zhen-nan/L2P
Researchers have developed a new transfer learning method called Latent-to-Pixel (L2P) that enables the creation of high-quality pixel-space diffusion models with significantly reduced computational and data requirements. L2P leverages pre-trained Latent Diffusion Models (LDMs) by freezing most of their layers and only training shallow ones to learn the transformation from latent space to pixel space. This approach allows for the generation of native 4K ultra-high resolution images and achieves performance comparable to the original LDMs on benchmarks like DPG-Bench and GenEval, using minimal GPU resources. AI
IMPACT Enables creation of high-resolution image generation models with significantly reduced training costs and data needs.