How jfayyad/lesiongen works
LesionGen trains a concept-guided diffusion model on dermatology image–caption data, generates conditioned lesion images, and evaluates classification performance on real or synthetic datasets. The README documents both LoRA and full-model training, while the sampled generator loads a pretrained Stable Diffusion pipeline and LoRA weights. Classification training uses a data-loading module; separate evaluation scripts assess generated or test images. Training-script internals and the referenced test loader were not sampled, so their relationships are shown only where the README or source index supports them.
Main entry points: generate.py Conditioned generation · train_lora.sh LoRA training · train_SD.sh Full-model training · eval_synth.py Synthetic evaluation
Components 16
Dermatology data
- HAM10000 / D7P
- Concept-rich captions
- Synthetic images
Image synthesis
- LoRA training
train_lora.sh - Full-model training
train_SD.sh - Trained diffusion model
train_lora.sh - Conditioned generation
generate.py - Stable Diffusion
Classification
- Dataset loading
classification/dataloader.py - Classifier training
classification/classifier.py - Classifier weights
Evaluation
- Synthetic evaluation
classification/eval_synth.py - Prediction evaluation
classification/eval_with_pred.py - Generated-image evaluation
classification/eval.py - Evaluation data loader
classification/test_loader.py
Other components
- Researcher
Connections 21
- Researcher → LoRA training: starts training
README.md - Researcher → Full-model training: starts training
README.md - HAM10000 / D7P → Concept-rich captions: paired with captions
README.md - Concept-rich captions → LoRA training: trains on pairs
README.md - Concept-rich captions → Full-model training: trains on pairs
README.md - LoRA training → Trained diffusion model: fine-tunes model
README.md - Full-model training → Trained diffusion model: fine-tunes model
README.md - Researcher → Conditioned generation: provides condition
README.md - Conditioned generation → Stable Diffusion: loads pipeline
generate.py - Conditioned generation → Trained diffusion model: loads LoRA weights
generate.py - Conditioned generation → Synthetic images: writes images
generate.py - HAM10000 / D7P → Dataset loading: loads image data
classification/dataloader.py - Dataset loading → Classifier training: provides batches
classification/classifier.py - Classifier training → Classifier weights: saves weights
classification/classifier.py - Synthetic images → Synthetic evaluation: evaluated as input
README.md - Evaluation data loader → Synthetic evaluation: provides batches
classification/eval_synth.py - Classifier weights → Synthetic evaluation: loads weights
classification/eval_synth.py - Evaluation data loader → Prediction evaluation: provides batches
classification/eval_with_pred.py - Classifier weights → Prediction evaluation: loads weights
classification/eval_with_pred.py - Synthetic images → Generated-image evaluation: evaluated as input
classification/eval.py - Classifier weights → Generated-image evaluation: loads weights
classification/eval.py