🎼 Orpheus Music Transformer
SOTA symbolic music transformer · 2.3M+ MIDIs · 8k context · multi‑instrument co‑composition
⚡ Advanced architecture 748M parameters, full attention, RoPE, Flash Attention – optimised for speed and quality.
🎹 Efficient encoding 3 tokens per note, 7 per tri‑chord – rich music within 8k tokens.
🥁 Instrument‑aware Learns all 128 MIDI patches separately + dedicated drum channel.
🔚 Natural endings Special outro tokens for graceful conclusions.
🧪 Batch generation 10 continuations at once – preview, select, combine.
🎚️ Three models Medium (fast), Large (quality), Fine‑tuned (optimal).
📘 How to use – step by step
1
Start a composition
Upload a MIDI file (.mid, .midi, .kar) – the model works best with a clear musical phrase.
Or select up to 5 prime instruments (the app builds a simple chord from them).
Or just click “Generate” – the model starts from the start‑of‑sequence token.
2
Set generation parameters
Adjust prime tokens (max 6656), generation length, temperature, top‑p, and optionally add drums or an outro token.
3
Choose a model
Medium Base (fast), Large Base (higher quality), Large Fine‑Tuned (recommended for final pieces).
4
Generate & preview
Click “Generate”. You get 10 continuations (batches). Each tab shows audio preview, MIDI plot (with seed‑end marker), and downloadable MIDI file.
5
Assemble your final piece
Pick a batch, click “Add batch” – it appends to your composition. Preview the full result below. Add more batches or remove the last one as you like.
6
Export
The final composition is shown as audio, full MIDI plot, and a downloadable MIDI file. (Local files are saved in output_midis/.)
⚙️ Generation options explained
| Option | Description |
|---|
| Prime instruments | If no MIDI is uploaded, builds a short chord from these instruments. Drums are separate. |
| Number of prime tokens | How much of the uploaded MIDI is fed to the model. Truncate to the exact phrase you want continued. |
| Number of tokens to generate | Length of each continuation. 512–1024 tokens ≈ 30–60 seconds of music. |
| Temperature | Randomness in sampling. 0.8–1.0 is typical. Higher = more creative, lower = more repetitive. |
| Top‑p | Nucleus sampling – only tokens summing to probability p are considered. 0.96 works well; 1 = disabled. |
| Add drums | Prepends a simple random drum pattern – useful if your seed lacks drums. |
| Add outro | Appends an [OUTRO] token, encouraging the model to wrap up the piece. |
💡 Best practices tips
- A good seed is everything! Upload a MIDI with a clear, trimmed phrase.
- Trim your seed – use the number of prime tokens slider to cut exactly where you want the continuation to start.
- Keep it focused: Remove excessive instruments – 4–5 prominent ones work best.
- Ideal run time: 30–60 seconds (~1024–1536 tokens) gives the model enough context.
- Experiment: You may need several generation attempts to get a continuation you like.
- Don’t chase perfection: Enjoy the creative process and the unexpected ideas the model offers.
📌 Please note: Orpheus is primarily a continuation/co‑composition model. It works best when given musical context. Random generation from scratch may not always produce great results.
🧠 Available models
Medium Base – faster, good for quick experiments.
Large Base – higher quality, more nuanced.
Large Fine‑Tuned – optimised for musical coherence and quality (recommended).
📦 Output & export
All generated MIDI files include metadata and are playable in any DAW. Audio is rendered with a professional SoundFont (SGM‑v2.01). Final compositions can be downloaded as MIDI files.
🎵 Orpheus Music Transformer – developed by Alex Lev (Tegridy Code) exclusively for Project Los Angeles.
🐙 GitHub ·
🤗 Hugging Face Space ·
📄 License: Apache 2.0 · Year 2026
If you use this model in your work, please credit the author and link to the Hugging Face Space.
❤️ Enjoy making music with Orpheus!
🌓 Automatically adapts to your system light/dark theme.