Candy Dungeon Music Forge (CDMF)

Candy Dungeon Music Forge (CDMF) is a local-first AI music workstation for Windows. It runs on your PC, uses your GPU, and keeps your prompts and audio on your hardware. CDMF is powered by ACE-Step (text → music diffusion) and includes a custom UI for generating tracks, managing a library, and training LoRAs.

Status: v0.1

What you can do

• Generate music from a prompt (optionally with lyrics)
• Use a built-in Music Player + library view (sort, favorite, categorize)
• Save and reuse presets
• (Optional) Stem separation to rebalance vocals vs instrumentals
• Train ACE-Step LoRAs from your own datasets
• Dataset helpers:
  • Mass-create _prompt.txt / _lyrics.txt files
  • (Optional) Auto-tag datasets using MuFun-ACEStep (experimental)

System requirements

Minimum:

• Windows 10/11 (64-bit)
• NVIDIA GPU (RTX strongly recommended)
• ~10–12 GB VRAM (more = more headroom)
• SSD with tens of GB free (models + audio + datasets)

Comfortable:

• RTX GPU with 12–24 GB VRAM
• 32 GB RAM
• Fast NVMe SSD
• Comfort reading console logs when something goes wrong

Install and run (recommended)

1. Download the latest release (installer)
2. Run CandyDungeonMusicForge-Setup.exe
3. Launch Candy Dungeon Music Forge from the Start Menu

Default install location:

• %LOCALAPPDATA%\CandyDungeonMusicForge

First launch notes

On first run, CDMF does real setup work:

• Creates a Python virtual environment (e.g. venv_ace)
• Installs packages from requirements_ace.txt
• Downloads ACE-Step and related models as needed
• Installs helpers like audio-separator

A console window (“server console”) appears and must stay open while CDMF runs. CDMF will open a loading page in your browser and then load the full UI when ready.

Using CDMF (high-level workflow)

1. Launch CDMF and wait for the UI
2. Go to Generate → create tracks from prompt (and lyrics if desired)
3. Browse/manage tracks in Music Player
4. (Optional) Use stem controls to adjust vocal/instrumental balance
5. (Optional) Build a dataset and train a LoRA in Training

Generation basics

• Prompt: your main ACE-Step tags / description (genre, instruments, mood, context)
• Instrumental mode:
  • Lyrics are not used
  • CDMF uses the [inst] token so ACE-Step focuses on backing tracks
• Vocal mode:
  • Provide lyrics using markers like [verse], [chorus], [solo], etc.
• Presets let you save/load a whole “knob bundle” (text + sliders)

Stem separation (vocals vs instrumentals)

CDMF can run audio-separator as a post-process step so you can rebalance:

• Vocals level (dB)
• Instrumental level (dB)

First use requires downloading a large stem model and adds a heavy processing step. For fast iteration: generate with both gains at 0 dB, then only use stems once you like a track.

LoRA training

Switch to the Training tab to configure and start LoRA runs.

Dataset structure

Datasets must live under:

<CDMF root>\training_datasets

For each audio file (foo.mp3 or foo.wav), provide:

• foo_prompt.txt — ACE-Step prompt/tags for that track
• foo_lyrics.txt — lyrics, or [inst] for instrumentals

CDMF includes tools to bulk-create these files (and optionally auto-generate them with MuFun-ACEStep).

Training parameters (examples)

• Adapter name (experiment name)
• LoRA config preset (JSON from training_config)
• Epochs / max steps
• Learning rate (commonly 1e-4 to 1e-5)
• Max clip seconds (lower can reduce VRAM and speed up training)
• Optional SSL loss weighting (set to 0 for some instrumental datasets)
• Checkpoint/save cadence

Experimental: MuFun-ACEStep dataset analyzer

MuFun-ACEStep can auto-generate _prompt.txt and _lyrics.txt files from audio. It’s powerful but:

• The model is large (tens of GB)
• Outputs aren’t perfect—skim and correct weird tags/lyrics before training

Troubleshooting

• First launch takes forever: check console for pip/model download errors; verify disk space and network
• No .wav files found: generate a track; confirm Output Directory matches the Music Player folder
• CUDA / VRAM OOM:
  • Reduce target length during generation
  • Reduce max clip seconds during training
  • Lower batch/grad accumulation if you changed them

Contributing

Issues and PRs welcome. If you’re changing anything related to training, model setup, or packaging, please include:

• what GPU/driver you tested on
• exact steps to reproduce any bug you fixed

(Consider adding CONTRIBUTING.md once you have preferred norms.)

License

This project’s source code is licensed under the Apache License 2.0. See LICENSE.

Note: Model weights and third-party tools used by CDMF (ACE-Step, PyTorch, audio-separator, MuFun-ACEStep, any LLM backend, etc.) are covered by their own licenses/terms.

Trademarks

“Candy Dungeon”, “Candy Dungeon Music Forge”, and associated logos/branding are trademarks of the project owner and are not granted under the Apache-2.0 license.

See TRADEMARKS.md for permitted use (e.g., descriptive references are fine; distributing a fork under the same name/logo is not).

Support

If you find CDMF useful and want to support development, you can:

• email support@candydungeon.com for more info
• Contribute to the creator's Ko-Fi and buy him a coffee/cigar if you want: https://ko-fi.com/davidhagar