Clarify PyFeat-Live code and model license scopes

LICENSE

@@ -22,15 +22,14 @@ SOFTWARE.
 
 ---
 
-Third-party model licenses
-===========================
-
-The MIT license above covers the Py-feat Live application code ONLY. It does
-NOT cover the facial-expression detection models that the app downloads and
-runs (via py-feat). Those pretrained models and their weights are licensed
-separately by their respective authors, and several carry NON-COMMERCIAL or
-research-only stipulations.
-
-You are responsible for reviewing and complying with the license of each model
-you use. See the py-feat documentation for the per-model licenses:
-https://py-feat.org/  (Models reference)
+Model and dataset notices
+=========================
+
+The MIT grant covers the original PyFeat-Live application code. Third-party
+libraries retain their own licenses. Downloaded detector and visualization
+weights, research datasets, and optional generator assets have separate
+terms; the application license does not grant their providers' rights.
+
+See LICENSE-MODELS.md for the application-specific notice and links to the
+classic-v1, multitask-v2, and dataset notices. Existing valid model license
+grants are not revoked by this clarification.

LICENSE-MODELS.md

@@ -0,0 +1,33 @@
+# PyFeat-Live licensing
+
+Reviewed October 8, 2026.
+
+## Application code
+
+The PyFeat-Live application source code is licensed under its
+[MIT License](LICENSE),
+including commercial use. Third-party libraries retain their own licenses.
+
+## Downloaded models and optional features
+
+The application downloads and runs Py-Feat models. Its MIT license does
+not replace the terms for those pretrained weights or their source data:
+
+- Classic detector configurations: [Py-Feat v1.0](https://github.com/cosanlab/py-feat/blob/main/LICENSES/PYFEAT-V1.md).
+- Multitask detector configurations: [Py-Feat v2.0](https://github.com/cosanlab/py-feat/blob/main/LICENSES/PYFEAT-V2.md).
+- Identity, gaze, pose, and visualization assets must be checked separately
+  for the chosen model and checkpoint. Disabling ArcFace alone does not
+  make the default multitask model commercially licensed.
+- Optional generator assets have their own model cards and upstream
+  dependencies. This review does not certify those separate generative
+  checkpoints for unrestricted use.
+
+The application wrapper neither removes nor adds to a provider's rights
+in the models. The [dataset register](https://github.com/cosanlab/py-feat/blob/main/LICENSES/DATASETS.md) records training and
+evaluation provenance and unresolved permissions. It does not require
+every application user to sign every dataset agreement merely to run
+inference; any such access requirement must come from the applicable model
+or provider terms.
+
+This notice does not grant dataset access, revoke existing valid licenses,
+or establish a blanket license over users' inference outputs.

README.md

@@ -115,10 +115,4 @@ If you run into installation issues with py-feat see [this issue](https://github
 
 ## License
 
-The Py-feat Live application code is released under the [MIT License](LICENSE).
-
-**The facial-expression detection models are not covered by MIT.** The app
-downloads and runs pretrained models (via py-feat) that are licensed
-separately by their authors, and **several carry non-commercial / research-only
-stipulations**. You are responsible for reviewing and complying with each
-model's license before use — see the [py-feat model reference](https://py-feat.org/).
+The PyFeat-Live application code is released under the [MIT License](LICENSE), including commercial use. Downloaded detector and visualization weights, third-party libraries, research datasets, and optional generator assets have separate terms. See [the application model notice](LICENSE-MODELS.md) and the [Py-Feat licensing guide](https://py-feat.org/pages/licensing/) for component-specific terms and unresolved permissions. The application's MIT license does not establish unrestricted use of every pretrained model.