Hi everyone!
Text diffusion models are getting faster. After seeing models like Gemini Diffusion and Mercury, ByteDance’s new Seed Diffusion Preview is another big leap in speed.
I think the diffusion approach is a really smart idea. For tasks like coding that need you to see the whole picture, planning it all out first is often faster than just generating word by word.
The speedup is real: it delivers a 5.4x speedup over similar-sized models that generate word by word, while performing just as well on key tests. A very interesting new way to build generative models.