
Workflow:
https://civarchive.com/models/2662358?modelVersionId=2989530
Huge improvement to ZIB, and a reduction in asian-ness.
12 steps 1CFG
Use this text encoder:
https://civarchive.com/models/2650573/thmuncz-imagete
or this one
https://civarchive.com/models/2662833/zimagetesolordz?modelVersionId=2990083
https://www.buymeacoffee.com/tripleheadg
We should have a Patreon or something equivalent set up in future for early access to our new model files directly. Don't bother sending me Buzz because I am located in a country which cannot recieve monetary rewards from Civitai.
We used our own custom training code and custom nodes for merging.
Stay tuned for more updates, we have a lot coming over the next few months.
Description
FAQ
Comments (11)
It's an interessting model. It's defined to ZImageBase but it works with er_sde/simple and 8 Steps very well. the prompt adherence is very impressiv and at the moment i use the standard clip. I have to test a bit more but i think it will be very good :)
I'm glad you like it :D Thanks :D
The prompting and understanding is crazy... It understand what you prompt and do exact this what you prompt (if it knows what it is what you mean), it's ultra sensible. But for good pictures you need 2 KSampler. First one i have er_sde/beta 8Steps, then i upscale to 160%, then second KSampler with er_sde/sgm_uniform 5Steps/0.56denoise (both cfg1.0). Sometimes it have problems with limbs and fingers, but it's rare and can be that er_sde not the best sampler for it. Good tests are also with euler_a but for me euler_a is a bit to clean for my taste. Everyone have to decide. For ultra realism skin i think you need a ZIT refiner and SD Upscale node with upscale_by 1. For me, my opinion, it's the first checkpoint that can do this good prompt adherence in so less steps. I like it very much. Thanks for the checkpoint.
@Gorean Thank you for your detailed feedback, it's very important to us. We are trying to improve this model as much as possible and see progress, but the opinion of users is very important. Have fun :)
@speach1sdef178 one point more, for my pictures i always used skinny lora and i needed very heavy str (-3.5), i think it's a ZIT lora and your base model is more dominant, thatsway the str. but it worked :) Unfortunately, I cannot add the LoRA to my sample images here on Civit, as Turbo LoRAs do not appear under the "Base Models" section. https://civitai.red/models/2229239/zit-skinny-fat-slider?modelVersionId=2509564
@Gorean Yes, LORA for ZIT works with our ZIB model, but sometimes you need a higher weight value and sometimes a lower one, everything depends on LORA itself. As for the fact that LORA is not visible as part of the workflow, I think this is some kind of civitai bug. There have been a lot of bugs lately.
@speach1sdef178 Nono... Civit doesn't have any bugs... those are all features. ;)
@speach1sdef178 For my taste it is a bit oversaturatet with color, now i use the dpo Lora with -0.4 to take out a bit of color. It works very well for me
https://huggingface.co/F16/z-image-turbo-flow-dpo
@Gorean glad to hear that :) and thanks for your feedback ))
@speach1sdef178 I’ve noticed a problem that I’d like for you to address. I really love slender women. Unfortunately, with your checkpoint, the lower ribs on the torso tend to become visible very quickly—something that Civit, unfortunately, did not intend. It would be great if you could fix this for the next version. I tried to resolve it using prompts, but unfortunately, that didn't work—at least, none of the prompts I tried seemed to do the trick.
This is pretty great! Doesn't really work super well for NSFW, but it does Great portraits.



















