These loras were trained on images from Megami Magazine from late 2018 to 2026. Anima already knows the Megami Magazine tag but it's riddled with japanese text. These loras were trained with the text manually removed. Please read the About This Version for more information.
To get the results you see in the images above, these are the settings I used.
The lora for Anima V1-2 was trained at 1792 resolution. So it's capable of generating around 1792x1792 or higher.(I use 1600x1984) P2 and P3 were trained at 1536 resolution, so they're capable of generating around 1536x1536. (I use 1408x1792 for most gallery images.) It can also generate small sizes just fine.
MeMaAni - V2
Trained on 68 (mostly) new and old images. Dataset was re-collected and edited for the best possible quality.
MeMaAni - V1
Trained on the same data used in Preview 3.
MeMaAni - Anima Preview 3
Trained on the same data used in Preview 2 with slight changes to improve results. Most of the improvements come from the Anima P3.
MeMaAni - Anima Preview 2
Trained on the same data used in 2d Gold Fish. I guess you could call it MeMa F, but I suppose that would make it A for this model. It uses a lot of images from the A dataset so it's close enough.
Comment On Anima
The final release of Anima 1.0 is fantastic. The model has so much capability. If you haven't made the switch, now is the time. It doesn't matter if all you post is one girl, the jump in quality, prompt comprehension, 16 channel VAE, newer anime knowledge and just raw stable performance makes it the worth while model to switch to as an anime genner.
Description
Version 2 brings newness at every stage of the training process.
Unlike all of the previous versions of MeMa, I went out and grabbed entirely new images to train on. Many of the newer images are from 2026 and back.
All of the images from the older dataset were.....well....okay-ish, but I wanted only the best for Anima training, so I re-edited all of the new images to provide the most cleans Megami Magzine style to date. I also changed the settings I used to train this lora. This is the first time I've departed so drastically from my traditional training methods. These actions were all in the name of providing you the most optimum results.
This lora was trained on 68 images at 1792 resolution. The dataset is comprised of mostly new images and a few familiar images I used in the classic A dataset. My overall goal was to improve the backgrounds and enhance finer details to properly reflect the style. I feel very confidence I achieved that, but I still believe there is room for improvement.
Due to the mixed nature of the images, eye styles will feel random at times. Try placing "colored pupils" or "ringed eyes" in the negative. In future versions, I will collector more images that are similar to stabilize styles.