- Python 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| assets | ||
| dvd | ||
| scripts | ||
| .gitignore | ||
| LICENSE | ||
| README.md | ||
| setup.py | ||
Adopting a human developmental visual diet (DVD) diet yields robust, shape-based AI vision
A plug-and-play training curriculum that progressively “ages” each image, mimicking the evolving visual sensitivities of the human visual system, from immature newborn vision through to mature adulthood.
1 Why Developmental Visual Diet (DVD)?
- Human vision begins with severely immature inputs, whereas current ANNs mainly train on instantaneous, high-fidelity inputs.
- We model developmental trajectories of visual acuity, contrast sensitivity, and chromatic sensitivity from birth to 25 years.
- A differentiable pipeline dynamically applies these sensory maturation on-the-fly during training.
- Guiding AI vision through this Developmental Visual Diet produces models that better align with hallmark features of human visual robustness:
- Near-human-level shape bias
- Enhanced robustness to corruptions
- Improved resilience to adversarial attacks
- Abstract shape recognition beyond state-of-the-art vision-language models (e.g., ChatGPT-4o, Gemini 2.0 Flash, LLaMA-4-Scout)
📈 Developmental visual trajectories
Note
For new projects, we recommend starting with the scale-free version in the
scale_freebranch.
It reformulates the DVD transformations, making behaviour more consistent across resolutions and different conditions.
The current branch remains useful for reproducing the original setup and results reported in the paper.
2 Installation
Typical install time on a standard desktop computer: ~1–3 minutes.
git clone https://github.com/KietzmannLab/DVD.git
cd DVD
pip install -e .
python - <<'PY'
import dvd, torch
print("DVD version:", dvd.__version__, "| CUDA =", torch.cuda.is_available())
PY
3 Quick demo - aging visual experience
Typical running time: ~1 minute on a single H100 GPU.
from pathlib import Path
from typing import List
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import torch
from dvd.dvd.development import DVDTransformer, DVDConfig # DVD Data Transformer (main API)
# ------------------------------------------------------------
# Configuration
# ------------------------------------------------------------
AGES: List[int] = [1, 4, 16, 64, 256] # ages in months
IMG_SIZE: int = 256 # resize target (px)
CFG = DVDConfig()
# Note: If your input images are not normalized to [0, 1], consider set 'by_percentile=True' in DVDConfig() to percentile-based thresholding, which adapts to the image’s actual intensity distribution.
# Input / output paths
ASSETS_DIR = Path("assets/example_stimuli")
IMAGE_PATHS = [
ASSETS_DIR / "example_1.jpeg",
ASSETS_DIR / "example_2.jpeg",
]
OUT_DIR = Path("results/dvd_demo_output")
OUT_PATH = OUT_DIR / "dvd_demo_output.pdf"
# Helper: load an image as [1, 3, H, W] float tensor in [0, 1]
def load_tensor(fp: Path) -> torch.Tensor:
img = Image.open(fp).convert("RGB")
img.thumbnail((IMG_SIZE, IMG_SIZE), Image.LANCZOS)
arr = np.asarray(img).transpose(2, 0, 1) / 255.0
return torch.from_numpy(arr).unsqueeze(0).float()
# Main: build demo
def make_demo(paths: List[Path], outfile: Path) -> None:
dvdt = DVDTransformer(CFG)
tensors = [load_tensor(p) for p in paths]
rows, cols = len(tensors), len(AGES)
fig, ax = plt.subplots(rows, cols, figsize=(3 * cols, 3 * rows))
for r, img_t in enumerate(tensors):
for c, age in enumerate(AGES):
out = dvdt(img_t.clone(), months=age) # DVD data transformation
vis = out.squeeze(0).permute(1, 2, 0).numpy().clip(0, 1)
ax[r, c].imshow(vis)
ax[r, c].axis("off")
if r == 0:
ax[r, c].set_title(f"{age} mo", fontsize=12)
fig.tight_layout()
outfile.parent.mkdir(parents=True, exist_ok=True)
fig.savefig(outfile, dpi=300)
print(f"Saved {outfile.resolve()}")
make_demo(IMAGE_PATHS, OUT_PATH)
4 Datasets
This project makes use of several datasets:
| Dataset | Description | Link |
|---|---|---|
| Ecoset | A natural image dataset introduced in Mehrer et al., 2021 | Ecoset Website |
| ImageNet | Our models were also trained on the initial release of ImageNet | ImageNet Website |
5 Trained checkpoints
We provide several trained checkpoints for direct use.
| Training data | Model variants | Notes | Link |
|---|---|---|---|
| Ecoset | Baseline, DVD-P, DVD-B, DVD-S | Default models trained on Ecoset. We recommend DVD-B as the default starting point. | OSF checkpoint files |
| Ecoset | Scale-free DVD-B | This version shows similar shape bias and higher accuracy than the DVD-B Paper Version, see the code in the scale_free branch |
OSF scale-free checkpoint |
| ImageNet-1K | Baseline, DVD-P, DVD-B, DVD-S | Models trained on ImageNet-1K (without hyperparameter sweeping) | OSF ImageNet-1K checkpoints |
Important
All checkpoints released here were trained with the same data augmentation pipeline, making comparisons within this repository straightforward.
For other publicly available models, training-time data augmentations may differ substantially, so results are not necessarily directly comparable.
6 Training with DVD | Example
python scripts/main.py /share/klab/datasets --arch resnet50 --epochs 150 --dataset-name ecoset_square256 --class-weights-json-path '/share/klab/datasets/optimized_datasets/lookup_ecoset_json.json' --batch-size-per-gpu 512 --image-size 256 --warmup-epochs 0 --development_strategy dvd --months_per_epoch 2 --contrast_amplitude_beta 1e-4 --contrast_amplitude_lambda 150 --seed 1
| Flag | Purpose |
|---|---|
--development_strategy |
dvd (full curriculum) or adult (control). |
--months_per_epoch |
Virtual ageing per epoch (smaller = finer). Example: DVD-P = 4, DVD-B = 2, DVD-S = 1 |
--time_order |
chronological or fully_random. |
--apply_* |
Toggle acuity / colour / contrast sensitivities. |
--contrast_amplitude_* |
Control the reference amplitude threshold in FFT. |
7 Core API
from dvd.dvd.development import DVDTransformer, DVDConfig, generate_age_months_curve
# Initialize transformer and generate age mapping curve
dvdt = DVDTransformer(DVDConfig())
age_curve = generate_age_months_curve(
epochs=args.epochs,
steps_per_epoch=len(train_loader),
months_per_epoch=args.months_per_epoch,
)
# Map current batch index to virtual age in months
step_idx = (epoch * len(train_loader)) + i
age_months = age_curve[step_idx]
# Apply age-based visual transformations
images_aged = dvdt(img_t.clone(), months=age, curriculum=age_curve)
8 Citation
@article{lu2025dvd,
title = {Adopting a human developmental visual diet yields robust, shape-based AI vision},
author = {Zejin Lu, Sushrut Thorat, Radoslaw M. Cichy, Tim C. Kietzmann},
journal = {arXiv preprint arXiv:2507.03168},
year = {2025},
doi = {10.48550/arXiv.2507.03168},
url = {https://arxiv.org/abs/2507.03168}
}

