Linum introduces Pyramid-JiT (P-JiT), a decoder-only pixel-space architecture that eliminates the need for a Variational Autoencoder (VAE). By predicting target images at multiple resolutions along the DiT trunk, P-JiT achieves comparable FD-DINOv2 scores to their previous latent diffusion model Linum v2 using 11.3x fewer training samples and 4.3x fewer GPU-hours, while supporting 4x higher pixel resolution.
VAE
变分自编码器
A generative model that learns data distributions via encoder-decoder networks
Curated Intelligence
The vLLM-Omni framework implements system-wide full-stack optimizations for the MiniMax H3 model, covering attention mechanisms, communication, and parallel VAE decoding. It further integrates FastVideo's FastH3 technique to reduce DiT forward passes from 49 to just 4. Benchmarks on eight B300 GPUs demonstrate that this approach generates a complete 10.125-second MP4 video in approximately 8.7 seconds, achieving 'real-time' performance where generation speed exceeds playback duration.
A new primer has been released that connects major generative AI architectures like PCA, VAEs, diffusion models, and GANs through compact mathematical derivations. Targeted at researchers and students with math backgrounds, the book focuses on revealing the unified mathematical structures behind these models rather than engineering implementation details.
Hugging Face proposes delegating the VAE decoding step of diffusion models to remote inference endpoints to alleviate VRAM constraints on consumer GPUs during high-resolution generation. This approach aims to replace traditional offloading or tiling methods, reducing local hardware requirements without compromising image quality. The code is open-source with privacy guarantees, currently in a community pilot phase.
OpenAI published the 'Variational Lossy Autoencoder' research paper in 2016, introducing a generative model architecture that combines variational inference with lossy compression. The model achieved state-of-the-art density estimation results on datasets like MNIST and OMNIGLOT, laying important technical groundwork for subsequent sequence modeling and generative AI.
OpenAI announced the addition of several top machine learning researchers, including Ian Goodfellow and Yura Burda, to its early team. The update focuses on personnel expansion rather than specific model releases or technical breakthroughs.