Please support Early Access with your Buzz's, otherwise I can't release the next version, thanks in advance.
Quartus - (can be considered) has nothing in common with previous versions. First, I removed some data from the model for subsequent correct training (don't ask about the nuances; so much was done that I don't remember).
Because of this (which is what I wanted), the illustrations/realism/rendering/anime "blended together," which I can easily fix.
I ran out of Buzz because I spent most of them experimenting with finding the right training settings.
The NSFW content of male-female interaction has been almost completely removed; I'll add it in future versions. Realism and different styles need to be trained simultaneously to ensure penis-pussy contact looks anatomically correct.
____________________________________________Brief history of changes____________________________________________
Quintus: Coming soon
Quartus: +Furry and Characters.
Tertium: More NSFW content, balance adjusted (anime is now anime, photos are photos, and so on)
Secundo: Improved NSFW content, improved realism.
Primum: Fixing basic anatomy and improving prompt understanding + some NSFW content.
Description
A multitude of NSFW poses have been added; you will need to describe the positioning of the female and male figures, as well as the specific action taking place. ZTurbo, however, currently does not respond to the captions I have assigned.
Realism has been enhanced, though anime, illustrations, and 3D renders remain fully supported as well.
In the next phase, I will continue adding more poses until I have completely eliminated all anatomical errors—a process for which BAZZ is required.
FAQ
Comments (9)
What settings to use? my ZIT defaults only produce garbage: Euler/Beta 9 steps
Every image involves a specific workflow: Zturbo produces low-quality latent noise after 8 steps, so a preliminary sampler is required—running for up to 7 steps—to establish the composition.
Alternatively, simply use a single sampler with 7 steps—the quality will be slightly lower.
Works with NV2080 8GB with ModelComputeDtype node, select fp16 on that. EulerA/beta 10 steps on about 1376x1376 images 90 secs. Use tiled VAE with enough low settings (like 576 tile, 32 overlap).
ZTurbo, however, currently does not respond to the captions I have assigned.
You posted this above, Please explain that for me as I am not as knowledgeable as others.
Before training, every image is described—and it is as if there is no difference whether I provide captions or not.
Main showcase image has penis body horror. You may want to train this more on testicles so they will appear.