Type | |
Stats | 651 27,047 |
Reviews | (70) |
Published | Nov 24, 2023 |
Base Model | |
Training | Steps: 3,636 Epochs: 25 |
Usage Tips | Clip Skip: 2 Strength: 0.8 |
Trigger Words | j4bc0m1x_jab style j4bc0m1x_laz style |
Hash | AutoV2 7212DCC4A2 |
Important things:
strongly recommend using speech bubble, simple background in negative prompt with at least 1.0 strength (I've been using 1.5)
You'll get the best results with setting minimum resolution to at least 768x768
Some of my showcase images use ADetailer for faces, but most don't
For showcase images I used 4x-AnimeSharp upscaler for Hires. Fix set to 0.4 denoising strength
This started with trying to answer a bounty but I found it more interesting so continued to work on it - this is a LoRA to try and reproduce the style of JAB Comix. I used this mostly as a learning experience to develop and get more familiar with making LoRA's. It's far from perfect and has lots of it's own issues but I was able to at least make some images that I felt were true to the style with it.
This was trained against the Aniverse 1.5 checkpoint model, so it seems to work best with that but it works to varying degrees with other models in my limited testing.
Version 1.2 Experimental
I tried training in a separate concept to reproduce Laz's art style with a new trigger word. I'm not particularly satisfied with it but I thought I'd just post it as someone may like it better. A few of the captions cause the styles to bleed into each other which modifies output which may not be desirable - you may prefer version 1.1 because of this.
Version 1.1
I originally made a major error in my assumptions about the comics and have corrected that for this version. The trigger word is changed to j4bc0m1x_jab and it should accurately and reliably reproduce his specific style now. Recommended strength is now 0.8.
Version 1.0
This is a mixing of multiple artists due to an error on my part. I'm leaving it available in case someone wants to play with it as it can still make images in that general style but it's not accurate to a specific artist like Jab. It's technically a mixing of Jab and Laz's art styles, you can control it to a degree by invoking other trigger words (based on the comics trained against), they're found in the trigger words info.