PRIVIMAGE, a novel method for efficiently generating differentially private synthetic images, leverages semantic-aware pre-training on a carefully selected subset of public data to achieve superior fidelity and utility compared to state-of-the-art approaches.
Combining the diversity of diffusion models, the efficiency of flow matching, and the effectiveness of convolutional decoders enables state-of-the-art high-resolution image synthesis at minimal computational cost.
A novel post-training quantization scheme, QNCD, that effectively mitigates both intra and inter quantization noise in diffusion models, enabling efficient low-bit inference while preserving high-quality image synthesis.
Innovative training-free approach FouriScale enhances high-resolution image generation by addressing structural and scale consistency through frequency domain analysis.
Pyramid Diffusion Model enables ultra-high-resolution image synthesis through innovative architecture and latent representation.
Improving training dynamics in diffusion models leads to better image synthesis results.
Pyramid Diffusion Model enables ultra-high-resolution image synthesis through a novel architecture and latent representation.
PCDMs incrementally bridge the gap between source and target poses through three stages, generating high-quality synthesized images.
Efficiently distilling large latent diffusion models for high-resolution image synthesis.
合成画像のアーティファクトを分類し、軽減するためにVision-Languageモデルを使用する方法を提案します。