Norman Rockwell style LoRA for Pony V6 XL and Illustrious-XL v0.1.
For best results, prompt using Booru tags.
See training data for image list, tags and training config.
Description
Most Common Bucket Sizes: 1088x1152 (x53), 896x1152 (x24), 960x1152 (x21).
I used the same script for as "V6" to train "V12" on Animagine XL V3.1.
FAQ
Comments (9)
I've been waiting for a good SDXL Norman Rockwell lora for ages! Thank you!
I have been testing this LORA (32/32 version) and I have not been able to generate images correctly. I use PonyXL v6 along with the LORA from Paseer XL Lighting (https://civitai.com/models/317638). Until I decided to try it without activating Paseer's LORA. That's when Norman Rockwell's style has started to properly generate for me.
I really like this LORA, how well it manages to imitate N.Rockwell's style.
But for reasons of limited GPU memory, it is better for me to use the Paseer accelerator. Until now I haven't had any problems with any LORA; but with this one it does not generate the style correctly.
However, if I don't use the accelerator, I have to put a minimum of 20 steps, while with Paseer I only need 6 steps.
Could there be a solution that allows me to use this LORA together with the Paseer one?
Congratulations on such a good LORA, knowsnose!
Thanks! I just tested out Paseer in A1111. All I needed to do was lower cfg to 3 and it seemed to work with 6 steps (although 8 worked better). I also used Euler ancestral as the sampler, as another one I tried didn't work at all. Does this work?
@knowsnose I'm sorry for answering delay. I have been testing with the parameters you recommend and it seems that what prevented it from generating correctly with this LORA was the sampler. When generating with the 'Euler a' sampler it has improved a lot!!! Thank you very much and again a thousand apologies for the delay in responding to you.
Great gen!
Does this work with image to image?
If you want to do hires fix or image to image, it should work just like any other model, as long as you use a denoise strength between 0 and 1. However, I didn't adjust the max and min timestep in the training settings, which is apparently supposed to help with image to image. Therefore, if you want to do something much more faithful to the input image, you'll need to also use something like control nets (e.g. xinsir) or ipadapter.
Amazing style absolutely love it, thanks!
Excellent style, very easy to create high quality images
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.