Jib Mix Zit is an upgrade for realistic images, sharper and prettier faces.
Jib Miz Zit V2:
Better fine small details (especially in backgrounds)
More variation in image and Style (Doesn't always produce photo realism but can be a nice surprise)
Less Zit mottled Skin.
More Varied faces and poses.
- A bit less photo-realistic.
I recommend the FP16 model for 12GB of Vram and up
But a smaller FP8 model is available.
For Best Image Quality try:
I recommend using 80% of the UltraFlux VAE merged on the fly with 20% of the Normal Zit/Flux VAE for sharper images
UltraFlux VAE: https://huggingface.co/Owen777/UltraFlux-v1/tree/main/vae
This custom nod: https://civarchive.com/models/2231351?modelVersionId=2638152 does a merge of the 2 VAE on the fly as the UltraFlux- VAE can be a little too sharp-looking on its own and definitely is if you use it for upscales.
V1 model can benefit from a small amount more of my Zit lora (around 0.15): https://civarchive.com/models/2194714/jib-mix-realistic-z-image-lora
I recommend ClownSharkSampler and Linear/ralston_2s for detail.
with some added noise option nodes.
It has been tuned against my workflow (I have a version posted here: https://civarchive.com/models/2194714?modelVersionId=2481800)
Description
I made a specialist workflow to get the best out of both Z-image and ZIT as a refiner.
FAQ
Comments (36)
The base to turbo workflow is incorrectly labeled as a checkpoint merge.
Thanks, I threw that up there quite quickly as people were asking for it, now corrected.
Is there a way to use the turbo LORAs in this workflow as I know it uses base first.
Yes there are separate lora slots for each model. Also you can kind of use ZIB loras on ZIT if you increase the weight to around 3, but not ZIT on ZIB.
Hey @J1B. So we're supposed to load our Lora trained on ZIT on the Load Lora node in the lower pipeline that's wired to JIB? And feel free to disable the one wired to Z-image Base? I trained my lora on ZIT (Ostris toolkit). I did try this but my Lora had zero effect. Stregnth was 0.8.
Because the tag is set to "workflow" for all files it Jib Mix ZIT does not anymore show up when filtering for "Z Image Turbo" and "checkpoint"...
In any case thank you for all your updates!
It's a workflow, so whats the issue? or what am I missing..
Yes, I see what you mean, I will have a play around with the ordering again.
I am an actual retard. It's a MODEL, not a fucking workflow. I am retarded perchance.
@Meowzilla No, its not really you, I have models and workflows under the same model page for convenience but civitai do not allow you to have different type labled of model/workflow on different sub tabs.
@J1B thanks daddy
V2.0 is still one of the best realism in my opinion. Good understanding, color and contrast and very high realism. My favorite Checkpoint at the moment.
Thanks, I have got a v2.5 model locally that is a merge of V1 and V2, I like it and I think I will release that, but just need to do more testing to see if I like the changes enough.
@J1B make a bit more nsfw, it's very interesting for many other creator i think
I am using your base to zit wf - with character lora - and cant seem to get sharp images. The first sample often corrupts a lot, and second sample is super soft. Any settings I can change to improve - tried every combo of VAE balance
I am not quite sure, someone on reddit posted that they had good outputs using a Z-image lora and a ZIT lora trained on the same dataset together at the same time, but seems like a lot of work and not sure they were talking about a character lora or not.
are you getting sharp images without the lora?
@J1B not really... I have been through a whole slew of lora making experiments - using onetrainer / AI Toolkit - so much noise about using fp32base to train etc... now mainly back to a zit 2 pass with detail demon... I do struggle to get nsfw clean images out though... now I am experimenting with abliterated text encoders.
@KurtcP yeah ZIB is good for intresting poses but a ZIT 2nd past is best for photo realism. With SDXL it took well over 1 year of train before models got really good at NSFW, so we are doing OK, I might focus on that a bit more next as a few people have asked for it.
@J1B yes
But I made about 5 loras since then - and a Lokr is by far the best results
for character lora, i am using a Z image base lora at 1.30 strength. Same prompt and setting with a ZIT lora only produces hallucinations.
noob here, how do I install it? I tried in the cmd but the git site doesnt exist anymore, and in manager sais the same thing. I put this folder in the custom nodes folder, but then what?..
The VAE Tech nodes? If you have unzipped and put the folder into the custom nodes folder then you just have to restart ComfyUI and they are installed.
@J1B I did exactly that and restarted but still says its missing..
@J1B yes is missing, git removed, VAE Merge doesn't work
@J1B Sorry I didn't read the description before, the node and even the VAE were there and I even already had it but it has different name right? it is the same "diffusion_pytorch_model"=="ultra flux for z image vae" Thank you, regards
@javierflores913 yes
Very good but unfornately not compatible with character lora :/ I Got som less detailed image with not good position when I use this checkpoint, everything is perfect with ZIT standard
If you were trying with Jib Mix ZIT V2 then give V1 a go as it is closer to the original ZIT, just less noisy.
@J1B I tried both 😅
i have really good results with a character (face) lora and a binary preprocessor controlnet (to get poses). better with v1 but even the few tests i've done with v2 can get good results.
as you say v1 is better than v2 with character loras but i am finding that i can do ok (sometimes better than ok!) with my workflow. my workflow has two sampling stages, dpmpp_2m/beta for both, one at 10 steps and the other at 1.3x upscale and 6 steps. and i have binary preprocessor controlnet running on the first stage. the first stage produces really cartoonish figures but when the second stage refines it, everything snaps back to realism. i am going to do some more tests but so far i like it! i would really love to see the 2.5 mixed version.
I get the following errors with Mod Manager on ComfyUi:
Failed to find the following ComfyRegistry list. The cache may be outdated, or the nodes may have been removed from ComfyRegistry.
comfyroll
chibi
Any idea how to fix?
i've tested my custom character LoRAs with various models, and they work best with JIB.
Hi, I've found a small issue in your workflow in sections "2. Two Stage ClownSharkSampler" and "2b. 2nd Flow Match Sampler instead".
In the KSampler (Advanced) nodes, the scheduler is set to FlowMatchEulerDiscreteScheduler. This entry seems to be invalid or causing an error (it might be a leftover from a specific environment/update). Changing it to a standard compatible scheduler (like normal) fixes the issue.
Just wanted to let you know so you can update the template. Great workflow otherwise!
Hi, could we get an FP8 for the v2.0?
Oh yeah, I can do, I am not at my PC until later next week.