Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces the Eisbach log-barrier, a novel confidence-based loss weighting mechanism that leverages entropy from the DiT output's spatial energy distribution to enhance generative model training. By applying this method to LoRA fine-tuning of Stable Audio 3 Medium on MusicCaps, the authors achieve significant improvements in thematic development, acoustic differentiation, and textural diversity, countering the typical mode collapse seen in generative models. The findings highlight that in supervised diffusion training, the gradient direction is aligned with ground truth, allowing for effective scaling of step sizes based on confidence levels without compromising output quality.
The Eisbach log-barrier transforms generative model training by using entropy to enhance musical diversity and thematic development, defying conventional wisdom about confidence-based loss weighting.
Confidence-based loss weighting is usually avoided in generative models because it accelerates errors when the model is confidently wrong, but this intuition breaks down in supervised diffusion training. We introduce the Eisbach log-barrier, a parameter-free weight derived from the entropy of the DiT output's spatial energy distribution: high entropy damps the gradient, while low entropy preserves it. Applied to LoRA fine-tuning of Stable Audio 3 Medium on MusicCaps, it unexpectedly yields stronger thematic development, clearer acoustic differentiation, and higher textural diversity than unweighted training, the opposite of mode collapse. This works because in supervised diffusion the gradient direction is locked to ground truth, so confidence only scales the step size, and because temporal entropy downweights flat samples while preserving high-contrast ones. The result is an online, self-referential data curriculum that emerges purely from the forward pass, with analyzed noise-level dynamics and testable predictions.