Sign In

Meia (Mobius Final Fantasy) | Character LoRA

76
414
3
Type
LoRA
Stats
414
Reviews
Published
Jul 9, 2023
Base Model
SD 1.5
Training
Steps: 2,780
Epochs: 5
Usage Tips
Clip Skip: 2
Trigger Words
mff meia, blue eyes, white bra, blue jacket, white corset
white panties
Hash
AutoV2
A8D2841CCD

Character LoRA for Meia the Witch from the game Mobius Final Fantasy.

Recommended weight: 0.7 - 1.0 (I use 0.8, but see which works best for you)

Trigger words: mffmeia, blue eyes, white bra, blue jacket, white corset, white panties Add/remove outfit tags based on image composition (e.g., portrait, upper body, cowboy shot, feet out of frame, full body) and whether you want her to wear something else.

Optional tags (work in progress, may either improve outfit details or make it worse): shoulder armor, short sleeves, elbow gloves, white gloves, blue corset (instead of white corset)

Trained on anyloraCheckpoint_novaeFp16

Notes (for those interested in rare character training info)

Even though Mobius was a Final Fantasy game, very few people were aware of it, and now that it was shut down in 2020, hardly nobody still remembers it. Having spent thousands of hours playing it over its lifespan, it still retains a special place in my heart as one of the best mobile games available at the time.

Since it was fairly certain that no one would bother making a LoRA for Meia, an unknown character (that has no pictures to be used for training) from a forgotten game, I decided to attempt training one myself.

The game's obscurity made it difficult to gather any training data; the only other images of the character are her in-game job cards, which are completely unsuitable for training due to their blurriness, clutter, and inconsistent outfits and faces. There are also only a few nude sketches, which were of little use. Because of this, the results were terrible at first, so I used videos on in-game cut-scenes to take screenshots of her, which helped quite a lot. However, the training images were still too limited, blurry and low-res to produce any good results.

To improve the dataset, I used Blender and the data-mined character's in-game 3D model & textures to render multiple images from various angles and in different poses. Although it wasn't the best 3D model (given that it is still a mobile game), the average at best renders added to the dataset were enough to greatly improve the quality of generated images.

The majority of the output images had a very "3D" and pastel(ish?) style, which I couldn't get rid of by using tags like "3d" or "realistic," which was an an issue since I wanted to generate drawn anime/cartoon images. This makes sense given that the entire training set was made up of screenshots of the 3D model used in the game.

Therefore, I used a low weight (0.4-0.6) of the prototype LoRa to generate a few thousands images using multiple checkpoints (for style variety), then handpicked a few dozen which had turned out well, and added them to the training data. I repeated this process a few times (using the output as input for the next version), removing the worst images and adding in the best new ones each time, until there were no more rendered/low quality images in the dataset.

Lastly, I improved training data with Photoshop (by manually removing backgrounds, clutter, fixing hands, lighting, etc. for each image), further pruned the dataset (keeping only the best images), and polished the tags as much as I could.

Since the training data was created using SD, there are a few minor problems with the accuracy of her default outfits, particularly the armoured parts. For those who have played the game and are familiar with the character, it was difficult to get her spiked armoured greaves and silver gauntlet just right because they appeared to conflict with other elements. Therefore I focused on her blue jacket/coat, corset, lace-trimmed white bra/panties and, of course, her face. The results seem to be good enough, but I believe they could be further improved with additional high-quality training data focusing on the details particular elements (especially her jacket/coat and armor), which I may do in the future.

Given that this all began from virtually nothing, I'm quite pleased with the results. It's unlikely that I'll do something similar again anytime soon as this process took a lot longer than a sane person should spend on an unknown character from an obscure game. In any case, I am satisfied for now.