long-tailed datasets, the head class can outnumber the tail by a large margin, and a vanilla diffusion probabilistic model (DPM) consequently loses fidelity and diversity on rare categories, undermining controllable generation. To address this gap, we present Tail Imbalance Diffusion Equalizer (TIDE), a training framework that restores class balance without discarding data or adding extra training stages. Our contribution is threefold: (i) score field rebalancing, where we embed class-prior knowledge into a mixing matrix that routes gradients toward minority classes, with this class-prior knowledge consisting of the empirical label frequencies in the training set together with a uniform target prior; (ii) stable optimization, achieved through a trio of lightweight regularizers: gradient energy equalization aligns per-class gradient norms, class-adaptive minimum signal-to-noise weighting highlights the noisy timesteps most informative for rare labels, and tail-balanced self-distillation transfers the smoother representation of an exponential moving-average teacher back to the student; and (iii) balanced generation in practice, where TIDE leads to noticeably improved generation quality across both head and tail classes, all without data resampling or auxiliary models. This strategy reframes class imbalance as a principled correction of the diffusion score field, enabling DPMs to produce crisp and diverse images across the entire label spectrum and laying a foundation for their use on naturally imbalanced datasets.