ZIT_LikeAReal for realism and ZIT_ANI for cartoon/anime styles.
I found that creating full models gives better results than LORAs, so I'm releasing these as standalone models.
The models are available as FP16 and FP8.
ZIT_LikeAReal
This model is better, especially for close-up or human-focused visuals.
If you just type "woman" in the basic ZIT model, you'll most likely get an Asian person by default. LikeAReal will most likely make a woman other than an Asian person.
It's much better for creating NSFW content compared to the basic model and has more realistic human skin textures.
Even when creating cartoon or anime footage with this model, you get more detail, better contours, and sharper lines.
Description
ZIT_LikeAReal
This model is better, especially for close-up or human-focused visuals.
If you just type "woman" in the basic ZIT model, you'll most likely get an Asian person by default. LikeAReal will most likely produce a woman other than an Asian person.
It's much better for creating NSFW content compared to the basic model and has more realistic human skin textures.
Even when creating cartoon or anime footage with this model, you get more detail, better contours, and sharper lines.
FAQ
Comments (8)
Hello I have some basic questions about:
-If I buy a RTX 3090 24Gb, let's assume using the ZIT standard version, what is the resolution I can get for the images? Only 1024x1024 or more?
-There is a function in Auto1111 or ComfUI to do a sort of Ima2Ima function? (on CivitAI there is not for ZIT)
Hi!
RTX 3080-Ti 12 Gb, up to 2k is great. So I guess in 3090 it will be up to 4k. But still you can cut image into parts and combine them in some image editing software with masks.
More than 2k gives artefacts/glitches on one square of an image - upper right, but the rest of the image is clean and good.
In ComfyUI img2img (if that is what you meant)
use "load image" -> "vae encoder" -> "latent image" in ksampler.
than use shift - fewer more close to original
use denoise - fewer more close to original
This model gives better and clean results than others, I tested some others - they give a lot of noisy dots in resulting image.
For example shift - 1.5 and denoise - 0.5 preserves all details of original image, but gives photoreal copy of original image
@ff84 Thanks for reply! Do you also know if ZIT images contain META data? (Prompt, model, version etc...) I cant see them with my usual off-line tools for Stable Diffusion...
не получится генерить 4к картинки, при разрешении больше 2500 по краям вылезают артифакты
генеришь картинки в 2K ,а потом апскейлишь их до 4к
Самое лучшее разрешение для ZIMAGE это 1080*1920
@settima_ai Sorry for the long answer) Yes it contains metadata, if you use comfyui, than it saves all workflow, not sure about other SD UI, SD Forge saves prompt, but not sure about other data
Does as advertised. Looks like and has similar promt adherence to the normal ZIT model, but better for NSFW, very nice to use.