@MonkeyForever Do you mind sharing it?

how did you manage to get a good result with character lora?

@MonkeyForever where we can we purchase this?

@MonkeyForever i'm having trouble 'trying' i'm on CPU , it's 'more grain' nothing more realistic here at least for me.

great question

@MonkeyForever let me tell you more after more try i done ,your Lora not only make real looking image it also breaking the loops and repetitive of result make completely random result like the on in Z image main model ( I hope you make more for (ZIT)

@MonkeyForever I didn't find it and I would like to read it

@MonkeyForever but ur lora will make base better , coz turbbo is finetuned from bas e so plz try if u find any good training on base , i love u 😍

Grab Z-Engineer from HF

@ronikush thanks dood

Your not wrong! Grrrrreat tip!

@MonkeyForever yeah seem so, i use ZiT Q8 GGUF and huihui text encoder still not perfect.... i dont find any fp8 or gguf model which is nsfw and works with multiple loras correctly

its great and much better for the general image-gen but i find V1 can give a sharper image sometimes while still retaining the amateur look

@MonkeyForever Link? I'm interested.

@MonkeyForever i am interested too

ZIT has major issues combining loras, try using Z-Image Base it is more stable.

@MonkeyForever It is hard to use, but it can make interesting outputs when ZIT gets a bit samey and boring: https://civitai.com/images/119966684

@MonkeyForever Yeah fair enough, everyone has different tastes, I think this lora reduces quality too much, makes it look like a phone camera from 2001.

@MonkeyForever also interested sent you a friend request on discord

@MonkeyForever fr

@MonkeyForever Good luck with that.

@MonkeyForever oh very interesting, i have been dabbling in the more traditional style of lora training, i think i got decent results but they still have issues with anatomy. plus i really do suck at natural language prompting i have been using local llm to refine it but it feel to cumbersome

That's interesting, because for me it's way harder to write tagging prompts rather than natural language prompts.

@meeowington i think its just habit i learned how to get sdxl and its many derivatives pony/illustrious etc to do things by learning how tags worked.

I am getting there but its not as easy to get certain effects because what you put before or after can change how the text encoder translates stuff... granted with is very much a me issue because i like making weird surreal artsy stuff alot :D

@purplelady Oh thats so cool! May I ask what loras/models you use for your surreal art? Because I cant really seem to find loras or models to do really surreal art so I rather stick to humans lol

@meeowington i trained a few loras but really didn't get anywhere, but here are a few that kinda work for it
https://civitai.com/models/282506/reticulating-noise-gradient-style?modelVersionId=2524468

https://civitai.com/models/2192426?modelVersionId=2468605
https://civitai.com/models/2334593/melancholy?modelVersionId=2626149

@purplelady Thank you! Also I'm trying out your loras right now!

@MonkeyForever could you send ne the PDFs?🥹

@MonkeyForever mind putting a tutorial out for this?

Use upscaling

use upscale like seedvr2 or just remacri

hoping soon lmao it's a great lora, keep up the great work!

@MonkeyForever he might be right, i did waaay to many tests and I'm like looking at every minor detail! This lora might work with single pass Turbo, but in double pass ZIB + ZIT (ZIT as refiner character lora booster) it gives doll like face, all the pores dissapears and removes details that ZIB made. Realistic Snapshot interfere too much on things it shouldn't interfere like eyes colors. So I just stick with Instagramification lora on 2nd pass with 0.90 weight and my psersonal character lora. Making lora that adds noise camera and 2000s quality photos doesnt make it more realistic.

@MonkeyForever There's nothing to prove, I did a simple test. I put in 1st ZIB pass my char lora 0.30 and FameGrid 0.70 (using FG somehow makes ZIB do what I actually want more accurate and natural) to set up character pose, camera angle, build up the environment and background details. 2nd pass ZIT my char lora 1.40 + yours on various weights, and Instagramification and some other realism ZIT loras (because I never said a word before about using ZIT loras on ZIB model since they are not compatible, not like ZIB lora on ZIT model) and from my testing RS v5 was in place 3. Mainly because of a doll face and changing eye color to blue, which Instagramification hasn't done.

ZIT model is not better than ZIB, they were made to do different things. ZIB makes detailed background more realistic, more variety of poses and you can use negative words which ZIT cannot. My golden mean when using ZIT as a 2nd pass was to use just the right amount of steps, enough for my character lora to properly shape at the right weight, but not so many that ZIT starts removing the details ZIB created in the 1st pass, because that's what ZIT does, it strips detail. And the right weight means the face doesn't look oversaturated or over-sculpted from pushing the lora too hard.

@MonkeyForever https://civitai.com/posts/27838047
here you go, thats what I meant by doll-like face using your lora, I don't know what else you might want from me, I already wasted too much time on you. And I saw your images, despite your prompt being 3 times longer than my, and on top of that sentences like "Their hands are anatomically correct with five distinct fingers" XD or "no grain" and tons other preemptive correction were quite funny - I personally think my photo is more realistic cuz as I said before pure ZIT photos most of the time looks like were shot by microvawe. I'm not saying ZIB is better all the way, these two models are good for different things, still using Base as your BASE model and refine using ZIT is much better.

@MonkeyForever na sorry, not the case. Dunno why.

@MonkeyForever okay, what bbox model would you recommend for automatic toes detailing?