Loading...
Compose multitrack music from a style, description or MIDI. Choose a version, then listen, edit and download.
Input
Style, description or MIDI
MIDI formats
Output
Multitrack MIDI · ZIP originals
Use case
Musical ideas, band arrangements and MIDI continuation
Compose multitrack music from a style, description or MIDI. Choose a version, then listen, edit and download.
Seed 42 · 35 s · 5 tracks
Playback is synthesized from MIDI. Its instrument sounds do not measure the model’s audio quality.
MIDILM
Import SF2/SF3, up to 512 MB per file. Files stay on this page locally and are not uploaded; transcription notes stay unchanged.
For audio or video projects at 48 kHz. Files are slightly larger.
WAV is synthesized directly at the chosen format using the current sounds, speed and volume. Sound quality also depends on the sound library; more bits cannot improve its source samples.
Original MIDIComplete original ZIPGeneration record
{
"model_profile": "midilm",
"arrangement_seed": 42,
"arrangement_mode": "create",
"arrangement_options": {
"prompt": "A warm jazz quartet in C major at 110 BPM, 4/4 time. Piano plays a clear lyrical melody, acoustic bass keeps a steady pulse, clean guitar adds soft chords, and drums keep a light swing groove. The harmony follows Dm7, G7, Cmaj7 and Am7, with balanced dynamics and recurring melodic phrases.",
"max_tokens": 2047,
"temperature": 1
}
}SHA-256: 60531fb6edff7267c65e932ec148adf26c319b9ca68e4bf9ebf7cd9eda42e7fc
Include the tool name, steps and the problem in your feedback. Screenshots can help us locate it.
Generate multi-instrument MIDI from a text description.
Paper result: MOS ↑ 3.2306 ± 0.9160 / 5
Computation: GPU · model · License: CC-BY-NC-4.0 (code)
Turn descriptions of style, mood and instrumentation into multitrack MIDI.
Paper result: FAD ↓ 0.173 · CLAP ↑ 22.1 · BF16
Paper Table 2: 896 test samples, FluidR3_GM, VGGish FAD and CLAP × 100. Speed uses 2K tokens on one L40S. No matching measurement is available for this site’s FP32 version.
Computation: GPU · model · License: Llama-3.2-Community
Large fine-tuned on quality data; not necessarily better for every style.
No comparable published music-quality score is available for this version.
Computation: GPU · model · License: Apache-2.0
Base checkpoint · 16 layers, 16 attention heads.
No comparable published music-quality score is available for this version.
Computation: GPU · model · License: Apache-2.0
Base checkpoint · 8 layers, 32 attention heads.
No comparable published music-quality score is available for this version.
Computation: GPU · model · License: Apache-2.0
4 / 8-bar context; medium checkpoint, 350.5 MB.
No comparable published music-quality score is available for this version.
Computation: GPU · model · License: Research / non-commercial weights; MIT code
4 / 8 / 12 / 16 bars; 4096-token context.
No comparable published music-quality score is available for this version.
Computation: GPU · model · License: Research / non-commercial weights; MIT code
4 / 8 / 12 / 16 bars; velocity and microtiming expression.
No comparable published music-quality score is available for this version.
Computation: GPU · model · License: Research / non-commercial weights; MIT code
4 / 8-bar context; small checkpoint, 82.5 MB.
No comparable published music-quality score is available for this version.
Computation: GPU · model · License: Research / non-commercial weights; MIT code
Length conditioning uses ten-bar ranges.
Paper result: PhraseFID ↓ 3.81
Computation: GPU · model · License: Apache-2.0 (code)
Unconditional generation without length or structural constraints.
Paper result: PhraseFID ↓ 3.89
Computation: GPU · model · License: Apache-2.0 (code)
Experimental structure conditioning; adherence still needs validation.
Paper result: PhraseFID (7.65)
PhraseFID and SRS under a fixed structure prompt are excluded from cross-version ranking.
Computation: GPU · model · License: Apache-2.0 (code)
Style and chord rules generate loops, parts and complete song structures.
No comparable published music-quality score is available for this version.
Computation: CPU · rules · License: GPL-3.0-or-later
Band instrumentation and form templates generate complete arrangements.
No comparable published music-quality score is available for this version.
Computation: CPU · rules · License: MIT
Scores use different datasets and methods. They are not a unified ranking or measured audio quality of this adapted runtime.
Ranks use the selected metric within the same evaluation conditions. Original records include reference music and ablations; these are not necessarily versions available here.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Paper Tables 1–3: test metrics average 10 random samples of 96 items; free text uses a separate set of 96. MOS is a blind 1–5 rating by 15 listeners; ± denotes standard deviation. Text conditions are ranked separately.
± indicates standard deviation.
| Metric rank | Model | Test setting | MOS / 5↑ |
|---|---|---|---|
| 1 | MIDILM | MidiCaps test set | 3.2306 ± 0.9160 |
| 2 | Text2MIDI-InferAlign | MidiCaps test set | 2.4861 ± 0.7675 |
| 3 | MuseCoco | MidiCaps test set | 2.0854 ± 0.9499 |
| - | Reference music Reference data; excluded from version ranking. | MidiCaps test set | 4.1010 ± 0.9590 |