DiffAM: Differentiable Appearance Modeling

Overview

DiffAM is a framework for learning and generating image appearances in a differentiable manner. It enables embedding extraction, conditional synthesis, and attribute-level editing of medical and natural images.

Repository	https://github.com/SSTDV-Project/DiffAM
License	MIT

Installation

pip install diffam

Quick Start

import diffam, torch
model = diffam.DiffAM(pretrained='medical_imaging')
image = torch.randn(1, 1, 256, 256)
embeddings = model.extract_embeddings(image)
synthetic = model.generate(embeddings, n_samples=5)
edited = model.edit_attributes(embeddings, attribute='contrast', strength=1.5)

Pretrained Models

Model	Best For
`medical_imaging_v1`	MRI, CT, X-ray
`face_v1`	Face images
`general_v1`	Natural images

Dependencies

torch>=2.0
torchvision>=0.15
einops>=0.6

Citation

@software{diffam,
  title   = {DiffAM: Differentiable Appearance Modeling},
  author  = {SSTDV Project},
  year    = {2024},
  url     = {https://github.com/SSTDV-Project/DiffAM}
}

← Back to Software References