Initial Audio has released Difro Melody AI, a standalone AI melody generator that runs entirely on the user’s machine.
Unlike cloud MIDI tools that send notes to a remote API, Difro uses a local VAE + Transformer model to generate and re-imagine MIDI in real time on CPU, or on GPU when a compatible card is available. No account, no cloud queue, no per-generation API bill.
The workflow is built for producers who already think in piano-roll terms. Hit record and play live into the piano roll, or drop in an existing MIDI clip. Difro analyses the musical DNA — melody, rhythm, harmony — and writes variations bar by bar. No MIDI file is required to start. Finished ideas drag straight into a DAW as MIDI, or bounce to a WAV loop via Render to WAV.
First Look
A walkthrough of the local AI co-producer, from live recording to dragging MIDI into a DAW:
YouTube.com/watch?v=WpfLTH8NHq8
How it works
Open the app, pick a prompt model, play or drop MIDI, then drag the result into Ableton, FL Studio, Logic, or Cubase:
YouTube.com/watch?v=OdYOhBL2JWA
Initial Audio says Difro is not a randomizer. It keeps the DNA of the original idea while writing something new. Record live or drag in chords / an old melody, then use the Creativity Slider to decide how far the model should take it:
YouTube.com/watch?v=oULNhCMnEXQ
Hit Render to WAV and Difro bounces the generated melody to an audio loop. Drag the WAV to the desktop, a folder, or straight into a DAW as a sample. Prefer MIDI? Click the output MIDI clip directly below the Render to WAV button and drag that out instead.
Initial Audio says that what sets it apart from typical “AI MIDI” utilities is that the model is not frozen. After a generation, the user can edit the output in the piano roll and rate it with a happy or sad face. Positive ratings — including ratings of edited MIDI — create a personal User Model: a copy of the original network that is nudged toward what that producer actually wants. Factory model and User Model can be switched at any time. If the User Model goes too far, it can be deleted from the menu (Delete User Feedback Model).
Difro ships with 10 prompt models built from 140+ MIDI files, with more models planned. The Pro version can also build custom prompt models from a folder of the user’s own MIDI, with no limit on how many files go in. Custom models stay 100% local, can be assigned a SourceLab preset, and can be shared as folders.
A built-in SourceLab Light engine (80+ presets) lets users audition ideas without loading an external instrument. Serum, Kontakt, Omnisphere and other instruments can still be hosted inside Difro (VST3 on Windows, Audio Units on macOS) so new melodies can be previewed with the user’s own presets before they hit the DAW.
GPU acceleration is optional. Compatible NVIDIA GPUs on Windows and Apple silicon GPUs on Mac can be selected from a CPU/GPU switch in the top-right corner. If no supported GPU is found Difro stays in CPU mode. Everything stays offline.
Under the hood, a VAE encodes the input MIDI into a latent map of rhythm, contour, and harmonic movement. A Transformer then predicts the next note — the same idea as an LLM, but for MIDI. Initial Audio describes it as a locally running MIDI GPT:
YouTube.com/watch?v=Buc_QKDVF6c
Features
- Local VAE + Transformer MIDI generation — 100% offline.
- Record live into the piano roll, or drop in MIDI — no MIDI file required to start.
- ReImagine variations with a Creativity Slider.
- Personal User Model: rate output (including edited MIDI) so the AI learns your taste.
- Switch between Original and User Model; delete the User Model anytime.
- 10 prompt models based on 140+ MIDI files (more to follow).
- Create custom prompt models from your own MIDI (Pro) — no file-count limit.
- Render to WAV and drag loops or MIDI into your DAW.
- Built-in SourceLab Light with 80+ presets.
- Host VST3 plugins on Windows, Audio Units on macOS.
- CPU or GPU: NVIDIA on Windows, Apple silicon on Mac.
- Multiple skins included.
Pricing & Availability
Difro Melody AI Pro is a one-time purchase at $49 (regular $69), including unlimited generations, custom models, royalty-free output, and lifetime updates.
A free version is available for Windows and macOS (50 generations per month, no credit card): Download for Windows / Download for macOS.
Read More