Realistic Slider
My old prompts from Z-image made 3d-characters from time to time.
So i made this slider to fix that problem.
This was trained to do the same as the word "Realistic" does in a prompt. But with this lora you can increase the strength of that word, in a sense.
Recommended range: -2 to 3
Description
FAQ
Comments (35)
This is great
The best! Works great, adds a real difference.
It's okay, thanks for all the work you've done on z-image. ♥
https://civitai.red/models/2728234/krea2filterbypass?modelVersionId=3066812
No custom nodes are required when used together :)
The node and that LoRA is not the same. But i will make some sample images to compare the results.
From the testing i have done, the Projector Scale LoRA does work pretty similar to the Node. But the LoRA very prone to extra hands and arms. And the LoRA also follows prompts much better.
@Loraholic that lora gives a black screen on my gguf setup :(
@brahianvalles I have not tested it on gguf. But its made with the latest version of ai-toolkit. I'll make sure to keep an eye out for potential updates for ai-toolkit if i decide to make more LoRAs in the near future.
@Loraholic not sure if you already used it this way, but the LoRA should never be used over 0.4 strength. And usually works in a lot of cases on 0.1-0.2.
In general, for me at least, it produced more consistent prompt adherence and less quality degradation than the node at 1-4 multiplier, with less "refusal" results.
@Loraholic
https://civitai.red/models/2728234/krea2filterbypass?modelVersionId=3066812
I think someone uploaded the original
@brahianvalles Just curious, why do you use gguf? If you use comfy, you can use models that is much larger then your VRAM, you know? Comfy just dump parts of the model to ram, and the speed is almost identical to having the whole model in vram.
@kennedysworks could you edit your original Discussion message to link to the correct LoRA. So people dont use the Projector LoRA by mistake, if they dont read all the comments here.
@Loraholic complete
I hope you port all of your z-image loras to krea. Your work is legendary. Thanks!
Most of my LoRAs can be prompted pretty easy in Krea, so im not sure they are all needed.
But a few "legends" will probably be remade in Krea. But i also have a hard time making stable slider LoRAs for Krea so only time will show if i find a good setup.
@Loraholic i can feel krea2 follow prompts better than zit
@ping26 A LOT better imo, but also very different to prompt for. It can be a bit tricky compared to Z, since Z does best with a long novel. But Krea also shine on very short prompts, and it can fall apart on very long prompts. At least from the little i have tested so far
It is a great pity; I hope to continue using the new Z-model LoRA.😣
Why not jump ship to Krea-2?
Its genuinely better in every single way. Better at prompts, looks way better and more realistic, much more robust to sampler/scheduler combos, nsfw out of the box, etc. The list goes on.
I use the nvfp4 and mxfp8 version myself, on my 12GB 4070 ti.
nvfp4 fit in vram and takes 17 seconds for a 1024x1024.
mxfp8 takes 22 seconds since its partally loaded to ram.
And even the fp4 looks much better compared to Z.
@wyxzddsjj919 I'm actually interested why you don't want to switch to a model that's roughly the same size, faster and better in any way. It was an insta-switch day one for me... it makes no logical sense to me why someone will not switch to Krea-2 if they are a Z-Image user.
@DriftPixel probably lack of lora's early
More Krea-2 LoRAs please!
Thanks for your great work, just a request please: This model really needs Breasts Sizes slider. it would be great if you would make one. 🌹🌹🌹🌹
Hehh. Its in the making. I want to make sure its top tier when i first release it.
So far the best lora I've found to reduce getting unwanted anime and 3d images in my outputs
Works great but be warned it changes the faces a lot, you get mid/average women the higher it goes.
This was just a rushed release the first day to get the model somewhat working. This was before there was any viable options for NSFW LoRAs.
I have trained a V2 that i just need to test that should be way more stable. Probably releasing later today or tomorrow.
Could we get a skinny/fat slider for Krea 2 please? Thanks so much!
@olternaut
I have made one, and released it. But i unpublished it because someone mentioned it can be against TOS regarding eating disorder etc.
And i dont want to take any chances so i will leave it unpublished until i get a answer in my ticket with staff.
@Loraholic Eating Disorder. That's a silly thing to go against TOS but okay. lol
@olternaut age slider "loli rays" totally fine. Weight sliders? Better be safe in case the one female AI gooner with an eating disorder gets upset
Anyway, you can find Loraholic's unpublished LoRA on civarchive just google it
Your lora sliders are amazing! Can you make a lora slider to control background blur? For example, -3 would blur the background (bokeh), +3 would completely sharpen the background. Is this possible?
I tried to use it with Krea2 identity, many tests at -2. Negative value to make it look less real don't work on real photos, or at least in the few I tested. But making it more real for 3d render, using positive values works.