These versions were trained on Moss Image Base. Example images were generated with Moss Image SFT.

These loras were trained on images from Megami Magazine from late 2018 to 2026. Moss already knows the Megami Magazine tag but it’s riddled with japanese text. These loras were trained with the text manually removed and the data curated to present a specific style or look.

To get the results you see in the images above, these are the settings I used.

Visit My Site For Models I Can't Post On CivitAI: https://qtgurl.com/

The loras for Moss V2 was trained at 2048 resolution while Moss V1 was trained at 1792 resolution. So they’re capable of generating at around 2048×2048 to 1792×1792 or lower. (I use 1600×1984)

MeMaMoss – V1 + V2

Trained on 68 new and old images. Dataset was re-collected and edited for the best possible quality.

About MeMaMoss V2

I was unsatisfied with the results of Version 1. Mainly because the 2d style continued to break as I was testing new prompts. In my frustration, I decided to retrain it. This time at a higher rank of 128 and at a resolution of 2048. With this new version, there was a noticeable quality increase but I was still running into the same problem of style loss.

I eventually discovered the culprit was the quality tags, particularly Masterpiece. After further trial and error, I arrived to a solid combination of positive and negative tags that give me a pretty consistent lora style.

Advice for Prompting Moss

When using this lora, only this in the positive:

high quality

And these in the negative:

best quality, normal quality, low quality, worst quality, photorealistic, realistic, 3d, film grain, bad anatomy, watermark, jpeg artifact, bloom, blurry, blurry background, depth of field, extra legs, too many fingers, too many toes, thick lineart, painterly

The tokens in Moss SFT can be very strong and if you use certain keywords like Masterpiece, it will over power the lora.

I hope you enjoy using this lora as much as I am. 2d style for life!