Instructions to use ResonateIntegrated/Resonate-MusiX-T1-50M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use ResonateIntegrated/Resonate-MusiX-T1-50M with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-to-audio", model="ResonateIntegrated/Resonate-MusiX-T1-50M")# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("ResonateIntegrated/Resonate-MusiX-T1-50M", device_map="auto") - sentence-transformers
How to use ResonateIntegrated/Resonate-MusiX-T1-50M with sentence-transformers:
from sentence_transformers import SentenceTransformer model = SentenceTransformer("ResonateIntegrated/Resonate-MusiX-T1-50M") sentences = [ "The weather is lovely today.", "It's so sunny outside!", "He drove to the stadium." ] embeddings = model.encode(sentences) similarities = model.similarity(embeddings, embeddings) print(similarities.shape) # [3, 3] - Notebooks
- Google Colab
- Kaggle
Resonate MusiX T1 (Piano & Melodic Edition)
Next-Generation Text-Conditioned Piano & Symbolic Music Transformer
Model Summary
Resonate MusiX T1 (Piano & Melodic Edition) is a 50.41-million parameter autoregressive decoder transformer specialized in generating expressive piano compositions, intricate keyboard counterpoints, and memorable melodic themes directly from natural language prompts.
Trained on a curated corpus of 10,000 multi-voice compositions, the model demonstrates strong proficiency in acoustic piano voicing, lyrical right-hand melodies, dynamic arpeggiated basslines, and rich harmonic chord progressions.
Natural language prompts guide key musical attributes including emotional tone, tempo, playing style, dynamic touch, and compositional complexity via continuous cross-attention conditioning.
Piano & Keyboard Capabilities
- Two-Hand Polyphony: Independently coordinates lower-register bass chords/arpeggios with expressive upper-register melody lines.
- Dynamic Phrasing: Supports 16 quantized velocity levels, reflecting nuanced key strike intensity from gentle pianissimo to striking forte.
- Zero-Latency Attack: Generation starts immediately on beat one without introductory pauses or empty measures.
- Anti-Monotony Guard: Built-in repetition penalty prevents repetitive note hammering, encouraging flowing scalar runs and harmonic transitions.
Local Inference (CLI)
Run generation locally on CPU or CUDA GPU:
1. Requirements
pip install torch miditok sentence-transformers symusic
2. Files Required
Keep these files in the same directory:
generate.pyresonate_musix_t1.ptresonate_tokenizer.json
3. Usage
python generate.py
Options:
- Prompt: Text description of the desired music style, tempo, or mood.
- Count: Number of tracks to generate (default: 1).
- Mode:
1: Dynamic Melody (balanced phrasing and transitions)2: Strict Harmony (tight chord alignment)3: Wild Solo (complex melodic variations)
Output .mid files are saved directly in the project directory.
Architecture Details
Resonate MusiX T1 replaces conventional GPT-2 building blocks with modern architectural primitives:
- Rotary Position Embeddings (RoPE): Operates on complex inner-product rotation ($\theta = 10000.0$), ensuring relative time and bar structure awareness across variable sequence lengths.
- SwiGLU Non-Linearity: Uses a gated linear unit with Swish activation ($d_{ffn} = 1728$, rounded to multiples of 64), providing superior representation capacity per parameter.
- RMSNorm (Root Mean Square Normalization): Placed pre-attention and pre-FFN with $\epsilon = 10^{-6}$ for numerical stability under mixed-precision regimes.
- Text Conditioning via Cross-Attention: Interleaved cross-attention blocks inject 384-dimensional semantic embeddings from
sentence-transformers/all-MiniLM-L6-v2directly into every transformer layer. - Anti-Repetition & Anti-Silence Engine: Built-in penalty ($1.28$) and leading rest suppression eliminate silence and note-spamming loops during generation.
Technical Specifications
| Parameter | Value |
|---|---|
| Model Type | Autoregressive Transformer with Cross-Attention |
| Primary Domain | Expressive Piano & Melodic Polyphonic Arrangements |
| Total Parameters | 50,412,800 (50.41M) |
| Checkpoint Size | 192.4 MB (.pt / Safetensors) |
| Context Window ($N_{ctx}$) | 1024 tokens (~30-45 seconds of continuous polyphony) |
| Hidden Dimension ($d_{model}$) | 640 |
| Number of Layers | 8 |
| Attention Heads | 8 ($d_{head} = 80$) |
| Intermediate FFN Size | 1728 (SwiGLU) |
| Text Embedding Dimension | 384 |
| Tokenization Format | REMI (Revamped MIDI) |
| Vocabulary Size | 1,024 tokens |
| Training Precision | bfloat16 / TensorFloat-32 (TF32) |
| Final Loss | 0.75 Cross-Entropy |
Piano Prompt Engineering Reference
| Style Category | Sample Conditioning Prompt |
|---|---|
| Emotional Ballad | emotional slow romantic bell piano melody with warm acoustic chords 80 bpm |
| Lyrical Classical | peaceful classical acoustic piano solo with flowing arpeggios and gentle dynamics |
| Nostalgic Minor Key | melancholic piano theme in minor key with deep bass octaves and delicate melody |
| Uplifting & Bright | bright joyful piano melody with fast upbeat chord changes and energetic rhythm |
| Cinematic Ambient | ambient cinematic piano soundtrack theme with spacious reverberant harmonies |
| Pop / Neo-Classical | modern catchy piano hook melody with rhythmic left hand syncopation |
Playback & Sound Quality
The model outputs standard symbolic MIDI (.mid). For high-fidelity acoustic playback, open generated files using virtual piano instruments:
- DAWs / VSTs: Keyscape, Pianoteq, Native Instruments The Grandeur, Addictive Keys.
- SoundFonts (Free): FluidR3_GM, Salamander Grand Piano, GeneralUser GS.
Limitations
- Instrument Assignment: Outputs universal MIDI event streams. Instrument realism depends on the downstream SoundFont or VST instrument loaded in your player.
- Context Boundary: Coherence remains optimal up to 1024 tokens (~45 seconds of audio).
Citation
@misc{resonate_musix_t1_2026,
author = {Resonate MusiX Team},
title = {Resonate MusiX T1: Dense Polyphonic Piano & Symbolic Music Transformer},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co}}
}
License
This project is distributed under the MIT License.
- Downloads last month
- 28
Evaluation results
- Training Lossself-reported0.750