Yee boi
I currently recommend 3.1 for its increased detail.
i recommend euler, simple, cfg 1
You can get away with complex NSFW poses using multiple characters if you use aura flow at 2-5. You can also turn it off completely different results. aura flow tends to make things a bit too perfect, it does corrections, so it is great for some things but not always desired.
I use my own workflow you can grab at: https://civarchive.com/models/2169381/donut-z-image-turbo-workflow-upscaleface-detailernsfw-fix
I also use https://civarchive.com/models/2191617?modelVersionId=2493436 for first pass and regular ae.sft for upscaling
To maximize flexibility and aesthetics use flexible model for steps 1-3 and aesthetics model for 3-9 (previews were generated this way)
props to: https://civarchive.com/user/epiTune/models for making moil.
Share your work with us at: https://discord.gg/CDkuG8knGG
Description
This is the actual V4. The version uploaded yesterday was a modified V3 which was uploaded by mistake. it is still better than V3 (and maybe even v4) so i renamed it to 3.1 and removed the 2 inferior models to keep it lean. Sorry that I keep fucking up the uploads every time. Most of the 3.1 previews were made with V4 cause i thought that one was being uploaded.
FAQ
Comments (10)
apolgy for bad english
where were u when donut delivered
i was at house eating dorito when civitai ping
"donut is deliver"
"yes"
You nailed it with v4.
great model!
The examples look great, though would appreciate it if you could share the prompts so we can have a better idea of which terms and concepts the model knows. Thanks!
It used standard zit clip, the encoder is not trained, so same prompting as all other zit models that don't ship with a custom trained encoder.
@DonutsDelivery Got it, but I was referring to the terms and words used in the sample images. For instance, what's the difference in the prompts for the images with the lady holding a dagger or glass of wine? Is it just "small breasts", "medium breasts", etc? Or is there more specificity?
@Nexdoor "holding a glass of wine" or "holding a short dagger", "she has small breasts" "she has large breasts" "she has very large breasts". i use natural language, not danbooru tags. theres not much to my prompting. sometimes i start the prompt with something like "a high quality vibrant dslr photo of ..." but that depends on what im trying to achieve.
This model is interesting. It's handled everything I throw at it with little issue. But the thing I like is that it generates flawed, realistic people, not the idealized supermodels many others will put out.
I absolutely agree and for that reason 3.1 is actually my favorite version. In fact this is by far my most favorite ZIT model. Totally underrated and needs more exposure.
hmm i was trying to use my character lora trained on normal zit but sometime the consistency is not there, anyway around it?
ive found out that ZIB loras work best with ZIB trained loras. theres incompatability issues with ZIT trained loras even though you apply it on a ZIT model.