vae

"Overall, from my tests, this vae produces identical results to original, but at 3x less vram, and at better speed."

"Instead of the usual CLIP + VAE + Diffusion setup we're used to in Stable Diffusion or FLUX, they’ve built a Native Unified Model called NEO-unify."

"Decided to make my own stable diffusion... in doing all of this on a CPU, using CFG with a bigru encoder, 32x32 images with 8x4x4 latent, 128 base channels for VAE and Unet"

""Let's leave everything—both the VAE's KL divergence (regularization) and the image reconstruction (decoder)—entirely to the diffusion model!""

"VibeVoice achieves 80x compression compared to Encodec with a 7.5 Hz tokenizer, enabling the synthesis of natural conversations up to 4 speakers and 90 minutes long within a single LLM context window, while surpassing competitors with an MOS of 3.76."

"Our VAE-based 2x upscaler strictly enlarges images within its range without hallucinations, delivering a purely true-to-source."

"This paper attempts to quantify the integration risk of corporate acquisitions (M&A) using differential geometry and machine learning (VAE)."

"Using VAEs, we construct a manifold of strategic space from social, political, economic, and marketing data, obtaining a smooth latent space from noisy data and performing differential geometry and control engineering simulations on that space."

""RAEs consistently outperform VAEs during pretraining across all model scales. Further, during finetuning on high-quality datasets, VAE-based models catastrophically overfit after 64 epochs, while RAE models remain stable through 256 epochs and achieve consistently better performance.""

"Variational autoencoders (VAEs) are known as image generation models, but can also be used for 'image correction tasks' such as inpainting and noise removal."

"The study focuses on the environmental dependence of Type Ia Supernovae in low-metallicity host galaxies."

"The paper is available on ArXiv."