BorDELLA v2.0
Do yourself a favor and zoom in. Managed to pack some flesh into this one. Thought about adding art or something, but there wasn't any space left. Invented a whole new 3 model at a time merge method for this, but didn't feel like changing the name. All old models are still in the mix. New models added include:
https://civarchive.com/models/897413?modelVersionId=2580484
https://civarchive.com/models/1787477/truevision-77r
https://civarchive.com/models/2231360/z-mania
And I used https://civarchive.com/models/2217636/perfectrsbmix-zit-real-anime-sfw
at negative strength to improve skin texture and reduce sameface.
Sorry about the early access, but I'm trying to feed my addiction to models here. You understand. Plus, this is the most NSFW Z-Turbo checkpoint out there so far.
BorDELLA v1.0
It's been awhile. Found a merge method called DELLA, decided to take it for a spin. This is a merger of 4 models and 2 Loras:
https://civarchive.com/models/2168935/z-image-turbo?modelVersionId=2442439
https://civarchive.com/models/2304577/eason-zimageturbo-realistic-nsfw-v1
https://civarchive.com/models/2237711/z-image-turbo-nsfw-photorealistic-zit?modelVersionId=2564182
https://civarchive.com/models/2206377/zit-mystic-xxx
Thank you to all these creators for your training. I took all the weights with the greatest distance from base Z-Image Turbo and DELLA merged them. Seems to work pretty well. It has good detail, anatomy, and posing, probably better than any of its ingredients, if I do say so myself. It likes lower model shift than most Z models, I only used 1.5 for all these images. Doesn't play great with other LoRAs at full strength, except for sliders, as it's weights are quite saturated.
ComfyUI workflow for all images is embedded, if you want to get results like these. I call it a delayed denoiser. Doesn't finish cooking the image until after its been upscaled unsampled and resampled, so you gotta trust it's gonna turn out good, but it usually does.
Description
FAQ
Comments (14)
how it work with loras?
works bad
u tested with character lora? it's not working with character?(
@Getcontact yeah chars and style dont work. 0.8 making mess and if i lower to 0.5 no effect
@alexsmileface thanks
@alexsmileface @Getcontact I did some consultations with some mathematicians, and they told me about a number between 0.8 and 0.5 called "0.75". So, I tried it, and I made you a Topless Ursula from Disney's The Little Mermaid. You're welcome. https://civitai.com/posts/26170279
Anyway, not saying every lora works well, but many are ok, and 0.75 is a good starting point for characters. Be descriptive in your prompt and you can pull out weak details. Sliders work well, and a lot of NSFW loras only need a nudge.
@bot_i_celli it's fine man all ppl on civit trying to make nsfw model but it's still cant do nsfw sry i tried i didnt like the result
Will there ever be fp8 version of rhe v2?
Very nice checkpoint, thank you!
I am not able to get V 2.0 to work well. When using res 2s, what scheduler and steps were you using? Thanks!
i am using 15 steps, 1.5cfg, euler beta
The output of V2 is completely different from V1; I'd say it takes some getting used to, but it also produces some interesting results. If you're using one KSampler, I've had good results with 8 steps of euler a/beta or 8 steps of dpmpp_2m_sde/beta. However, you'll get better results with two KSamplers. For the first one, I use 8 steps of res_multistep/simple, and for the second, 5 steps of er_sde/sgm_uniform. Otherwise, feel free to experiment more; you might find some nice combinations :)
My workflows are embedded in the example images. I use res_2s bong_tangent to in the final KSampler of every image, though my initial steps are often euler_a beta, or euler beta. I use a basic guider with no negative conditioning, so equivalent to 1.0 cfg. I think I used 20-21 steps for most images, at 1120x1440. My sampler choice depends on what I'm trying to do, and how complicated the poses are. If I watch the live previews with res_2, and see large changes between each steps, where it looks like it's having difficulty converging, I'll switch to euler_a. If that doesn't converge, I'll switch to euler.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.