see-through/annotators

annotators

Annotator plugins for the inference pipeline. Installed as a Python package via the unified requirements.txt at repo root (-e ./annotators).

Package Structure

Module	Purpose	Extra required
wdv3_tagger	WDv3 image tagging	(base)
gradcam	GradCAM heatmap generation	(base)
lama_inpainter	LaMa inpainting	(base)
bizarre_tagger	Bizarre pose estimation + BG segmentation	[bizarre_tagger]
lang_sam	Language-guided SAM segmentation	[lang_sam]
animeinsseg	Anime instance segmentation (mmdet3)	[animeinsseg]
anime_face_detector	Anime face detection (legacy)	Separate ann_mmpose env

Optional extras

The base annotators install with the unified env. For heavier extras:

# Body parsing + language SAM (~10 min, needs C++ compiler)
pip install -e annotators[bizarre_tagger,lang_sam]

# Instance segmentation (needs CUDA toolkit for mmcv build)
pip install -e annotators[animeinsseg]

# Or use the pre-configured tier file:
pip install -r requirements-inference-mmdet.txt

# Everything
pip install -e annotators[all]

Anime face detection (separate env)

anime_face_detector requires legacy dependencies (PyTorch 1.13, mmcv-full 1.7) incompatible with the unified env. Set up a dedicated conda environment:

conda create -n ann_mmpose python=3.10
conda activate ann_mmpose
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 \
  --extra-index-url https://download.pytorch.org/whl/cu117
pip install openmim==0.3.6 numpy==1.26.4 opencv-python==4.10.0.84
mim install mmcv-full==1.7.0
mim install mmdet==2.28.2
mim install mmpose==0.29.0
pip install -e ./common

Run from the repo root:

conda run --no-capture-output -n ann_mmpose \
  python inference/scripts/parse_live2d.py facedet \
  --exec_list workspace/datasets/.../exec_list.txt