V. Lokhande

Hallucinations in diffusion models can be reduced by up to 25% without sacrificing image quality, thanks to a novel score modulation technique.

Diffusion language models inherently signal factual uncertainty during generation, outperforming even trained hallucination detectors, and OSCAR leverages this to reduce hallucinations.

Achieve state-of-the-art dataset distillation without any training by guiding diffusion models along learned latent manifolds, yielding synthetic datasets with superior fidelity and representativeness.