metadata
license: mit
tags:
- medical
- image-generation
- diffusion
- flow-matching
pipeline_tag: unconditional-image-generation
SiT FAF Generation & Editing
This repository contains the trained model weights for the SiT FAF Generation and Editing project. See the Github repository here
File Structure
sit/
sit_checkpoint_0110000.pt: Main SiT model checkpoint (1.1M steps).- Architecture: SiT-XL/2
- Resolution: 256x256
- Dataset: FAF Retinal Scans
stylegan/
stylegan_checkpoint.pkl: Pre-trained StyleGAN2-ADA model.- Used for baseline comparisons and FID calculation.
judges/
10k_age_judge.pt: Regression model used to evaluate age preservation.laterality_judge.pt: Classification model used to evaluate eye laterality (Left/Right) preservation.
Usage
These weights are designed to be used with the scripts in the GitHub repository.
# Example: Generate Synthetic Dataset of 10k Samples
python scripts/prepare_datasets/generate_synthetic_dataset.py \
--ckpt {path to sit/sit_checkpoint_0110000.pt} \
--data-path data/images_256_cleaned/metadata_cleaned_256.csv \
--mapping-file data/class_mapping.json \
--output-dir data/synthetic_10kSamples \
--num-samples 10000 \
--batch-size 32 \
--cfg-scale 4.0 \
--seed 42