I took Nvidia’s new generative AI model called Sana for a spin.
It’s still rough around the edges: just to set it up and get it to work in ComfyUI was an ordeal, with some unexpected curveballs.
It also seems to have large, gaping holes in its training data, and of course it lacks the kind of robust tool, node, and control ecosystem of more mature models.
But putting that aside, there are 2 things about it that are quite remarkable and immediately apparent:
- The degree of prompt adherence.
- The speed. Dear Lord, the speed.
Even at 50 steps, something like the cat image below generates in 2.4 seconds for me, which considerably outpaces both Flux and SD at a comparable resolution and parameter set.
Will keep an eye on this one for sure.
