How koda-dernet/side-step works
Side-Step is a standalone ACE-Step audio-model adapter fine-tuning toolkit. Users prepare audio datasets and sidecar metadata, optionally enrich or analyze them, preprocess audio into training tensors, configure and run adapter training, then inspect runs or export compatible adapters. The GUI, CLI/wizard, and TUI are alternate entry points into these capabilities. The graph combines README-documented product stages with sampled call sites; wiring for unsampled subsystems is omitted where it cannot be established.
Main entry points: trainer.py Adapter trainer · task_manager.py GUI task manager · server.py GUI API server · train.py CLI entry point
Components 22
User interfaces
- Desktop GUI
frontend/js - GUI API server
sidestep_engine/gui/server.py - Interactive wizard
sidestep_engine/ui/wizard.py - CLI entry point
train.py - Terminal interface
sidestep_engine/tui/app.py
Dataset preparation
- Dataset builder
sidestep_engine/data/dataset_builder.py - Sidecar enrichment
sidestep_engine/data - Offline audio analysis
sidestep_engine/analysis/audio_analysis.py - Audio preprocessing
sidestep_engine/data/preprocess.py - Dataset manifest
Analysis and training
- Fisher PP++ analysis
sidestep_engine/analysis/fisher - Model loading
sidestep_engine/models/loader.py - Adapter trainer
sidestep_engine/core/trainer.py - Trained adapter
- ComfyUI export
sidestep_engine/core/comfyui_export.py
Run operations
- GUI task manager
sidestep_engine/gui/task_manager.py - Run history and files
sidestep_engine/gui/file_ops.py - Training monitor
frontend/js/training.js
Other components
- User
- Audio files
- Caption and lyrics providers
- Model checkpoints
Connections 26
- User → Desktop GUI: uses
- User → Interactive wizard: uses
- User → CLI entry point: runs
- User → Terminal interface: uses
- Desktop GUI → GUI API server: requests
- GUI API server → GUI task manager: dispatches tasks
- GUI task manager → Audio preprocessing: runs preprocessing
- GUI task manager → CLI entry point: spawns training
- CLI entry point → Adapter trainer: starts training
- Interactive wizard → CLI entry point: exports command
- Audio files → Dataset builder: provides files
- Dataset builder → Dataset manifest: writes manifest
- Audio files → Sidecar enrichment: provides files
- Caption and lyrics providers → Sidecar enrichment: supplies metadata
- Audio files → Offline audio analysis: analyzes
- Dataset manifest → Audio preprocessing: guides preprocessing
- Audio preprocessing → Adapter trainer: prepares tensors
- Model checkpoints → Model loading: loads weights
- Model loading → Adapter trainer: provides model
- Dataset manifest → Fisher PP++ analysis: supports analysis
- Fisher PP++ analysis → Adapter trainer: provides rank map
- Adapter trainer → Trained adapter: saves adapter
- Trained adapter → ComfyUI export: converts PEFT adapter
- GUI API server → Run history and files: requests run history
- GUI task manager → Training monitor: streams progress
- Training monitor → GUI API server: receives updates