| |
Researchers developed JiT-DDT, a new encoder-decoder architecture for text-to-image generation that trains 3.6× faster than the previous Linum v2 model while generating images at 4× higher resolution (512×512 vs 256×256 pixels). The key innovation replaces the traditional two-stage Latent Diffusion Model approach (separate VAE for compression and DiT for generation) with a unified pixel-space diffusion model that constructs its own efficient latent space specifically optimized for image generation rather than reconstruction.
Read Full Article →
← More Tech news