The official turbo LoRA dropped, so I'm sharing the new version of the merge I'm using. This one does best at 8 steps. Full checkpoint, no text encoder or VAE needed, and it's small and fast. The results look nicer to me than what v1 was putting out, so yeah. Keeping this version.

This is a merge of the fp8 diffusion model of Anima Base, the new official turbo LoRA by the folks who made Anima, the Heretic-Abliterated version of the Qwen text encoder (so you can prompt for pairs of small blue birds without being censored for it), and of course the standard VAE.

Stick it in a checkpoint loader, set it for CFG ~1-1.3 (yes, some negative prompting ability survives, I use CFG 1.3 all the time) and 8 steps or so, and have fun.

I also like some of the outputs I get at 4 steps, but then that's just me, probably. You'd probably find them too washed-out. 8 steps is good.