Upload 4 files

Mix Directive Stem Splitter v1 (4-stem)

LICENSE.md

@@ -0,0 +1,78 @@
+MIX DIRECTIVE STEM SPLITTER — MODEL LICENSE (v1.0)
+
+Copyright (c) 2026 Mix Directive. All rights reserved.
+
+This license covers the "Mix Directive Stem Splitter" model weights and
+the files published with them in this repository (the "Model"). The Model
+is proprietary. It is made available for download only so that the Mix
+Directive application can fetch and run it. By downloading, copying or
+using the Model you agree to these terms. If you do not agree, do not
+download or use the Model.
+
+1. PERMITTED USE
+   You may download the Model and run it only through an unmodified,
+   genuine copy of the Mix Directive application, under that
+   application's own license, to separate audio that you have the right
+   to process.
+
+2. RESTRICTIONS
+   Except as allowed in section 1, you may NOT, and may not permit any
+   third party to:
+   (a) use, load or run the Model in any other software, service, script
+       or product, whether commercial or not;
+   (b) copy, redistribute, publish, sublicense, sell, rent, lend or
+       otherwise make the Model available to anyone, including by
+       re-uploading it to any website, hub or mirror;
+   (c) modify, fine-tune, retrain, quantise, convert, merge or distil the
+       Model, or create any derivative model from it or from its outputs
+       used as training targets;
+   (d) use the Model, its weights, architecture details or outputs to
+       develop, train, evaluate or improve any competing product, model or
+       service;
+   (e) reverse engineer the Model to extract information about the data it
+       was trained on;
+   (f) remove or alter this license or any proprietary notice in the
+       Model files;
+   (g) use the Model in any way that violates applicable law or the rights
+       of others.
+
+3. YOUR AUDIO AND OUTPUTS
+   You keep all rights you have in the audio you process. Mix Directive
+   claims no ownership of the separated stems the Model produces from
+   your audio. You are responsible for having the right to process that
+   audio and to use the results.
+
+4. OWNERSHIP
+   The Model is licensed, not sold. Mix Directive retains all right,
+   title and interest in the Model, including all intellectual property
+   rights. No rights are granted by implication, estoppel or otherwise
+   except as expressly stated here.
+
+5. NO TRAINING DATA
+   The Model does not contain or distribute the recordings it was trained
+   on. Nothing in this repository grants any right to those recordings.
+
+6. WARRANTY DISCLAIMER
+   THE MODEL IS PROVIDED "AS IS" WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+   IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, ACCURACY OR
+   NON-INFRINGEMENT. SEPARATION RESULTS MAY CONTAIN ARTEFACTS, BLEED OR
+   MISSING CONTENT.
+
+7. LIMITATION OF LIABILITY
+   TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, MIX DIRECTIVE WILL
+   NOT BE LIABLE FOR ANY INDIRECT, INCIDENTAL, SPECIAL, CONSEQUENTIAL,
+   EXEMPLARY OR PUNITIVE DAMAGES, OR FOR ANY LOSS OF PROFITS, DATA OR
+   GOODWILL, ARISING OUT OF OR IN CONNECTION WITH THE MODEL OR THIS
+   LICENSE, WHATEVER THE LEGAL THEORY.
+
+8. TERMINATION
+   This license ends automatically, without notice, if you breach any of
+   its terms. On termination you must stop using the Model and delete
+   every copy in your possession or control.
+
+9. GOVERNING LAW
+   This license is governed by the laws of the jurisdiction in which Mix
+   Directive is established, excluding its conflict-of-law rules.
+
+Questions about this license: hello@mixdirective.example

README.md

@@ -1,5 +1,50 @@
 ---
 license: other
 license_name: mix-directive-model-license
-license_link: LICENSE
+license_link: LICENSE.md
+library_name: onnx
+pipeline_tag: audio-to-audio
+tags:
+  - music-source-separation
+  - stem-separation
+  - mix-directive
 ---
+
+# Mix Directive Stem Splitter (4-stem, v1)
+
+The stem separation model used by the [Mix Directive](https://mixdirective.example) desktop
+app. It splits a stereo mix into **drums**, **bass**, **other** and **vocals**.
+
+**This model is proprietary.** It is published here only so the Mix Directive app can
+download it on demand. Using it outside the app, redistributing it, or building on it is
+not permitted: see [LICENSE.md](LICENSE.md).
+
+## Files
+
+| File | What it is | SHA-256 |
+|---|---|---|
+| `model_fp16.onnx` | the model (ONNX, fp16 weights, 56 MB) | `f45b4b79dd08b9e0237bd7420b735659b043edb25b572949dbb4ae5ce3fe110b` |
+| `model_config.json` | sources, sample rate, segment length, input normalisation | `af642c9ac4e3812449c847ec5bc969b15fa791faa4a5d14f2d2de3b7968db809` |
+
+The app pins a specific commit of this repository and checks each file's SHA-256 before
+using it.
+
+## Model
+
+- Architecture: Hybrid Transformer Demucs (the open-source `demucs` 4.1 architecture, MIT),
+  22.6M parameters, 44.1 kHz stereo, 6 s segments.
+- Weights: trained from random initialisation by Mix Directive. No pretrained third-party
+  weights were used.
+- Training data: music owned by the Mix Directive author, used with full rights. The model
+  does not contain or reproduce those recordings.
+
+## Intended use
+
+Separating music the user has the right to process, inside the Mix Directive app (stem
+extraction before mixing and Auto-Mix). Results can contain bleed between stems and
+artefacts, especially in the "other" stem.
+
+## License
+
+Mix Directive Stem Splitter Model License v1.0 — see [LICENSE.md](LICENSE.md). All rights
+reserved.

model_config.json

@@ -0,0 +1,52 @@
+{
+ "architecture": "HTDemucs (demucs 4.1.0, MIT)",
+ "sources": [
+  "drums",
+  "bass",
+  "other",
+  "vocals"
+ ],
+ "samplerate": 44100,
+ "audio_channels": 2,
+ "model_cfg": {
+  "channels": 44,
+  "t_layers": 5,
+  "t_heads": 8,
+  "bottom_channels": 0,
+  "segment": 6.0
+ },
+ "model_name": "medium",
+ "nparams": 22612960,
+ "checkpoint": "best",
+ "step": 120450,
+ "valid_sdr": {
+  "drums": 11.869,
+  "bass": 8.689,
+  "other": 2.352,
+  "vocals": 10.811,
+  "mean": 8.43,
+  "n_songs": 44
+ },
+ "valid_sdr_mix": {
+  "drums": 7.507,
+  "bass": 8.648,
+  "other": 0.749,
+  "vocals": 5.304,
+  "mean": 5.552,
+  "n_songs": 41
+ },
+ "provenance": "own weights: random init, trained on the user's Suno stems (sum-of-stems mixtures); stage 2: fine-tuned from our stage-1 weights (best.pt, step 85702, sha256 ad21ed8115643425\u2026) with Suno's delivered mixes as part of the inputs and a multi-resolution STFT loss",
+ "load": "separation_train.build_model(model_cfg, sources).load_state_dict(torch.load('weights.pth'))",
+ "input_normalisation": "x = (mix - mean(mono)) / std(mono); output * std + mean (as demucs apply)",
+ "onnx": {
+  "file": "model_fp16.onnx",
+  "input": "mix",
+  "shape": [
+   1,
+   4,
+   2,
+   264600
+  ],
+  "max_abs_diff": 0.0002949833869934082
+ }
+}

model_fp16.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f45b4b79dd08b9e0237bd7420b735659b043edb25b572949dbb4ae5ce3fe110b
+size 56343975