AnimeSeg

Anime Character Segmentation using Mask2Former and DINOv2 + U-Net++ with LoRA fine-tuning. Also integrates Background Removal via anime-segmentation.

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")

# Background Removal (powered by anime-segmentation)
bg_pipe = AnimeSegPipeline.from_bg_remover().to("cuda")
no_bg_img = bg_pipe("path/to/image.jpg")
no_bg_img.save("no_bg_output.png")

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

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)

# Background Removal (powered by anime-segmentation)
bg_pipe = AnimeSegPipeline.from_bg_remover().to("cuda")
no_bg_img = bg_pipe("path/to/image.jpg")
no_bg_img.save("no_bg_output.png")

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. 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:

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.