README.md

metadata

license: creativeml-openrail-m
tags:
  - text-to-image
  - stable-diffusion
  - anime
  - aiart

This model is trained on 33 different concepts from Bofuri: I Don't Want to Get Hurt, so I'll Max Out My Defense (防振り: 痛いのは嫌なので防御力に極振りしたいと思います。)

Example Generations

Prompt: BoMaple uniform BoSally unfirom, yuri, in classroom, 4K wallpaper, beautiful eyes

Prompt: 2girls, BoMay BoYui, yuri, half body, floating in the sky, cloud, sparkling eyes, 4K wallpaer, anime coloring, official art

Prompt: BoKanade casting magic, 4K wallpaper, outdoors

(Negative is mostly variations of: bad hands, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry)

Usage

The model is shared in both diffuser and safetensors format. Intermediate checkpoints are also shared in ckpt format in the directory checkpoints.

Concepts

The 33 concepts are listed in concept_list and demonstrated below.

BoMaple + BoSally + The following use the full name of the concepts

Expect bad results for BoMaple sheep form and non-human concepts. Espeically the model clearly does not understand the anatomy and the size of syrup (well it is true it's size is not fixed). For BoKasumi sarashi adding bandages seems to help. For BoMaple pajama you can add stripe for more similarity to the pajama appearing in anime. The remaining concepts should go through smoothly.

Prompt format

During training the concept names are put at the beginning of the images separated only by spaces, but not doing so seems to work as well. Put aniscreen after the concept names would reinfoce the anime style. Having two concepts in a same image is fairly doable as demonstrated above. However, expect weird blending to happen most of the time starting from three concepts. This is partially because this model is not trained too much on multi-concept scenes. Below is roughly the best we can get after multiply tries (there is still clothes blending).

Prompt: (BoMaple black armor) BoSally turtleneck BoKasumi, 3girls, 4K wallpaper, ahoge, black hair, brown hair, outdoors, long hair

More Generations

Prompt: BoMaple black armors aniscreen, 1girl solo, Hydra in the sky, light purple eyes, 4K wallpaper

Prompt: BoMaple black armors near small turtle syrup, sitting with knees up on rock looking at viewer, turtle shell, beautiful hand in glove, in front of trees , outdoors, close-up, 4K wallpaper

Prompt: BoMaple pajama stripe, sitting on bed with barefoot, in girl's room, detailed and fancy background, sparkling purple eyes, hand on bed, 4K wallpaper

Prompt: BoFrederica, cowboy shot, in rubble ruins, ((under blue sky)), cinematic angle, dynamic pose, oblique angle, 4K wallpaer, anime coloring, official art

Prompt: Turtle Syrup Fox Oboro next to each other simple background white background, animals

Failures are of course unavoidable

Finally, you can always get different styles via model merging

Dataset Description

The dataset is prepared via the workflow detailed here: https://github.com/cyber-meow/anime_screenshot_pipeline

It contains 27031 images with the following composition

• 7752 bofuri images mainly composed of screenshots from the first season and of the first three episods of the second season
• 19279 regularization images which intend to be as various as possible while being in anime style (i.e. no photorealistic image is used)

Note that the model is trained with a specific weighting scheme to balance between different concepts so that every image does not weight equally. After applying the per-image repeat we get around 200K images per epoch.

Training

Training is done with EveryDream2 trainer with ACertainty as base model. I use the following configuration thanks to the suggestion of 金Goldkoron

• resolution 512
• cosine learning rate scheduler, lr 2.5e-6
• batch size 4
• conditional dropout 0.05
• change beta scheduler from scaler_linear to linear in config.json of the scheduler of the model

The released model is trained for 57751 steps, but among the provided checkpoints all the three starting from 34172 steps seem to work reasonably well.