Ditto's resources:
https://www.pixiv.net/en/users/62713424
https://www.patreon.com/DittoChad
===
NoobAI v-pred update Read Carefully!
DISCLAIMER (applies only to NOOB-AI version): To achieve best results, use settings that I use, checkpoint that I use, HRfix settings and upscaler that I use, enhance prompt string that I use, negative prompt that... You got the Idea.
Trained on https://civarchive.com/models/833294?modelVersionId=1190596, dataset consists of latest Ditto's artworks and some older artworks which in my opinion depict his style well.
Was tested on https://civarchive.com/models/1201815/nai-xl-vpred10-2d-accelerated, both "Final" and "style2" versions. Better CFG for "Final" is between 2.1-2.5, for "Style2" is 2.5-2.75. The sampling method I use is Euler Dy CFG++ and the scheduler is KL optimal.
You simply can drag and drop any of the preview images to your webui to use my settings or to adjust it.
Doesn't work on model the Lora was trained on (and doesn't perform well on other v-preds, I tested few of them). If you find a way to make it work on different base model as well as on 2d Accelerated one - share your generating parameters with others.
The model gives good results with and without enhance tags string:
masterpiece, best_quality, very_awa, absurdres, very_as2, hdr, High_dynamic_range, rich_details, clear_shadows_and_highlights, intense, enhanced_contrast, highly_detailed, amazing_quality, very_aesthetic
I suggest you to keep negative tags like this:
worst_quality, low_quality, bad_anatomy, blurry, text, Signature, 3D, pixel, watermark, fused_fingers, mutation, amputee, monochrome
You may want to remove "monochrome" one to make colors more plain and less saturated.
===
Illustrious update
I simply added few newer Ditto's arts and trained a Lora on it for Illustrious with slightly different training settings. Check "about this version".
===
Pony release
About 3 weeks ago I decided to make Dittochad style Lora. And soon after that day I made it. And I wasn't satisfied with result. It's been 3 weeks of trials and errors and I am still not satisfied. But I got really tired of those repetitive manipulation and now you can use Lora of the style that already exists on this website.
Honestly, the process gave me several micro depressions. Not really sure If I want to do things that I really barely understand.
The Lora is not bad, I just wanted to make it way better than this. sigh
Have to keep fighting.
Description
Slightly retagged and changed some parameters. Foreskin tag causes less errors now.
This version was trained locally and I overall like it more than previous one.
Sorry for uploading large sized Loras, I'll fix it soon I hope....
Recommended settings: 5.5 CFG, 24 steps, Euler A, Karras