2603 ZIT — By Stable Yogi
Fast photoreal Z-Image Turbo with a natural, real-life look — believable skin, faces, and lighting instead of the over-sharp "stock studio" feel. ~8-step renders, SFW → spicy.
Highlights
📷 Natural, real-photo realism — soft, believable lighting
🧬 Clean skin texture, crisp faces, reliable anatomy (holds up on complex poses)
🎨 Versatile — portraits, full-body, editorial, candid
🧩 Excellent LoRA base
What's new
Retuned for true-to-life realism — dialed back the over-processed studio look
Ships in fp8 (Forge + ComfyUI) and Q8_0 / Q4_0 GGUF (ComfyUI)
Recommended settings
Steps 8–9 · CFG 1.0 · Sampler Euler · Scheduler Simple
Resolution 896×1152 or ~1024 square
Which file — Q8_0 GGUF = best quality (ComfyUI) · fp8 = Forge or ComfyUI · Q4_0 = smallest / low-VRAM
Want the stronger, higher-detail Pro build? → Get the Pro version of my models here
Description
FAQ
Comments (30)
When will we get Z Image Base by Stable Yogi?
Work in progress.
@Mr_MAQ The benefit of Z-Image-Base now being out is, that loras trained for Z-Image-Base are very compatible with the Turbo checkpoint.
I have trained a few character loras on z-image base and by bumping the strength up a notch to 1.25-1.5 they look perfect on z-image-turbo without getting ruined by other loras and/or the checkpoint.
I just tried using this checkpoint again with my newly trained loras and it works great.
Now i can super easily do NSFW with the character loras i trained. Before it looked really bad because the loras kept bleeding badly into the checkpoint or other loras.
And this way it looks so much better than on z-image-base itself since it's not really finetuned in any way and most likely will take a bit of time before good finetunes come out, just like it did with SDXL.
@vslinx That’s great info, thanks! What settings are you usually running on Z-Image Base for these results? I'm getting bad result specially on full body and long shot, face getiing distorted.
@Mr_MAQ you mean for the training or the generation?
I'm really unsatisfied with the results of the z-img-base generated images, which is why i tried using the z-img-turbo checkpoints instead with the z-img-base loras, and that works way better.
It'll take some time before a good checkpoint comes along that elevates z-img-base to a good useable checkpoint imo.
For training i use the Ostris ai-toolkit with these settings:
Transformer & Text Encoder -> float8
Target type -> LoKr
LoKr Factor -> 4
Data Type -> BF16
Batch Size & Gradient Accumulation -> 1
Optimizer -> AdamWBit
Learning Rate & Weight Decay -> 0,0001
Leave the rest as-is, uncheck every box on the right in the "Training"-Area and then activate only "Unload TE".
Unfold the "Advanced"-Section and activate "Do Differential Guidance" and set the value to 3.
I save every 500 Steps and keep the last 8, i do around 7000 steps in total, you can also do less if you increase the learning rate.
I use no triggerword and no captions whatsoever in the training data, in the dataset region i only activate the 512 Resolutions.
And then i deactivate sampling to save some time.
I have done 3 character loras with different settings and when i did these it always came out the best. These are my settings you can paste into the "Show Advanced"-Section of ai-toolkit.
@vslinx Thankyou very much. I getting vad results on z image base model. I used almost evevry sampling method with higher steps still not get desire results. I only got good results for closeup and macro shots. I used res_multistep\simple
@Mr_MAQ Yeah i'm sadly having the same experience ^^
No matter the resolution, 2nd pass, enhanced samplers/schedulers or even seedvr2 - the images just look grainy, bad and with bad details.
Only really closeups of the face look good.
Until good ZIMGBase checkpoints are out i'll stick to Turbo and just use base loras :D
@vslinx All we can do is wait fr 😶🌫️
@vslinx correct me if i'm wrong, but from what i've read people don't understand that base is not as good as turbo because turbo is like a completed checkpoint that works right out of the box. The reason why base doesn't produce as good image is because it's like a soup. Base z-image is just like water in a soup. People still have to put their desired ingredients in to produce something. And turbo is a finished soup with a custom set of ingredients.
@Melodic_Possible_582589 you‘re not wrong.
ZIMGBase is the foundation for turbo, but turbo was distilled using human intervention. Meaning the outputs were actually evaluated by people and then decided what the desirable output is. So it‘s very good at realism because it was decided which output is desirable for this. That‘s also why it‘s not very good at variety and therefore bad at combining multiple loras.
But just like SDXL it‘s a very good and solid foundation and with re-training of desirable outputs while maintaining the variety it offers, it has the potential to be the next „foundation“ model that most models could be based upon.
2602 - BF16 on the way? :-D
Yes
Bf16, FP8, Q8, & FP4 coming here on Civit
Fp16 - Q4 - Q5 - NVFP4 now available on My Patreon
@Stable_Yogi Cool, thanks! I'm digging your work lately my friend!! :-D
That's a truly exceptional checkpoint of outstanding quality, you've done a really great job there :)
Exceptional work. Any chance we can get the NVFP4 or FP4 versions? Thanks Stable_Yogi. You're the GOAT!
Thank you for the kind words.
FP4 and NVFP4 incoming !
@OzzyOsman NVFp4 + Fp16 + Q4 + Q5 are now available on My Patreon
Bf16 , Q8, Fp4 scheduled for Civit
@Stable_Yogi Awesome! Can't wait
@Stable_Yogi when did patreon become something to sell your shit / paywall your shit on? wasn't it supposed to be support a creator by free will with minimal perks ? p2w reached the creative community (we had commissions before, but thats non comparable)
@BlackForestGypsy Ah yes, the golden age where GPUs ran on pure passion and electricity was paid in compliments.
Some drops are free.
Some drop early for supporters.
All of it funds the next round of madness.
Nobody’s locked out — just different doors, different timing.
Pick the one you like and step through.
Now if you'll excuse me, I must go convert kindness into VRAM.
@Stable_Yogi I guess you're right. Haven't looked at it through that lens. And I did download this zImage model from this thread with buzz. Just don't like how things have changed online and everything is being sold. Back then you did good work and money came.
based on your description, it shows tag style prompts. does this checkpoint prefer tag style or story telling prompts?
Its natural Language ready as default, Ii tried injecting Tags, and they work, but not that accurate for now.
needs more work.
Thank you for sharing!
I tested 2602-NSFW-FP8 and in my opinion, for an FP8 it hast good color and contrast and a very good realism. At the moment it is one of the best with good standard nsfw paired with realism and quality. I tested tags and it reacted to them, but far from a Pony or Illu-like effect but I also don't think something like that was expected. HOWEVER, as you would expect from an LLM, it tried to make the best of it, and some very nice pictures still resulted. But perhaps it's also due to the tag understanding that I feel the LLM understanding is weakened. I like it for standard NSFW and can only recommend it as an FP8.
Hi, thank you for sharing so far it has been good. I am having one little difficulty though I can seem to generate anal, I have tried various ways but it refuses any help would be appreciated. Cheers!
The Xmas checkpoints have the best looking girls and realistic lights/background. Seems to mostly reflect modern looking instagram girls.
The 2600s seems to generate elastic skin and old looking faces from the early 2000s era lol. Does anyone else experience this?
idk, but using same prompts i used for qwen models; i get disfigured sex images and they hardly follow the prompt.
Noticed!! Whatever is going on in Qwen being such a larger model really makes a difference. Apparently, you do lose things when you go to less than 1/2 the model parameters. Some of this is obviously the difference in prompting styles the model was trained with, but not all.
Unfortunately, I'm not sure Z or Klein will move much past 1GirlUniversity, and at the same time Qwen is resistant to success.