See Through Anime Layer Decomposition

See-through: Anime Character Layer Decomposition

Skill by ara.so — Daily 2026 Skills collection.

See-through is a research framework (SIGGRAPH 2026, conditionally accepted) that decomposes a single anime illustration into up to 23 fully inpainted, semantically distinct layers with inferred drawing orders — exporting a layered PSD file suitable for 2.5D animation workflows.

What It Does

• Decomposes a single anime image into semantic layers (hair, face, eyes, clothing, accessories, etc.)
• Inpaints occluded regions so each layer is complete
• Infers pseudo-depth ordering using a fine-tuned Marigold model
• Exports layered .psd files with depth maps and segmentation masks
• Supports depth-based and left-right stratification for further refinement

Installation

# 1. Create and activate environment
conda create -n see_through python=3.12 -y
conda activate see_through

# 2. Install PyTorch with CUDA 12.8
pip install torch==2.8.0+cu128 torchvision==0.23.0+cu128 torchaudio==2.8.0+cu128 \
  --index-url https://download.pytorch.org/whl/cu128

# 3. Install core dependencies
pip install -r requirements.txt

# 4. Create assets symlink
ln -sf common/assets assets

Optional Annotator Tiers

Install only what you need:

# Body parsing (detectron2 — for body attribute tagging)
pip install --no-build-isolation -r requirements-inference-annotators.txt

# SAM2 (language-guided segmentation)
pip install --no-build-isolation -r requirements-inference-sam2.txt

# Instance segmentation (mmcv/mmdet — recommended for UI)
pip install -r requirements-inference-mmdet.txt

Always run all scripts from the repository root as the working directory.

Models

Models are hosted on HuggingFace and downloaded automatically on first use:

Key CLI Commands

Main Pipeline: Layer Decomposition to PSD

# Single image → layered PSD
python inference/scripts/inference_psd.py \
  --srcp assets/test_image.png \
  --save_to_psd

# Entire directory of images
python inference/scripts/inference_psd.py \
  --srcp path/to/image_folder/ \
  --save_to_psd

Output is saved to workspace/layerdiff_output/ by default. Each run produces:

• A layered .psd file with semantically separated layers
• Intermediate depth maps
• Segmentation masks

Heuristic Post-Processing

After the main pipeline, further split layers using heuristic_partseg.py:

# Depth-based stratification (e.g., separate near/far handwear)
python inference/scripts/heuristic_partseg.py seg_wdepth \
  --srcp workspace/test_samples_output/PV_0047_A0020.psd \
  --target_tags handwear

# Left-right stratification
python inference/scripts/heuristic_partseg.py seg_wlr \
  --srcp workspace/test_samples_output/PV_0047_A0020_wdepth.psd \
  --target_tags handwear-1

Synthetic Training Data Generation

python inference/scripts/syn_data.py

Python API Usage

Running the Full Pipeline Programmatically

import subprocess
import os

def decompose_anime_image(image_path: str, output_dir: str = "workspace/layerdiff_output") -> str:
    """
    Run See-through layer decomposition on a single anime image.
    Returns path to the output PSD file.
    """
    result = subprocess.run(
        [
            "python", "inference/scripts/inference_psd.py",
            "--srcp", image_path,
            "--save_to_psd",
        ],
        capture_output=True,
        text=True,
        cwd=os.getcwd()  # Must run from repo root
    )
    if result.returncode != 0:
        raise RuntimeError(f"Decomposition failed:\n{result.stderr}")
    
    # Derive expected output filename
    base_name = os.path.splitext(os.path.basename(image_path))[0]
    psd_path = os.path.join(output_dir, f"{base_name}.psd")
    return psd_path


# Example usage
psd_output = decompose_anime_image("assets/test_image.png")
print(f"PSD saved to: {psd_output}")

Batch Processing a Directory

import subprocess
from pathlib import Path

def batch_decompose(input_dir: str, output_dir: str = "workspace/layerdiff_output"):
    """Process all images in a directory."""
    result = subprocess.run(
        [
            "python", "inference/scripts/inference_psd.py",
            "--srcp", input_dir,
            "--save_to_psd",
        ],
        capture_output=True,
        text=True,
    )
    if result.returncode != 0:
        raise RuntimeError(f"Batch processing failed:\n{result.stderr}")
    
    output_psds = list(Path(output_dir).glob("*.psd"))
    print(f"Generated {len(output_psds)} PSD files in {output_dir}")
    return output_psds

# Example
psds = batch_decompose("path/to/my_anime_images/")

Post-Processing: Depth and LR Splits

import subprocess

def split_by_depth(psd_path: str, target_tags: list[str]) -> str:
    """Apply depth-based layer stratification to a PSD."""
    tags_str = " ".join(target_tags)
    result = subprocess.run(
        [
            "python", "inference/scripts/heuristic_partseg.py",
            "seg_wdepth",
            "--srcp", psd_path,
            "--target_tags", *target_tags,
        ],
        capture_output=True, text=True,
    )
    if result.returncode != 0:
        raise RuntimeError(result.stderr)
    # Output naming convention: original name + _wdepth suffix
    base = psd_path.replace(".psd", "_wdepth.psd")
    return base


def split_by_lr(psd_path: str, target_tags: list[str]) -> str:
    """Apply left-right layer stratification to a PSD."""
    result = subprocess.run(
        [
            "python", "inference/scripts/heuristic_partseg.py",
            "seg_wlr",
            "--srcp", psd_path,
            "--target_tags", *target_tags,
        ],
        capture_output=True, text=True,
    )
    if result.returncode != 0:
        raise RuntimeError(result.stderr)
    return psd_path.replace(".psd", "_wlr.psd")


# Full post-processing pipeline example
psd = "workspace/test_samples_output/PV_0047_A0020.psd"
depth_psd = split_by_depth(psd, ["handwear"])
lr_psd = split_by_lr(depth_psd, ["handwear-1"])
print(f"Final PSD with depth+LR splits: {lr_psd}")

Loading and Inspecting PSD Output

from psd_tools import PSDImage  # pip install psd-tools

def inspect_psd_layers(psd_path: str):
    """List all layers in a See-through output PSD."""
    psd = PSDImage.open(psd_path)
    print(f"Canvas size: {psd.width}x{psd.height}")
    print(f"Total layers: {len(list(psd.descendants()))}")
    print("\nLayer structure:")
    for layer in psd:
        print(f"  [{layer.kind}] '{layer.name}' — "
              f"bbox: {layer.bbox}, visible: {layer.is_visible()}")
    return psd

psd = inspect_psd_layers("workspace/layerdiff_output/my_character.psd")

Interactive Body Part Segmentation (Notebook)

Open and run the provided demo notebook:

jupyter notebook inference/demo/bodypartseg_sam.ipynb

This demonstrates interactive 19-part body segmentation with visualization using the SAM body parsing model.

Dataset Preparation for Training

See-through uses Live2D model files as training data. Setup requires a separate repo:

# 1. Clone the CubismPartExtr utility
git clone https://github.com/shitagaki-lab/CubismPartExtr
# Follow its README to download sample model files and prepare workspace/

# 2. Run data parsing scripts per README_datapipeline.md
# (scripts are in inference/scripts/ — check docstrings for details)

Launching the UI

# Requires workspace/datasets/ at repo root (contains sample data)
# Recommended: install mmdet tier first