Fiz-Rot - LyCORIS LoCon - Illustrious

LoCon trained on Fiz-Rot's artstyle for Illustrious models

Why two versions?

When first training on Illustrious I had several problems with cross-model compatibility. Turns out a setting I had ticked (DoRA) was causing these problems, but before I realized this I would train versions for each individual model. This resulted in two versions: One on base illustrious and one on NoobAI. Even after fixing the problem I'd figured I would train for both and see how it turns out and so I'm posting both versions. Going forward I'll probably only train on base Illustrious since it seems to be best in terms of accuracy and merge compatibility. This also means the old version is a DoRA but I can't change it in the settings without splitting it into a new model page so ¯\_(ツ)_/¯

Which version should I use?

• The base illustrious version seems to adhere strongest to the style across merges without affecting the performance of the merge too strongly.

• The NoobAI version is the version I tested the most, but is mostly a result of an error on my part. Both versions have their merit but I would probably recommend the base illustrious version due to wider compatibility.

General usage Notes:

For more specific issues check the "about this version" section of the specific model.

Thus far my most successful permanent positive/negative combo is:

masterpiece, best quality, good quality  

lowres, worst quality, low quality, oldest, jpeg artifacts, blurred, artist name, signature

This likely changes depending on settings and merges, but I've not needed to modify them.

Strength depends on merge lora compatibility and what you're trying to do, generally:

• If you're using it alone on a highly compatible model or base then I recommend .8-1

• If using with artist tags I recommend .4-.8, but this is to your discretion

• If using on stricter merges I would recommend not going over .7-.8

General Workflow for Sample Images (on ComfyUI):

• Sampler from images is done with Restarts with segments of [3,2,0.06,0.30], which are values from the paper on Restarts

  • Note: Using restarts tends to smooth out the image more but makes it more coherent, your mileage may vary and it's up to you if you want to use them

• No special scheduler is used

• Conditioning is split (A1111 BREAK) into 3 parts, generally BG and camera/Subject Face and attributes/Subject body and clothing

• Initial resolution is selected from the SDXL whitepaper appendix I: https://arxiv.org/pdf/2307.01952

• Image is upscaled by 1.5x using Lanczos resampling

• Second sample pass is the same, but at around 15 steps at .3 denoise

• A facedetailer is used after the upscale