Here we are with a epiCRealism Version for ZImageTurbo.
Since it is a pretty awesome base model on its own, i tried to give it some of the epiCRealism feeling. Try it out yourself.
I'm using my loved ForgeUI Neo, so i can't say much about ComfyUI Workflows or stuff, but would like to see what you can come up with.
Under ForgeUI i ♥️ these
Settings
CFG: 1-2 (2 sometimes give better/cleaner details but adds generation time)
Steps: 9 (8-10 are good)
Sampler: DPM++ SDE (Slow but best realism)
Schedule Type: Bong Tangent (Pretty cool variations, Beta also works good)
Time Shift: 6 (3-6 where 3 can sometimes give pretty nice compositions)
Keep in mind that every tiny bit in the settings (Steps, Time Shift) can produce really different outputs.
Description
FAQ
Comments (53)
Great to see such a ZIT model from you. Was using your EpicRealismXL model quiet often within the last months. Will give that a try!
But are u able to provide some information on the training data you used for that? Did you tried to fix some of the typical nsfw issues that come with ZIT?
Hey Sir, thank you for using my model. At least i tried to refine for nsfw but it was getting worse instead of getting better. To be true it is pretty hard to train this distilled model with concepts.
Stunning style, magnificent results. This will now be my top choice for image generation. Thank you for your hard work.
TY! Hope it lasts
Where's the setting for time shift in forge neo btw?
Choose UI preset lumina in the top left, then you have "Shift" Setting next to the "CFG" Slider
finally it is OUT ;)
this is Way Better. Much realistic it gets.
is it working with lora character??
Insta-download!
It's a great model. Do you have any plans for fp16?
No
amazing model, super realistic, super amazing..........
TY! ❤️
Looks good will be trying it out
BF16 please
He probably trained it on fp8 so converting it back to bf16 would be a waste
There isn’t much of a difference, and from my testing FP8 actually performs better while keeping generation time and storage much lower.
"BF16 please" It seems someone has more memory to waste
Does it support well male genitals, or generally speaking NSFW?
nope not much good results for nsfw
@IchigoKurosaki2408 you must be blind
@anon1962_ Yeah, maybe I'm because I don't see good nipples and pussy in any of the images I generated
@IchigoKurosaki2408 Key words "I generated". People can always complain, could be prompt, sampler, other settings, etc.. so many things could make it your fault vs the models. Keep learning
@anon1962_ Fair
there should be a 6-bit version for Low-End GPUs
Honestly though, what i like the most of this finetune is that it will not default to asian ethnicity. Good work :)
Hi, can anyone link a nice ComfyUI workflow for this? Want to try it out, thanks! 🩷
Since have not used Z-image checkpoints yet really, pretty sure these have some nodes/decoder or whatnot thingies that i better just leave to missing nodes autoinstalling than trying to build one myself .. 😅 so a link here or dm, either would be greatly appreciated.
Seems like a downgrade visual-wise. Hmm.
To stay close to base model with improved style to my likeness, better use my Loras instead.
Is this trained or a merge? Cause it feel like a merge when I use loras...
Since we only can train Loras for Z-Image it is finetuned by trained Lora’s of my epiCRealismXL dataset
Why it’s not working with Draw Things app?
It does. I use a Mac Mini M2 with 16GB and it works perfectly fine. Maybe you need better hardware or change settings like CFG scale or another sampler.
Z-Image base is out... epiCRealism update when?!
**totally normal none crazy twiching**
He's gonna nail it for sure.
@6vidit9 need it soon!!!
Hi,
I see you mentioned Time Shift: 6 as a recommended setting. However, I cannot find this node in my comfyui, dragging and dropping one of your provided supply images also did not fetch me any information.
Could you please tell me where I can find this setting/node name. I am new to Comfy : )
Thanks for your time.
I think it might be the usual shift value, that is bveing refferrred to. So use modelsamplingauraflow node and set it to 3
@benshapiro679 ty
@the__Realist I am actually not sure anymore, because the scheduler that the creator recommends (Bong_tangent) is unaffected by shift. The only use of shift ( the shift i refer to ) is to change/shift sigmas. But the bong_tangent scheduler is designed such that its immutable to shift values.
So i am not sure, what the author means when he says "timestep shift"
@benshapiro679 alright hope he responds in that case then : )
@benshapiro679 as I said, I use Forgr Neo I don’t know how to set it up in comfy but you can just use a zit workflow that works for you in comfy
Load Diffusion Model -> ModelSamplingAuraFlow -> KSampler
Honestly, this is hands down my favorite checkpoint so far. After trying tons of models (Redcraft and Beyond reality my fav in top 3 ) and generating hundreds of images, this one just nails the vibe — it adds that perfect amount of noise and texture. Most others give you super smooth, plasticky skin unless you use a LoRA, but I’ve been getting amazing results with this one even without it :)
Thank you very much! Your work is stunning as always. Really appreciate your work with my model. ❤️
@epinikion Thanks for making such an awesome model—without you, I’d probably keep jumping between different models. But now I’m all in with yours, and I don’t even bother with others anymore. That "Er_sde" setting? Game-changer. The images are way better, and it works insanely well with your ZIT LoRA—plus, it plays nice with other Loras too. You’re the real MVP, the GOAT, the legend—my friend 🤘🥂
Does this model need a separate CLIP load