Text-to-image diffusion models often rely on strong guidance to align outputs with prompts, but this can backfire: independently sampled noise vectors collapse onto the same high-density mode, producing near-duplicate images. The problem is not just a lack of variety between pairs of images, but a collective failure to cover the full range of plausible visuals.

A new paper on arXiv proposes tackling this with determinant maximization, a technique that rewards diversity across the entire batch of samples rather than only penalizing pairwise similarities. The authors argue that this collective view is key to escaping the prototype-collapse trap while preserving the benefits of strong guidance.

The work is presented as an advance beyond pairwise repulsion, which has been a common but limited strategy for encouraging variation. By framing diversity as a global property, the method aims to keep diffusion sampling both faithful to the prompt and genuinely varied across multiple outputs. The abstract does not yet provide experimental details, so the practical gains remain to be seen, but the conceptual shift is clear.