unStable Revolution ZIB
Z-Image Base, improved with realism and NSFW
Recommended Settings:
Sampler: Res Multistep, others work too
Scheduler: Bong Tangent, others work too
25-30 Steps
cfg 4-5
shift 3
Use any resolution, i like 896x1152 the most
NSFW and SFW
Mostly for realism, but you can try some 3D or other Stuff.
Description
FAQ
Comments (28)
its doesnt work
like around 15% of world population. explaint it.
it shows a black box whenever i try it.
send an example
@Peli86 its just a black screen. I cant send an image through comments. I am using draw things app and using a turbo lora for it.
@Kitten123 check if your app is updated, on forge everything is OK.
@Kitten123 So, my checkpoint is the problem, not your app or turbo lora? Cool.
@Peli86 I got it working, no need for rude comments. I found the solution from developer.
@Kitten123 I'm also getting black images on Neoforge. It works great on ComfyUI though. Can you comment on how you fixed the issue?
@3r3de you have to look at custom.json for z image base how its coded and copy some code and add it to unstable model code.
@3r3de you mahmy have to ask the developer for help.
@Kitten123 Okay, thanks for the reply.
Edit: Its Sage Attention. If anybody else has issues like me on NeoForge, needed to disable SageAttention, then its work perfect fine. Great model btw! Thanks. @Peli86
@3r3de Thanks for letting us know! Had no problems in forge. how to solve this? just in case someone has a similar problem.
Didn't work for me, either in Forge Neo. Tried the proposed settings on top of multiple other combinations, with and without Lightning / DeDistilled Loras - images never really "converge", noisy and not aesthetically sound. I have Sage enabled.
@nofilter ZIB in general doesn't like sageattention ( https://civitai.com/models/2342797/z-image-base?dialog=commentThread&commentId=1094656 )
@mphobbit yes, I was unsure about sage, but downloaded the model again and rerun the example with --disable-sage in Forge Neo, worked like a charm. Thank you!
Thank you for sharing.
May we gonna have a distill/lightning version of it?
Here are distiiled 4 or 8 steps loras: https://civitai.com/models/2362961/z-image-fun-distill-8-step-lora
Is this a diffusion model like other ZIT & ZIB. Because I only can load it via checkpoint loader. It seems wrong to use other workflow just for this model. And is it legit ZIB?
IDK man, I tried 20 images, it all comes out overbaked, undercooked or body horror. Tried Res Multi, tried euler, cfg between 3 to 6, ive tried multiple combos of these settings over 25 images and i cant seem to get a good image out of it. I see the potential in it but I'm not finding that sweet spot. Any tips beyond what you wrote down already? do you plan an FP8 Version?
Try using natural language in your prompts. See example images. Strong descriptions of poses, especially when there's more than one person in the picture, can produce terrible results.
It's a general ZIB's issue (probably a payment for output diversity or its requirements for wild amount of steps ).
1. Check your shift value. On Forge you need to decrease it to 3-4 from default 6. If you're Comfy user, check ModelSamplingAuraFlow node.
2. Try a turbo lora: https://civitai.com/models/2362961?modelVersionId=2657506
3. Check schedulers. Simple and sgm_uniform definetely work, beta works worse.
4. Don't forget about negatives, remember old times with SD1.5/SDXL.
5. NLP helps with quality issues but it depends from picture, some pictures were improved, some not.
@mphobbit nice comment, thx. i had problems with negatives, made it worse in most images. i leave it blank .
@Peli86 it was random for me with vanilla, locally I had to use not so many negatives, but online had difficulties without them.
@mphobbit interesting. same sampler/scheduler/steps etc?
@Peli86 Yes. Typically ZIB likes to add extra limbs or break overall anatomy giving disproportions without negatives. For example in this picture it liked to add extra legs.
@mphobbit you are right. I had similar issues using zeroed out conditioning for negative (As I do with ZIT). Once I started using negatives like SDXL this model started performing way better.
This model completely blows away character lora trained with Zi Base.