[CVPR 2026] Illustrator’s Depth: Monocular Layer Index Prediction for Image Decomposition

Nissim Maruani^1,2, Peiying Zhang³, Siddhartha Chaudhuri⁴, Matthew Fisher⁴, Nanxuan Zhao⁴, Vladimir G. Kim⁴, Pierre Alliez^1,2, Mathieu Desbrun^1,5. Wang Yifan⁴

¹Inria, ²Université Côte d'Azur, ³City University of Hong Kong, ⁴Adobe Research, ⁵École polytechnique

Paper & more available here.

News

• 05/26 📟 Code & model released
• 03/26 🔥 Our paper was accepted to CVPR 2026

Abstract

We introduce Illustrator’s Depth, a novel definition of depth that addresses a key challenge in digital content creation: decomposing flat images into editable, ordered layers. Inspired by an artist’s compositional process, illustrator’s depth infers a layer index for each pixel, forming an interpretable image decomposition through a discrete, globally consistent ordering of elements optimized for editability. We also propose and train a neural network using a curated dataset of layered vector graphics to predict layering directly from raster inputs. Our layer index inference unlocks a range of powerful downstream applications. In particular, it significantly outperforms state-of-the-art baselines for image vectorization while also enabling high-fidelity text-to-vector-graphics generation, automatic 3D relief generation from 2D images, and intuitive depth-aware editing. By reframing depth from a physical quantity to a creative abstraction, illustrator's depth prediction offers a new foundation for editable image decomposition.

Installation

Below are the setup instructions for Linux and macOS environments. Start by cloning this repo:

git clone --recursive https://github.com/nissmar/illustrators_depth.git

Linux

Create and activate the Conda environment, then install Depth Pro (tested with CUDA 12.4):

conda create --name argos python=3.11 ipython -y
conda activate argos
pip install -r requirements.txt 
cd ml-depth-pro
pip install -e .
cd ..

MacOS

Create and activate the Conda environment, then install Depth Pro (tested on Sequoia 15.7.2):

conda create --name argos python=3.11 ipython -y
conda activate argos
cd ml-depth-pro
pip install -e .
cd ..
pip install -r requirements_macos.txt
conda install -c conda-forge cairo

Model Download

The model's weights are accessible here. Please download them using:

mkdir checkpoints
cd checkpoints
gdown 12BRCnHIQutSThfJhhHn_sYOBxFeq-NB3
unzip mmsvg_model.zip
rm mmsvg_model.zip 
cd ..

🚀 Quickstart

For an easy quickstart, use our provided gradio demo:

python gradio_demo.py

Illustrator's Depth Predictions

Running from command line

We provide a script that computes Illustrator’s depth predictions for all images in a folder SOME/PATH. It will generate a new directory, SOME/PATH_predictions, containing the model outputs. On macOS, be sure to set --device 'mps'. Run:

python -m predict --model_path checkpoints/mmsvg_model/checkpoints/id_model.ckpt --device 'cuda' --src SOME/PATH 

Running from Python

from src.model.illustrators_depth_model import load_illustrators_depth_model

model_path = 'checkpoints/mmsvg_model/checkpoints/id_model.ckpt'
device = 'cuda'

img_path = 'data/examples/a.png'
preserve_size = True
gaussian_sigma = 0 # Blurring the input might help for images with textures (canvas, paper...)

model, cfg = load_illustrators_depth_model(model_path, device=device)
model.eval()

np_img, pred_depth = model.infer_single_image(
            path, preserve_size=preserve_size, gaussian_sigma=gaussian_sigma)

np_img # The input image   
pred_depth # The predicted illustrator's depth

Vectorization

Running from command line

python predict_and_vectorize.py --device 'cuda' --src SOME/PATH 

Running interactively

See interactive_vectorization.ipynb

Rasterize SVG images: RGB and Depth

We provide a script to rasterize SVG images, along with their ground-truth illustrator's depth:

python rasterize_svg_with_depth.py --src SOME/PATH 

Evaluation

Our paper includes two evaluation tracks:

• Depth Prediction Evaluation: We assess raw depth predictions after normalizing them using median and mean deviation.
• SVG Reconstruction Evaluation: We evaluate both the reconstructed SVGs and their associated depth. Although depth predictions guide the layering, we ultimately rasterize the reconstructed SVGs’ depth maps before computing metrics.

Depth Prediction Evaluation (Table 1 in the paper)

Step 1: Rasterize Ground Truth SVGs

python rasterize_svg_with_depth.py --src data/MMSVG100/

Step 2: Run illustrator's depth prediction on raster images

python -m predict --model_path checkpoints/mmsvg_model/checkpoints/id_model.ckpt --device 'cuda' --src data/MMSVG100

Step 3: Compute depth metrics

python evaluate.py --gt_src data/MMSVG100 --pred_src data/MMSVG100_predictions/

SVG reconstruction

Step 1: Rasterize Ground Truth SVGs

python rasterize_svg_with_depth.py --src data/MMSVG100

Step 2: Vectorize input images

python predict_and_vectorize.py --model_path checkpoints/mmsvg_model/checkpoints/id_model.ckpt --device 'cuda' --src data/MMSVG100

Step 3: Rasterize reconstructed SVGs

python rasterize_svg_with_depth.py --src data/MMSVG100_vectorization

Step 4: Run SVG evaluation

python evaluate.py --gt_src data/MMSVG100 --pred_src data/MMSVG100_vectorization --eval_SVG True

Training the model

The model is trained using 8 A100 GPUs. To launch training, run:

python -m src.train.train --cfg src/configs/train.yml

Citation

If you find this project useful, please consider citing:

@article{maruani2025illustrator,
  title={Illustrator's Depth: Monocular Layer Index Prediction for Image Decomposition},
  author={Maruani, Nissim and Zhang, Peiying and Chaudhuri, Siddhartha and Fisher, Matthew and Zhao, Nanxuan and Kim, Vladimir G and Alliez, Pierre and Desbrun, Mathieu and Yifan, Wang},
  journal={arXiv preprint arXiv:2511.17454},
  year={2025}
}