Ten arguments immediately come to mind as to why you're doing almost everything right and just needed a little encouragement from him. But I don't think that's what you meant.

z-image-turbo really does not like loading multiple loras. but i've found that negatively weighting certain loras (style loras have worked best for me so far) will allow you to positively weight more loras. i load amateur photography and analog core at -0.55 each and that allows me to load more than one lora. although if you're trying two character loras or more they might be fighting for the same space, so idk how it'd work in that case.

@Soup_Fusion hey,
thanks for the response. I only ever use one character and one action/style. one character and xxxmystic, i cant seem to get working...
so you recommend that i put an extra lora on at negative? and it could make mystic and the character work?

@timothyjsmith that's what i've been doing and it's been working for the most part. I haven't really used any character loras so i'm not sure it will work in your case but i've been able to load loras together that previously wouldn't work. i think style loras might paint broad strokes across the weights so by weighting them negatively you have room for the other loras.

maybe ZIT creator just likes bdsm

I agree actually most of the images I make are 1024 x 1536 it's a great LORA tho. Definitely the best NSFW one at the moment.

@alcaitiff how do you recommend stacking Lora sometimes I use it on full 1.0 settings if I'm using one like yours that has lot's of different things. But other ones that specify on one thing and a one part of the body I would reduce it to to 0.50 or lower As I asked Gemini and it said the road of thumb is for it to be around 1.0 to 1.2 overall on a distilled model like z Image. Also I increased the steps to about 12 to 16 or even 20 to give it more breathing room to add the detail and Lora in and works quite well even on the default workflow from ComfyUI. I usually use Res_2m and the Flux ultra vae that Jib recommended on his workflow. As it does make a big difference I've also tried NAG Normalized Attention Guidance Which works pretty well, although the difference isn't that much apart from the overall look.

I think the biggest difference that I've got you just found out today is actually doing the prompt in Chinese (simplified Chinese I think) As I think these Chinese image models work better using the native language as I found it as amazing prompt adherence I've been using a LLM to prompt the English version to Chinese, making sure it follows the right structure from the Fal Z Image prompt guide (I set up my own custom instruction on a LLM chatbot to be a master prompter).

All these things I think have helped, although I have just stuck to the original default ComfyUI workflow as I find that that is sufficient enough especially with your new awesome lora my favourite chest lora too is the one from peterkickasspeter (natural b**b one) although I've noticed that when I asked for a rear view, it does muckup a bit as I don't know if the lora has enough rear view training data.

But yeah can't wait for the base model and I hope you have a great rest of the year. Take care man

@alcaitiff Thank you for the explanation, appreciate it :)

most people wont be using base. it is a lot bigger model. but it will mean we can train loras with better quality and finetune base to distill a model that is as easy to run as ZIT with better NSFW understanding.

@alcaitiff that appeared huge on my new display :)

You mean, you're not into penis hands?? >.>

@Latent_Dreamscape This is because ZIT 'isn't censored' /s

(This is a consequence of it being censored by deleting male anatomy from the human body)

I don't think many LORAs will. You're better off using a checkpoint or just waiting for the base model. As with this turbo distilled model, you can only really have about two LORAs or more but you have to make sure the overall LORAs strength is up to around 1.0 to 1.2 overall because if you go beyond that's the prompt adherence will start going and also it may look a bit deformed. and increase the steps to about 16 to 20... I'm currently trying to make a character Lora. If you really want to try, you could try with this lora at 0.50 to 0.40 and use your charecter Lora at 0.70. It's a bit of a balancing act so you need to prioritise what you want in it image and then decrease. It will increase it depending.

You are incorrect, it works fine with my consistent character LoRAs, I am a creator check my profile and use my workflow and ready made consistent character lora for testing. it is working like a charm.

Edit: sorry forgot to mention, try using the latest loras, the old ones doesnt work.

@coolstrad Can you drop the link to your workflow, i cant seem to find it. Thanks mate

@nobitaadidas My workflow and prompts are already attached to the images of the character you are looking. To use it, download the image, open the image on comfyui and you will have full access to my workflow and prompt.

@coolstrad 100% -- There are still many that have not heard that some weeks ago now the "de-Turbo" ) (de-distilled) version was released by Ostris. Any Loras trained on the de-turbo are liberated! -- I don't even think a base model will be released now as the De-Turbo release has already done the job. That's probab;ly why its been so long without a base release. You may see an official base release with the Z-Image edit.

Dude. Don't make a fool out of yourself. Any time you are having problems with a LoRA or a Checkpoint, ask nicely for help. Don't come to a page full of people who have clearly shown a thing to work perfectly fine and go, "DURRR! IT DON'T WORK FOR ME! (Fart. Fart. Fart.)"

@alcaitiff Averything okay with prompt and setup because everything works WITHOUT this lora. Just activating it breaking generation.
Seems your lora incompatible with most models.

@jazgalaxy581 actually reporting models not works - only way i ever got any respond.

lora makes image into disfigured mess too. Workflow from ComfyUI

I dont think this is an issue with your lora, I am getting this with most loras. Anything above 0.25 strength makes things convoluted. I am using your workflow. If i set the Lora strength to 1, for example, it basically looks like a water painting.

But I am seeing this with all z-image-turbo models and loras, so I dont think it is you!

@alcaitiff Happy to, but it might not be until monday or so. Do you have an FP8 Z-Image-Turbo checkpoint? My laptop gpu only has 12GB of vram :D

@alcaitiff 

I did a little digging. I definitely dont think there is anything wrong with your Loras.

I dig through the comments on some of the checkpoints, and descriptions and there definitely seems to be something with fine tunes and using Loras. Generally in the past this hasn't been an issue for me.

However if I use a checkpoint, and someone made a lora with the the standard z-image-turbo, it doesn't seem to mix well.

Using your Lora, and the standard FP8 z-image-turbo, I get WAY better results (though something is still off) than if I use someone's z-image-turbo checkpoint here on civit.

@alcaitiff Thanks, makes sense.

Do you know if running the standsrd z-image-turbo in FP8 and running a lora would cause an issue? My nvidia 3500 ada has FP8 tensor cores, so it seems like a good idea.

@alcaitiff thanks! Sorry not to hijack the thread...

But why not use FP8? If you have a 4000 series card or newer, you have FP8 tensor cores which will be faster. (And obviously FP8 != Q8)

@alcaitiff Using your lora makes my comfy get disconnected. not sure if your lora is the cause. but everything else works. just when I want to use your lora this happens

@alcaitiff he is talking about his local comfy server. happens to me too sometimes. regardless which lora/ckpt I use.

@alcaitiff it's not an OOM issue. I got 192GB and it even happens when the server is "idle". but that's rare. it mostly occurs quite randomly during generations. the terminal window just says "press any key to continue".. and when you do, it closes the terminal. so you have to reopen comfyUI. what I noticed is that even downloads get stuck on certain steps during the process. like there is no network at all for a few seconds. no clue what it causes. I kinda got used to it.

@alcaitiff I wish. haha. 4090 (24GB VRAM) + 192GB RAM (glad I bought those in may 2024 for 744€. they are now 2400 in same store :D)

@alcaitiff @janssenbutton Yeah I meant the server.
Anyway I found the source of problem.. while I was searching about z-image and looking for workflows, I saw people talking about Lora Loader.

It seems Z-Image doesn't work with normal Lora Loaders, that's why people had to use a node pack called CRT (https://github.com/PGCRT/CRT-Nodes). then the Comfy's main Lora Loader was fixed and you don't need the CRT's LoraLoader anymore.

but as you see here I am :D. So I gave it a try and... it's working now.
The node is called Lora Loader Str (CRT)

I'm not technical, I don't know why was that happening. but It was not oom I have 16vram + 32ram. (I don't know maybe it is xd) anyway that's the fix.

You won't have disconnect issues if you stop using comfyui

@alcaitiff Sorry, I'm super needy. But, do you also plan to do a Flux Klein version? 🥹🤣

Z-Image

Do you find bro? I want it

LORAs are chained together. But my loras didnt work well with this one......

do let me know if it worked with stacking other loras

@LightSleek yes it works well with stacking LoRas. I have used two other Loras along with it for cinematic and detailing. But you have to adjust the strength. If you put all the loras at 1.00 then it creates images with waxy skin texture. Ajust according and other Loras strongly influence each other so you have to try the same prompt with this lora alone and then add one Lora and try again. Also adjust sampler, cfg and resolution. But a the bottom line, this works with other Loras just fine. I used rgthree node for stacking Loras instead of adding two separate LoadLoraclips

@akringerzekrom656 thanks a lot, do you have a workflow that i can use for that?

@LightSleek yeah sure. But how do I send it?

@akringerzekrom656 you can post it on gofile.io its temproary ill download it

@akringerzekrom656 Which workflow do you use to work with this??

@LightSleek Which workflow do you use to work with this??

@renmustafin8709 use the standard workflow from templates in ComfyUI. Add LoadLora checkpoint and connect it to Ksampler. You can also use my workflow


https://huggingface.co/datasets/alexrider656/z_image_workflow

@LightSleek I posted it on huggingface. Let me know if its working

https://huggingface.co/datasets/alexrider656/z_image_workflow 

@akringerzekrom656 I don't know, maybe I'm too sure about this, but it won't open, it says "Unable to find workflow in z_image_turbo_gguf_workflow_with_rgthree_lora"

@akringerzekrom656  It dissapeared again YoY

@alcaitiff Thanks for responding, I sent you a DM!

Simple, dont use ZIT. It doesn't like multiple LoRa's. I know SDXL is old at this point but you'll have much better results with a character lora usecase.

prompt for Asian women. its easy.

@alcaitiff thanks for your wonderful work, really love it. thumbs up

@alcaitiff ty

well said alfa .. yes its another clueless clown just trashing peoples hard work !

personally i'd like to thank you for your hard work and continued development of the lora.

IMO the main reason the lora is giving bad results for you is because your more than likely using a checkpoint with the lora already baked into it , therefore increasing the strength of its effect and causing said bad results.

So keep up the good work alfa .. i'd send some gold buzz but since civit stopped giving me any i can't do .. now if it was blue buzz i'd be giving you thousands 🤣🤣🤣

I retract my review for the Mystic XXX. This was a mistake on my part, as I tried to use it with non-standard versions of ZIT and those were prone to their own weaknesses in working with LoRAs and I blamed Mystic XXX. I wholeheartedly beg for pardon from the creator. I feel sorry if my review mislead someone beleiving that this LoRA is bad. What I can say, the reason for this bad review was because every other LoRA (around 6 or 7) I tried worked well but when I tried Mystic my output was botched so I thought this LoRA is the culprit. But with standard ZIT it has been amazing and I am loving it. Thank you alcaitiff and sorry again. I wish you best.