This is simple detailer trained on model NoobAI-XL.
V-Pred versions work better on v-pred model. If you don't know what it means - read checkpoint description. If not explicitly stated - 99% it is eps prediction version.
V-pred related specifics:
V0.5 - just slap it on with weight 1.5 makes image better.
V1.0 - slightly alters image giving more details. Weird interaction that I found - it darkens dark. V-pred NOOB-AI is first SDXL model that I see that can make really dark imagery. ESPECIALLY with short prompts, like 1girl, dark, night. If you are feeling that that is not what you want - just be more descriptive. describe face, vlothes and you won't get those black on black silhouettes anymore. Btw that applies to the NOOBvp1 itself too. See example here: https://civarchive.com/images/47951726
On the upside it can darken stuff that base model has issues with darkening, like some poses.
V1.0
Merged with custom flat color lora, recommended weight is 1. Slightly alters image, especially with no prompts like 1girl.
V0.5
Weight range -6 - +6
I recommend using 1-2.
Description
This took waay more effort then it was supposed to. Ended up with tweaking OneTrainer to work with v-pred SDXL (thanks to guy from Discord for guiding me). But here it is:
Detailer V1, V0.5 merged with custom synthetic undertrained minimalist style lora. Please read main description for more info on specifics. Designed to be used with weight 1.
FAQ
Comments (34)
Stabilizes generations very well! Great job!
Thanks! Checked your images and have a suggestion. Please try adding lighting description and removing non v-pred loras, you'll be amazed by result. That's exactly the reason why I made this lora. See yourself: https://civitai.com/images/49504727
@Volnovik What tags for lighting do you suggest?
@Hannibal_Lecter Did not investigate deeply, but seems basically anythig will work, linked image has metadata. Even basic booru like: light, shadow will work.
Generation guide with all my parameters for NobbAI-Xl v-pred:
Using this seems to make my Gens far too saturated. Any tips to prevent this.
I have to look at image/prompt/ui setiings to give constructive anything constructive. If it is just the lora - most probably some prompt specifics. I advise you to read my article, it has one example of oversaturation which is not actually oversaturation + a lot of tips:
Made a lora to battle oversaturation of base v-pred model by oversaturating image https://civitai.com/models/1555532?modelVersionId=1760239 check it out
Workflows for comfy for vpred (generation, inpaint, upscale) with metadata saved in civitai format
Made a lora to battle oversaturation of base v-pred model by oversaturating image https://civitai.com/models/1555532?modelVersionId=1760239
Don't ask me why, but I use this v-pred version with NovaFurryXL (which is EPS afaik), and it still works great together. I actually prefer the look more than using the EPS detailer. Really weird. But I am happy shrugs
If it suits you it is good ¯\_(ツ)_/¯ I did some testing but not extensive one tbh. Dataset is the same, process and weights during creation are also identical. But base model is different and in my article you can clearly see that v-pred is more capable, maybe it inherited some of that. Judging by my last lora and spirals most of my creations turns out to have some unexpected positive side effect. Hope it will continue to be so.
A little oversaturated. Works exceptionally well with VAEs.
Thanks! Try mixing in this one: https://civitai.com/models/1555532?modelVersionId=1760239
I was like "Good god this keeps killing my images..." And this I realized I was using the wrong type for the model I was using by accident. The nightmare fuel you can create by using vpred on eps and eps on vpred is astonishingly creepy. Ima do it again
Not working, using Invoke: "The SDXL LoRA could only be partially converted to diffusers format."
I guess you should report it to invokeai, not here. Not sure how it works I. Invoke, but you can merge the lora manually (creating a separate checkpoint) and then load it there as a workaround
it just works!
Beautifull
a really great Lora, very nice details. i think it's a pity that the gallery of the eps pred version is mostly full of futa furry abominations, although you can generate such great pictures with it. Thank you for the lora.
Thanks, I tried to do periodic cleanups, just got tired cleaning that day after day. Felt like a chore at certain point. Also you have to take down your filters to clean everything, making it even worse 🤷
does this LORA work with ILL Lora like WAI 14.0?
Yes. EPS should give better result on common wai, vpred should give better result on wai-shuffle-noob. Purely because of their shared dataset. But see yourself. Users reported that it is not working with invoke, but that's on inwoke team. Since it is a pure concept cancellation I think it can work even with default sdxl to certain degree. But best suited for noob variants
I love how well this works :D
GOOD
Nice
This LoRA is excellent — Test on WAI-SHUFFLE-NOOB, compared to other LoRAs that enhance details, it does not add excessive noise.
Thanks!
Great
great
pretty good! Compared to other details enhance lora, it won't change art style.
ty its so good
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.