Classifier-free guidance (CFG) sharpens visual generation quality but doubles sampling cost by requiring multiple model evaluations. Prior guidance-free methods compress CFG into a single inference, yet they stay bounded by CFG's own weaknesses. Researchers at KAIST propose SCALE, a training objective that reparameterizes the likelihood target to include a self-contrastive term: the model compares its prediction on a real sample against its prediction on a self-generated sample, then adjusts toward the data manifold.
This yields SCG-enhanced predictions in one forward pass, eliminating CFG's overhead. On class-conditional and text-to-image generation, SCALE consistently beats CFG-mimicking baselines and, notably, keeps improving even when combined with DPO, a preference-optimization method. The authors argue the self-contrastive signal is not merely a CFG substitute but an additive source of guidance.
The paper is accepted at NeurIPS 2026. A single source means there is no conflicting view to weigh; the claims are reported as presented, pending further replication or discussion. The main open question is how SCALE scales beyond the tested diffusion and autoregressive architectures.