Skip to content

Releases: koda-dernet/Side-Step

v1.0.0-beta

05 Mar 03:52
@koda-dernet koda-dernet
v1.0.0-beta
b352086
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v1.0.0-beta Pre-release
Pre-release

v1.0.0-beta - "Last Light"

Complete overhaul of Side-Step

Full Feature Set

Training & Adapters

  • LoRA + LoKR + LoHA + OFT adapters with DoRA support
  • Auto-configured timestep sampling — Turbo gets discrete 8-step, Base/SFT gets continuous logit-normal + CFG dropout
  • Gradient accumulation, mixed precision, gradient clipping
  • Learning rate scheduling — cosine, linear, constant, warmup
  • VRAM presets — 8/12/16/24 GB+, Quick Test, High Quality, Recommended
  • Live VRAM estimation — segmented bar (model + activations + optimizer)

Data Pipeline

  • Two-pass preprocessing — peak/LUFS normalization, chunking, tensor export
  • Dataset building — folder scan + sidecar metadata → dataset.json
  • Audio analysis — duration detection with 4-level fallback (soundfile → torchcodec → mutagen → ffprobe)
  • Caption enrichment — local (Qwen-Omni), Gemini, OpenAI, lyrics-only modes
  • Sidecar management — bulk tag operations, format conversion, instrumental flag

Analysis & Optimization

  • Fisher Information analysis — per-layer importance for rank allocation
  • PP++ (Preprocessing++) — adaptive chunking based on audio characteristics
  • Cross-projection targeting — fine-tune specific attention layers
  • Export to ComfyUI — native, PEFT, and LoKR formats with alpha normalization

User Interfaces

  • Desktop GUI — Ez Mode, Advanced, Monitor, Lab (History, Audio Library, Preprocess, PP++, Export)
  • Interactive Wizard — terminal prompts with back-navigation and flow chaining
  • Full CLI — every argument documented with defaults, subcommands for train/preprocess/analyze/captions/dataset/export/gui
  • Live progress telemetry — JSONL streaming for GUI updates

Platform & Ops

  • Cross-platform installers — Windows (bat/ps1), Linux/macOS (shell) with uv, Python 3.11, Node.js, model checkpoints
  • GPU detection — CUDA, MPS (Apple Silicon), XPU (Intel), CPU fallback
  • TensorBoard integration — per-layer gradient norms, learning rate tracking, versioned log directories
  • Settings persistence — platform-aware config directory (APPDATA/XDG_CONFIG_HOME)

Download: Clone the repo and run uv run sidestep gui or uv run sidestep for the wizard.

Note: This is beta software maintained by a single developer. If you encounter issues, please open an issue on GitHub. Commercial use requires written permission — see the LICENSE file for details.

UI & Themes

  • New default theme: "Last Light" — desolate urban glow aesthetic
  • Additional themes: Amber Terminal, Phosphor Green, Neon Rose. Support for custom themes

License & Legal

  • CC BY-NC-SA 4.0 — attribution required, commercial use needs permission
  • User liability disclaimer — tool neutrality for training data/outputs
  • Clear commercial contact — GitHub profile for licensing inquiries

Documentation

  • Updated README — uv run commands, screenshots, license summary
  • Requirements clarified — canonical source is pyproject.toml
  • Obsidian docs — cleaned up references

Breaking Changes

  • License change — from MIT to CC BY-NC-SA 4.0
  • Git history reset — full overhaul, previous commits archived

No binaries yet as this is a beta.

Loading