How invoke-ai/invoke-training works
Project type and purpose
• This is an open-source Python library (invoke-training) for configuring and launching custom Stable Diffusion training workflows (fine-tuning, LoRA, textual inversion, etc.), usable via both a CLI and a GUI.
• The README highlights core use cases: stable diffusion (SD) training modes, SDXL training modes, and integration with InvokeAI.
File structure
• Top-level directories:
• src/invoke_training – the main codebase
• src/invoke_training/_shared – shared utilities (accelerator, data, checkpoints, optimizer, flux integrations, stable_diffusion helpers)
• src/invoke_training/config – pydantic-style pipeline and data/optimizer config definitions
• src/invoke_training/pipelines – orchestration of various training modes (stable_diffusion, stable_diffusion_xl, flux, experimental)
• src/invoke_training/model_merge – tools to extract and merge LoRA weights
• src/invoke_training/scripts – convenience wrappers and experiments
• src/invoke_training/ui – GRadio-based GUI (app, config groups, blocks, pages)
• src/invoke_training/sample_configs – YAML examples for different GPUs and modes
• tests – unit tests for every module
• CI/CD: .github/workflows (test, deploy), pre-commit, mkdocs.yml for documentation.
• Docs live under docs/ (architecture, getting started, guides, API reference).
README for additional insights
• Architecture is config-driven: pipelines accept YAML configs, build data loaders, optimizers, model components, run training loops.
• Dual interface: CLI (“invoke-train”) and GUI (“invoke-train-ui”) expose the same underlying pipeline engine.
• Training progress streamed to TensorBoard. Models saved as LoRA or full checkpoints, compatible with InvokeAI.
Detail
• Layer: interface, config, orchestration, compute, persistence, monitoring, extensions.