AI GitHub
AI Toolkit for configurable diffusion training jobs
Run documented adapter and model-training recipes using explicit job configuration.
Key Takeaways
- The toolkit organizes training in job files that specify the base model, dataset, output and sample generation.
- Start from a job example that matches the model architecture.
- Use a small captioned dataset and save fixed samples at intervals.
- Supported architectures and memory requirements vary by recipe.
How it works
The toolkit organizes training in job files that specify the base model, dataset, output and sample generation. That makes a run inspectable before it consumes accelerator time.
What you need
Start from a job example that matches the model architecture. Treat source images, captions and sample prompts as part of the job's versioned inputs.
What to try first
Use a small captioned dataset and save fixed samples at intervals. Inspect the configuration and compare checkpoints against the same base prompts.
Include a prompt outside the dataset's common composition. Compare the adapter and base model so you can see whether the change is general or tied to memorized framing.
Where to look for failures
Supported architectures and memory requirements vary by recipe. Review the model terms and use a representative holdout to spot overfitting.
Related Franklin pages
Separate adapter configuration from evaluation. Use held-out prompts and keep a base-model output so changes in memorization and generalization are visible.
- Adapt a supported model by training a subset of parameters and saving compact artifacts. PEFT for training and loading model adapters
- Configure image-model adaptation jobs through a graphical training interface. kohya_ss for configuring diffusion model training
Franklin’s image tests give you examples of outputs to compare across visual models. Franklin’s image tests
Project documentation
Try it for your workflow
Explore the project’s documentation for setup and usage, or save this page to a collection for later.
Read the project documentation ↗