Hacker News Reader

prior generations usually take fewer steps than vanilla SDXL to reach the same quality.

But yeah, the inference speed improvement is mediocre (until I take a look at exactly what computation performed to have more informed opinion on whether it is implementation issue or model issue).

The prompt alignment should be better though. It looks like the model have more parameters to work with text conditioning.

in my observation, it yields amazing perf at higher batch sizes (4 or better 8). i assume it is due to memory bandwith and the constrained latent space helping.

However, the outputs are so similar that I barely feel a need for more than 1. 2 is plenty.

I think that this model used consistency loss during training so that it can yield better results with less steps.

...because they feel that at 20+10 it achieves a superior output than at 50 steps for SDXL. They also benchmark it against 1 step for SDXL-Turbo.