SDXL ONLY
weight: 0 to 5.0
positive: more realistic
negative: less realistic, cartoon, painting, etc
I noticed the more bizarre your prompt gets, the more SDXL wants to turn it into a cartoon. This helps give you the ability to adjust the level of realism in a photo. All images were generated without refiner. I refuse.
If you like my work, I am not asking for coffee, but a kind review is always appreciated.
Description
FAQ
Comments (19)
Those are the best example images ever
Nice! Question. Is there a trick to training sdxl loras? Or is it actually as vram hungry as it seems.
I'm retraining a character lora right now, with just a single batch and I'm sitting at 23.2/24.0 GB used on my 4090.
I have a custom script that is hacky as hell right now for sdxl, but from what i have seen, so is everyone else's. For these, I am not training off images, just doing latent space manipulations. It trains, but I can unload the text encoder, though I think sd-scripts is doing that too. The only issue on training I am having is huge vram spikes on back propagation. It peaks at about 18GB during that, and then I unload again and it drops down to 6-7gb, This is with lion8bit.
you guys need to use Prodigy.
@nonamebrand Does prodigy even work on SDXL? It is so ram hungry that I didn’t even bother.
Thank you for releasing this. Haven't had a chance to experiment with XL yet but it's super exciting to see models getting dropped already. Keep up the great work, folks!
Mmm. I am not getting the realistic results I was expecting, as you can see in my uploaded pictures. Your bear looks more realistic than my tiger.
EDIT: scratch that below, I only saw the first picture. On your scale you are going up to 1.0. This is designed to be more granular and go up to 5.0. 1.0 is a low weight this LoRA
You put a negative weight on it. Negative makes it less realistic. Change this <lora:sdxl_photorealistic_slider_v1:-1> to <lora:sdxl_photorealistic_slider_v1:2> or something like it
@Ostris Yeah, I did that. I tried a different prompt (The kangaroo one of the gallery) and it worked better.
Anyway, thanks for creating this fix. :)
so you wont do any stable v1.5 loras in the future ??
I will definitely keep doing 1.5 LoRAs. I don't think 1.5 is going anywhere. It is small, powerful, can run on a macbook,etc. sdxl is a beast and really has less of a need for custom LoRAs as its prompting is so good.
@Ostris niice, dude i have your profile page on my homepage XD i check your loras everyday, thx keep it up
@solom123 Thank you! That means a lot, truly it does!
@Ostris Thanks, based on my tests, a lot of low end to mid end systems can only work with SD 1.5 workflows(without running into OOM), even with my RTX 3060 desktop, it takes about 16-17 secs to generate an image and that's without controlNet. With SD 1.5 models and LORA's, it's only around 4-5 secs, if its with controlNet around 9-10 secs, which is marginally faster specially in the conception/sketch/drafting phase where there's a lot generating that needs to be done.
For SD 1.5: at 20 images at 5 secs per generation = 100 secs or 1 mins 40 secs
For SDXL : at 20 image at 16 secs per generation = 320 secs or 5 mins 20 secs
The extra resolution isn't really that important until the final phase of processing the image.
One thing I do love about SDXL is that it seems to have better base prompting.
great work!!!!
I also found that when your prompt words are not included in XL's training data (or rarely refer to the data),XL likes to use cartoon pictures to deal with them, otherwise it is very good. For example: cats, dogs, may only need to write one word, the effect is very good, very realistic. But in a Velociraptor, it's a cartoon-like style.
I'm sure your weight loras are good, but your weight scales go off the chart. What ui you need to use to go over +2?
As far as I know of, I think only invoke limits this. Every other one I can think of lets you do whatever you want.
Hello,
Thank you for this. I have a query.
I am new to this AI generation, I would like to know if your lora is only for cartoonish/anime kind of models/renders or can i use it with a realistic type of model such and RealVisXL.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.