How git03-khushi/jit-cifar10-image-generation works
This repository is a notebook-based, reduced-scale JiT-inspired diffusion experiment for CIFAR-10. Its documented workflow loads and preprocesses data, trains a patch-based Transformer denoiser under multiple patch-size and time-embedding configurations, then evaluates and visualizes results and saves checkpoints. The notebook itself is the only implementation file in the tree; component boundaries and edges below reflect README descriptions, not inspected notebook cells.
Main entry points: JiT Project.ipynb Model training
Components 16
Data preparation
- Image preprocessing
JiT Project.ipynb - Noisy input
JiT Project.ipynb
Denoising model
- Patch Transformer
JiT Project.ipynb - Time embedding
JiT Project.ipynb - Clean-image prediction
JiT Project.ipynb - Velocity MSE objective
JiT Project.ipynb
Training experiments
- Patch-4 run
JiT Project.ipynb - Patch-8 run
JiT Project.ipynb - Time ablation
JiT Project.ipynb - Model training
JiT Project.ipynb
Evaluation outputs
- Test-loss evaluation
JiT Project.ipynb - Loss visualization
JiT Project.ipynb - Sample and denoising views
JiT Project.ipynb - Checkpoint saving
JiT Project.ipynb
Other components
- Researcher
- CIFAR-10
Connections 15
- Researcher → Image preprocessing: runs notebook
README.md - CIFAR-10 → Image preprocessing: loads data
README.md - Image preprocessing → Noisy input: prepares images
README.md - Noisy input → Patch Transformer: feeds denoiser
README.md - Time embedding → Patch Transformer: conditions model
README.md - Patch Transformer → Clean-image prediction: predicts clean image
README.md - Clean-image prediction → Velocity MSE objective: derives prediction
README.md - Patch-4 run → Model training: trains configuration
README.md - Patch-8 run → Model training: trains configuration
README.md - Time ablation → Model training: trains configuration
README.md - Model training → Velocity MSE objective: minimizes loss
README.md - Model training → Test-loss evaluation: evaluates model
README.md - Model training → Loss visualization: visualizes loss
README.md - Model training → Sample and denoising views: generates visualizations
README.md - Model training → Checkpoint saving: saves weights
README.md