These loras were trained on images from Megami Magazine from late 2018 to early 2024. Megami Magazine has different styles of 2d images and I chose to separate them by types. (Hence Type A, B, C, etc.) Please read the About This Version for more information. Please use the lora based on the model for the optimum results.
To get the results you see in the images above, these are the settings I used.
Short List
Type A:
Links: v5.0 A [WaiShuffle-Vpred] | v5.0 A [Noob1-Vpred] | v4.0 A [Zootlustrious-v2] | v3.0 A [Hadrian DeliceXL] | v3.0 Type A [AutismMix] | v3.0 A [Ponymagine] | v3.0 A [Illustrious-v1]
Type B:
Links: v5.0 B [Noob1-Vpred] | v4.0 B [Zootlustrious-v2] | v3.0 Type B [AutismMix] | v3.0 Type B [Hadrian DXL]
Type C:
Links: v5.0 C [Noob1-Vpred] | v3.0 C [Illustrious-v1]
Type D:
Links: v5.0 D [Noob1-Vpred] | v4.0 D [Zootlustrious-v2]
Old:
Links: v2.0 [AutismMix] | v1.0 [Pony]
Long List
V5 Type D: Trained using 53 images selected for their similar eyes. The base for this lora is Noob 1.0 V-Pred.
Links: v5.0 D [Noob1-Vpred]
V5 Type C: Trained using 29 images selected for their similar eyes. The base for this lora is Noob 1.0 V-Pred.
Links: v5.0 C [Noob1-Vpred]
V5 Type B: Trained using 52 images selected for their similar eyes and hair. The base for this lora is Noob 1.0 V-Pred.
Links: v5.0 B [Noob1-Vpred]
V5 Type A: Trained using 43-50 images selected for their similar eye style and detailed backgrounds. The base for this lora is Noob 1.0 V-Pred and WAI-SHUFFLE-NOOB Vpred 2.0.
Links: v5.0 A [Noob1-Vpred] | v5.0 A [WaiShuffle-Vpred]
V4 Type D: Trained using 53 images selected for their similar eye style. Trained on Zootlustrious v2.
Links: v4.0 D [Zootlustrious-v2]
V4 Type B: Trained using 52 images selected for their similar eyes and hair Trained on Zootlustrious v2.
Links: v4.0 B [Zootlustrious-v2]
V4 Type A: Trained using 50 images selected for their similar eye style and detailed backgrounds. Trained on Zootlustrious v2.
Links: v4.0 A [Zootlustrious-v2]
V3 Type C: Trained on 58 images based on similar eye types. Compared to Type A and B, the eyes are more round and more slender. The base for this lora is Illustrious-XL v0.1.
Links: v3.0 C [Illustrious-v1]
V3 Type B: Trained using 52 images selected for their similar eyes and hair. Trained on Hadrian DeliceXL 2.0, and AutismMix.
Links: v3.0 Type B [AutismMix] | v3.0 Type B [Hadrian DXL]
V3 Type A: Trained using 50 images selected for their similar eye style and detailed backgrounds. Trained on AutismMix, Hadrian DeliceXL 2.0, Ponymagine , Illustrious-XL v0.1.
Links: v3.0 A [Hadrian DeliceXL] | v3.0 Type A [AutismMix] | v3.0 A [Ponymagine] | v3.0 A [Illustrious-v1]
V2: This lora was trained using 150+ images that I painstakingly removed the japanese text from. Trained on AutismMix.
Links: v2.0 [AutismMix]
V1: Trained using 1,200+ images. (All I had.) Trained on Based Pony.
Links: v1.0 [Pony]
Description
This lora version was trained on 150+ images from Megami Magazine mostly from 2023 to early 2024. All of the images from the dataset have descriptive text in the corner, so I painstakingly went through and removed all the text from the corners of every image used.
My goal for the lora was to exceed the previous version completely. Well, it's more of a diagonal upgrade than a direct one. This lora was trained on Autism Pony instead of Base Pony and I tried using slightly different settings to accelerate the training. The results have been mostly positive, but noticeably different than the previous version.
The Good
-Better understanding of the prompt thanks to AutismMix
-No more annoying japanese gibberish in the corners
-The style is more consistent (A mix of the lora and the influence from AutismMix)
-Much improved backgrounds (I made sure to include detailed background images and it paid off)
The Bad
-Hands come out wrong more often than I like (this is a result of the dataset used, I tried to avoid this from happening but you can't win em all)
-Faces tend to lack detail the further away they are (this effect hires results)
-Due to the limited data I used, the style is noticeably different from the previous version, this might not be desirable for some
Over all, I am happy with the results. Some people might be wondering if I will go back and correct the remaining 1,000+ images to make a true version upgrade. The answer to that question is, no. No I will not.
Enjoy!