Organize ONNX model export tooling

- Restructure the repo around model families, moving YOLO assets under yolo/
  and adding first-class Cutie and EfficientSAM export directories.
- Add Cutie scripts for exporting OpenCV-compatible quality tiers and probes for
  validating model loading/forward passes. Add EfficientSAM packaging for Tiny and
  Small seed-mask variants, including OpenCV validation and static Small export
  conversion.

Author: jonoomph

.gitignore

@@ -8,11 +8,23 @@ __pycache__/
 .vscode/
 
 # Generated model artifacts. Keep these local or publish via Git LFS/releases.
-models/pt/*.pt
-models/onnx/*.onnx
-models/text/*.ts
-models/**/*.tmp
-releases/
+yolo/models/pt/*.pt
+yolo/models/onnx/*.onnx
+yolo/models/text/*.ts
+yolo/models/**/*.tmp
+yolo/releases/
+
+# Generated EfficientSAM assets
+efficient-sam/models/
+efficient-sam/build/
+efficient-sam/releases/
+
+# Generated Cutie assets and local upstream checkout
+cutie/vendor/
+cutie/weights/
+cutie/models/
+cutie/build/
+cutie/releases/
 
 # OS/local noise
 .DS_Store

MODEL_ARTIFACTS.md MODELS.mdMODEL_ARTIFACTS.md renamed to MODELS.md

@@ -2,25 +2,31 @@
 
 Generated model files are intentionally ignored by Git:
 
-- `models/pt/*.pt`
-- `models/onnx/*.onnx`
-- `models/text/*.ts`
-- `releases/*.zip`
+- `yolo/models/pt/*.pt`
+- `yolo/models/onnx/*.onnx`
+- `yolo/models/text/*.ts`
+- `yolo/releases/*.zip`
+- `efficient-sam/models/*.onnx`
+- `efficient-sam/releases/*.zip`
+- `cutie/weights/*.pth`
+- `cutie/models/*.onnx`
+- `cutie/releases/*.zip`
 
 The ONNX files are reproducible with:
 
 ```bash
-python scripts/export_yolo_seg_onnx.py
+python yolo/scripts/export_yolo_seg_onnx.py
+python cutie/scripts/export_cutie_quality_tiers.py
 ```
 
 They are not committed by default because pretrained YOLO checkpoints and
-derived ONNX exports are licensed by Ultralytics, and some generated files are
-large enough to require Git LFS or GitHub Releases instead of normal Git
-storage.
+derived ONNX exports are licensed by Ultralytics, Cutie checkpoints and derived
+ONNX exports are governed by upstream Cutie terms, and some generated files are
+large enough to require Git LFS or GitHub Releases instead of normal Git storage.
 
 The exporter code in this repository is MIT licensed. That license does not
 change the license terms of Ultralytics checkpoints, MobileCLIP weights, or
-derived model artifacts.
+Cutie weights, or derived model artifacts.
 
 Supported outputs:
 
@@ -35,9 +41,9 @@ If you want to publish generated ONNX files with a GitHub repository, use Git
 LFS or attach them to a GitHub Release after reviewing Ultralytics licensing
 requirements for your intended use.
 
-Class labels are not duplicated per model. Shared label sets live in
-`models/labels/`, and `models/labels/manifest.json` maps each ONNX filename to
-the correct label file and output width.
+YOLO class labels are not duplicated per model. Shared label sets live in
+`yolo/models/labels/`, and `yolo/models/labels/manifest.json` maps each ONNX
+filename to the correct label file and output width.
 
 Release packages duplicate the required label file inside each zip as
 `classes.names`:
@@ -48,7 +54,12 @@ yolo26n-seg.zip
   classes.names
 ```
 
-The root `models.json` file is the consumer-facing release catalog. It stays
-small on purpose: friendly dropdown names, zip asset filenames, and release
-download information. The exporter refreshes zip checksums when release packages
-are generated.
+The `yolo/models.json` file is the YOLO consumer-facing release catalog. It
+stays small on purpose: friendly dropdown names, zip asset filenames, and
+release download information. The exporter refreshes zip checksums when release
+packages are generated.
+
+Cutie release packages are generated under `cutie/releases/` and contain the
+four ONNX slices needed by libopenshot's Cutie Object Mask backend. The Cutie
+exporter also writes `cutie/releases/cutie-models.json` with asset names,
+dimensions, and checksums.

NOTICE.md

@@ -3,8 +3,9 @@
 Copyright (c) 2026 OpenShot Studios, LLC
 
 This repository contains helper scripts and documentation for downloading
-official Ultralytics YOLO segmentation checkpoints and exporting them to ONNX
-for OpenCV DNN consumers.
+official Ultralytics YOLO segmentation checkpoints, EfficientSAM checkpoints,
+and Cutie video object segmentation checkpoints, then exporting them to ONNX for
+OpenCV DNN consumers.
 
 The repository code and documentation are licensed under the MIT License. See
 [LICENSE](LICENSE).
@@ -16,3 +17,17 @@ weights, derived ONNX exports, or using them in commercial products:
 - https://github.com/ultralytics/ultralytics
 - https://github.com/ultralytics/assets
 - https://www.ultralytics.com/license
+
+Cutie source code and pretrained model weights are provided by the Cutie
+authors. Review upstream Cutie licensing and release notes before
+redistributing weights or derived ONNX exports:
+
+- https://github.com/hkchengrex/Cutie
+- https://github.com/hkchengrex/Cutie/releases
+- https://github.com/hkchengrex/Cutie/blob/main/LICENSE
+
+EfficientSAM source code and pretrained model weights are provided by the
+EfficientSAM authors. Review upstream licensing and release notes before
+redistributing weights or derived ONNX exports:
+
+- https://github.com/yformer/EfficientSAM

README.md

@@ -1,137 +1,63 @@
-# YOLO ONNX Exports for OpenShot
+# OpenShot ONNX Exports
 
-This repo builds OpenCV-friendly ONNX exports of official Ultralytics YOLO
-segmentation models for use in OpenShot and libopenshot.
+This repository builds OpenCV-friendly ONNX model packages for OpenShot and
+libopenshot. The generated ONNX files and zip packages are static runtime
+assets: OpenShot should not need Python, PyTorch, or the upstream training
+frameworks after export.
 
-The generated models are static ONNX graphs: no Python, PyTorch, or Ultralytics
-runtime is needed after export. OpenShot/libopenshot can load them with OpenCV
-DNN and handle thresholds, NMS, labels, boxes, and masks in C++.
+## Model Families
 
-## Getting Started
+| Family | Purpose | Directory |
+| --- | --- | --- |
+| YOLO | Object detection and instance segmentation | [`yolo/`](yolo/) |
+| EfficientSAM | Prompted seed-mask generation for Object Mask | [`efficient-sam/`](efficient-sam/) |
+| Cutie | Video object mask propagation for Object Mask | [`cutie/`](cutie/) |
 
-```bash
-python3 -m venv .venv
-source .venv/bin/activate
+XMem remains in `experiments/` as historical scratch work and is not promoted as
+a supported release family.
 
-python -m pip install --upgrade pip
-python -m pip install --upgrade torch torchvision --index-url https://download.pytorch.org/whl/cpu
-python -m pip install --no-deps -r requirements.txt
+## Release Artifacts
 
-python scripts/export_yolo_seg_onnx.py
-```
+Generated model binaries are ignored by Git. Build them locally and upload the
+zip files as GitHub Release assets after reviewing upstream model licenses.
 
-Default output goes here:
+Common outputs:
 
 ```text
-models/pt/       downloaded official checkpoints
-models/onnx/     exported ONNX models
-models/labels/   shared label files and manifest
-releases/        generated GitHub Release zip assets
-models.json      friendly release/download catalog for consumers
-```
-
-Large model binaries and generated release zips are ignored by Git. `models.json`
-is the file apps and scripts should read: it lists the friendly model names,
-release zip assets, and download base URL. During release packaging, the exporter
-refreshes each generated zip checksum.
-
-## Models
-
-Default export:
-
-```bash
-python scripts/export_yolo_seg_onnx.py
+yolo/releases/          YOLO zip packages and yolo/models.json
+efficient-sam/releases/ EfficientSAM zip packages
+cutie/releases/         Cutie zip packages and cutie-models.json
 ```
 
-Exports nano, small, and medium segmentation models for:
-
-- `yolov8` COCO 80 classes
-- `yolo11` COCO 80 classes
-- `yolo26` COCO 80 classes
+## Quick Commands
 
-Optional broad-vocabulary YOLOE export:
+YOLO:
 
 ```bash
-python scripts/export_yolo_seg_onnx.py --families yoloe26
+python yolo/scripts/export_yolo_seg_onnx.py
 ```
 
-YOLOE-26 uses the official prompt-free checkpoint vocabulary, embeds its 4,585
-labels with MobileCLIP during export, and writes a static ONNX model. MobileCLIP
-is not needed at runtime.
-
-Export everything:
+Cutie:
 
 ```bash
-python scripts/export_yolo_seg_onnx.py --families yolov8 yolo11 yolo26 yoloe26
+python cutie/scripts/export_cutie_quality_tiers.py
 ```
 
-## Release Packages
-
-The exporter writes one self-contained zip per model under `releases/` and
-updates `models.json`:
-
-```text
-releases/yolo26n-seg.zip
-  model.onnx
-  classes.names
-```
+EfficientSAM:
 
-The checked-in `models.json` intentionally stays small so OpenShot can populate
-a compact dropdown such as `YOLO26: Nano`, `YOLOv11: Nano`, and `YOLOv8: Nano`.
-
-## Details
-
-All exports use:
-
-- `640x640` static input
-- batch size `1`
-- ONNX opset `17`
-- `dynamic=False`
-- `nms=False`
-- `end2end=False`
-- `simplify=True`
-
-The ONNX output is raw YOLO segmentation output. Consumers are responsible for
-post-processing.
-
-Label files are shared instead of duplicated per model:
-
-```text
-models/labels/coco80.names
-models/labels/yoloe26-4585.names
-models/labels/manifest.json
-```
-
-Release zips duplicate the required labels as `classes.names` so every download
-is self-contained.
-
-## Examples
-
-```bash
-python scripts/export_yolo_seg_onnx.py --sizes n
-python scripts/export_yolo_seg_onnx.py --families yolo26
-python scripts/export_yolo_seg_onnx.py --families yoloe26 --sizes n
-python scripts/export_yolo_seg_onnx.py --force-download
-python scripts/export_yolo_seg_onnx.py --skip-opencv-validate
-```
+See [`efficient-sam/README.md`](efficient-sam/README.md). Its export/release
+tooling will live there as it is promoted from experiments.
 
 ## Links
 
-- [Ultralytics GitHub](https://github.com/ultralytics/ultralytics)
-- [Ultralytics model assets](https://github.com/ultralytics/assets/releases)
-- [YOLOv8 docs](https://docs.ultralytics.com/models/yolov8/)
-- [YOLO11 docs](https://docs.ultralytics.com/models/yolo11/)
-- [YOLO26 docs](https://docs.ultralytics.com/models/yolo26/)
-- [YOLOE docs](https://docs.ultralytics.com/models/yoloe/)
-- [Ultralytics license](https://www.ultralytics.com/license)
 - [OpenShot](https://www.openshot.org/)
 - [OpenShot GitHub](https://github.com/OpenShot/openshot-qt)
 - [libopenshot](https://github.com/OpenShot/libopenshot)
 - [OpenCV DNN](https://docs.opencv.org/4.x/d2/d58/tutorial_table_of_content_dnn.html)
 
 ## Project Notes
 
-This is an export utility, not an official Ultralytics or OpenShot project.
-See [MODEL_ARTIFACTS.md](MODEL_ARTIFACTS.md) and [NOTICE.md](NOTICE.md) for
+This is an export utility, not an official upstream model project. See
+[`MODELS.md`](MODELS.md) and [`NOTICE.md`](NOTICE.md) for
 artifact and licensing notes. The repository code is MIT licensed; upstream
-model weights and generated ONNX files remain subject to Ultralytics licensing.
+model weights and generated ONNX files remain subject to their upstream terms.