Is the merge recipe for LoRAs used for Aesthetic V1.0 available?
I'd like to know the LoRAs used and the recipe.
They are all custom trained loras not really useful outside the context of modifying the style/composition of the underlying aesthetics tune. Besides, I am making a v1.1 with a more standard style and that will be ready later today or tomorrow.
I guess I can share, in general terms, the categories of these loras and what they do.
- DMD-inspired distillation lora, but applied at low weight. So it doesn't really do any distillation, but collapses the distribution just a bit while also biasing the style and adding details. This is the main one I'm altering with v1.1, to make it have smoother lines and less noise.
- Anime coloring lora, trained on a couple thousand high-rated danbooru images with "anime coloring" tag. Again, applied at weight less than 1, just to push the othewise 2.5D-ish style to be a little flatter.
- Lora trained on "heavy sweating" tag, applied at low negative weight. This removes excessive tendency to add sweatdrops (a side effect of most types of DMD loras), while leaving the image otherwise almost pixel-identical for the same seed without the lora. The full Turbo model also has this which is how I fixed the excessive sweatdrops that are in the turbo lora.
They are all custom trained loras not really useful outside the context of modifying the style/composition of the underlying aesthetics tune. Besides, I am making a v1.1 with a more standard style and that will be ready later today or tomorrow.
Will v1.1 also come with a v1.1b release?
There were some cases where v1.0b produced more natural or cleaner results than v1.0
No, since v1.0b is just the underlying aesthetics finetune with nothing else added. That's a fairly large finetune, and that part isn't changing. v1.1 will change some of the loras merged in, to make it cleaner and less noisy.
will we see a turbo version of 1.1?
Haven't seen any update yet, so did you run into any problems, bro 😂
I have a version I could release now, but I've been spending a day or two retraining one of the loras with different settings to see if I can improve it further. Idk maybe tomorrow.