anime-seg 0.2.0

Anime Character Segmentation with DINOv2

AnimeSeg

Anime Character Segmentation using DINOv2 + U-Net++ with LoRA fine-tuning.

sample image

Installation

pip install anime_seg

Usage

from anime_seg import AnimeSegPipeline
pipe = AnimeSegPipeline.from_mask2former().to("cuda")
mask = pipe("path/to/image.jpg")
mask.save("output.png")

AnimeSegPipeline() default constructor is deprecated. Use from_mask2former() or from_dinoV2().

Optional: output size

# Same as input size (default)
mask_same = pipe("path/to/image.jpg")

# Fixed output size
mask_fixed = pipe("path/to/image.jpg", width=1024, height=1024)

# Width/height can be specified independently
mask_w = pipe("path/to/image.jpg", width=1024)
mask_h = pipe("path/to/image.jpg", height=1024)

Advanced Usage

# Load specific file from HF repo
pipe = AnimeSegPipeline.from_mask2former(
    repo_id="suzukimain/AnimeSeg",
    filename="models/anime_seg_mask2former_v3.safetensors"
).to(device="cuda")

# DINOv2 backend
pipe_dino = AnimeSegPipeline.from_dinoV2(
    filename="models/anime_seg_dinov2_v2.safetensors"
).to("cuda")

# Use PIL Image
from PIL import Image
img = Image.open("image.jpg")
mask = pipe(img)

Model Files

Models should follow the naming convention:

models/anime_seg_{architecture}_v{version}.safetensors

Example:

• models/anime_seg_dinov2_v2.safetensors
• models/anime_seg_mask2former_v3.safetensors

Resolution order:

1. models/model_config.json
2. fallback scan by models/anime_seg_{architecture}_v{max_version}.{ext}

Segmentation Classes and Mask Colors

Default from_mask2former() returns 12 classes:

ID	Class Key	RGB	Color
0	background	(0, 0, 0)	Black
1	skin	(255, 220, 180)	Pale Orange
2	face	(100, 150, 255)	Blue
3	hair_main	(255, 0, 0)	Red
4	left_eye	(0, 255, 255)	Cyan
5	right_eye	(255, 255, 0)	Yellow
6	left_eyebrow	(150, 255, 0)	Yellow Green
7	right_eyebrow	(0, 255, 100)	Emerald Green
8	nose	(255, 140, 0)	Dark Orange
9	mouth	(255, 0, 150)	Magenta Pink
10	clothes	(180, 0, 255)	Purple
11	accessory	(128, 128, 0)	Olive

from_dinoV2() returns 13 classes (includes unknown as ID 12).

DINOv2 Compatibility Note

Earlier versions primarily used DINOv2. Current recommendation is from_mask2former(), while from_dinoV2() remains for compatibility.