First try at a skin improver lora for Z-Image Base, it sort of works sometimes. recommend 0.4-0.7 strengh on ZIB
It can also improve ZIT images somewhat is used at a higher strength on ZIT 1.5 - 2.1
If you get bored of the slowness of Z-Image Base then checkout my V2 ZIT Model: https://civarchive.com/models/2231351/jib-mix-zit
Description
Initial Version
FAQ
Comments (15)
Phenomenal. Even at 1.0 strength, it improves ZIT outputs for "free" - I can stack it with a character LoRA and it doesn't distort the face. ZIB feels like the start of another golden age for local image gen.
Love it. Way better skin and less plastic chic.
Any chance we'll see a JIB Mix ZIB checkpoint?
Maybe, if someone can show that ZIB can get up to same photo realism as ZIT reliably, for now I am enjoying generating with ZIB and then doing a 2nd refiner pass with ZIT, it gives great results: https://civitai.com/images/119368457
I posted a workflow: https://civitai.com/models/2231351?modelVersionId=2644538
But I am still working on improving it now.
@J1B Looks impressive. As a Swiss pasta lover, I'll surely give this a try.
@J1B The more I play around with different schedulers the more potential I see with ZIB. Check this https://civitai.com/images/119958658 or this https://civitai.com/images/119958070
@LFMX I have been playing around with ZIB for hours today with the new ZIB Turbo lora, but what I have found is there is not really any free lunch and if you want a really good image I will have to wait at least 100 seconds to get it which is a shame. That $4000 RTX 5090 upgrade is looking very tempting to me right now...
@J1B That's my impression as well. I'm on a cloud based setup. Even with an H200 the good stuff (dpmpp_2m_sde + beta or ddim_uniform at 40 steps+) will take 20+ seconds.
I created this image with your prompt using Z-Image Turbo FP8 [Kijai] at my native resolution of 1.8 megapixels without LoRa upscaling or any other effects; this is how it comes out of my sampler. Would ZImageBase be even better? Image post by nedolo77 | Civitai
@nedolo77 i would think so: https://civitai.com/images/120013249
Check the details of her skin and hair. Prompt adherence such as spilling out cleavage. Only Lora used was Jibify at .3
@LFMX Then I'll probably test the FP8 version. I could also use FP16 with my RTX 4070, but since my workflow is over 400 nodes and many different models are running, that would take too long.
@nedolo77 above image was done with bf16 at 50 steps. Looking forward to see a comparison with fp8
@nedolo77 This is what I got out of my 2 stage ZIB to ZIT workflow:https://civitai.com/posts/26393252
ZIB follows prompts better and you did put "Digital Artwork" so it has made it a bit less detailed than it would have been I think.
But yes the fp8 Z-Image model seems to be similar quailty to the standard one.
@J1B I don't usually upload native images; they go through several upscalers and my photo effects beforehand. But I'll check how your workflow performs on my PC with the RTX 4070 12GB tomorrow or so. I took another picture using your prompt, natively with FP8 Turbo. However, I didn't have LoRa or anything else enabled, so don't be surprised if your LoRa data is in the files; it's embedded in my hash data because I usually have it enabled. Image posted by nedolo77
I am kinda curious on how you managed to train it to work with both base & turbo.. I thought Base LORAs don't work on Zit & Turbo LORAs don't work on ZiB.. what's the trick to train a LORA that can work on both ZiT & ZiB?
Yeah, base loras do work on ZIT you (sort of) you just have increase the weight to about 3. ZIT loras don't work on base thought, also the images I create at the moment are a 2 stage workflow using ZIT and ZIB and Civitai don't let you add 2 base models, I'm usally just using it on ZIB (although I might try it on ZIT again today).