Feat: AI refactor pipeline

.github/workflows/release-build.yml

@@ -147,6 +147,7 @@ jobs:
             --icon assets/app_icon.ico `
             --paths src `
             --hidden-import core.build_info `
+            --hidden-import mediapipe `
             --add-data "src/ui/dark_theme.qss;." `
             --add-data "assets/app_icon.ico;." `
             --add-data "assets/app_icon.png;." `
@@ -161,6 +162,8 @@ jobs:
             --hidden-import torch `
             --hidden-import sklearn `
             --hidden-import sentence_transformers `
+            --collect-data mediapipe `
+            --collect-data pyiqa `
             --add-data "models;models" `
             --runtime-hook runtime_hook.py `
             src/main.py
@@ -214,6 +217,7 @@ jobs:
             --icon assets/photosort.icns \
             --paths src \
             --hidden-import core.build_info \
+            --hidden-import mediapipe \
             --add-data src/ui/dark_theme.qss:. \
             --add-data assets/app_icon.ico:. \
             --add-data assets/app_icon.png:. \
@@ -236,6 +240,8 @@ jobs:
             --hidden-import torch \
             --hidden-import sklearn \
             --hidden-import sentence_transformers \
+            --collect-data mediapipe \
+            --collect-data pyiqa \
             --add-data models:models \
             src/main.py
 

README.md

@@ -19,7 +19,7 @@ PhotoSort is a powerful desktop application focused on speed designed to streaml
   * **Fast Processing**: Intensive operations (scanning, thumbnailing, analysis) run once in batch to ensure fast image scrolling.
   * **Optimized Image Handling**: Supports a wide range of formats, including various RAW types, with efficient caching.
   * **Intelligent Image Rotation**: Smart rotation system that automatically tries lossless metadata rotation first, with optional fallback to pixel rotation when needed.
-  * **AI Best-Shot Ranking**: Compare stacks with either the bundled multi-model pipeline or an OpenAI-compatible vision model (e.g. Qwen3-VL).
+* **AI Best-Shot Ranking**: Compare stacks with either the bundled MUSIQ/MANIQA/LIQE pipeline or an OpenAI-compatible vision model (e.g. Qwen3-VL).
   * **AI Star Ratings**: Ask the configured AI engine to score individual photos with 1–5 stars.
 
 - **Update Notifications**: Automatically checks for new releases and notifies users when updates are available, with direct download links.
@@ -102,25 +102,15 @@ The application will automatically detect and load the model when you use the ro
 
 ### AI Best Shot Ranking & Engines
 
-PhotoSort can rank similar shots and assign AI ratings using either a local
-multi-model pipeline or an OpenAI-compatible vision model; switch engines in
+PhotoSort can rank similar shots and assign AI ratings using either the local
+MUSIQ/MANIQA/LIQE pipeline or an OpenAI-compatible vision model; switch engines in
 **Preferences → AI Rating Engine** (`F10`). Settings persist between sessions.
 
 **Local pipeline (default)**  
-Runs entirely offline with three Hugging Face checkpoints:
-BlazeFace face detector (`qualcomm/MediaPipe-Face-Detection`), eye-state classifier
-(`MichalMlodawski/open-closed-eye-classification-mobilev2`), and the aesthetic predictor
-(`shunk031/aesthetics-predictor-v2-sac-logos-ava1-l14-linearMSE`). Place each bundle
-under `models/` and choose **Local Pipeline** in preferences.
-
-Required downloads (install into `models/`):
-
-1. **Face detector** – [`qualcomm/MediaPipe-Face-Detection`](https://huggingface.co/qualcomm/MediaPipe-Face-Detection)  
-   Extract `model.onnx` to `models/job_*/model.onnx` (or e.g. `models/MediaPipe-Face-Detection_FaceDetector_float/model.onnx`).
-2. **Eye-state classifier** – [`MichalMlodawski/open-closed-eye-classification-mobilev2`](https://huggingface.co/MichalMlodawski/open-closed-eye-classification-mobilev2)  
-   Copy all files into `models/open-closed-eye-classification-mobilev2/`.
-3. **Aesthetic predictor** – [`shunk031/aesthetics-predictor-v2-sac-logos-ava1-l14-linearMSE`](https://huggingface.co/shunk031/aesthetics-predictor-v2-sac-logos-ava1-l14-linearMSE)  
-   Copy all files into `models/aesthetic_predictor/` (includes the CLIP backbone plus regression head).
+Runs entirely offline by blending three state-of-the-art no-reference IQA models:
+**MUSIQ**, **MANIQA**, and **LIQE**. These metrics are loaded
+through [`pyiqa`](https://github.com/chaofengc/IQA-PyTorch); no manual model
+downloads are required.
 
 **LLM engine**  
 Connect PhotoSort to any OpenAI-compatible endpoint that accepts images
@@ -138,8 +128,7 @@ the API key blank.
 - **AI star ratings**: To score every visible image, run **View → AI Rate Images**
   (`Ctrl+A`). The ratings are stored in your XMP sidecars/metadata cache so
   they survive reloads, and you can filter the library using the standard rating
-  controls. (Detailed breakdowns from the AI response are kept internally for future
-  UI integrations.)
+  controls.
 
 ### Exporting Logs
 

assets/keyboard-layout.html

@@ -785,7 +785,7 @@ <h3 style="float: left;" class="editable">Photosort Shortcuts (ctrl/cmd)</h3>
             <div data-key="KeyX" class="key">X<span class="userText"></span></div>
             <div data-key="KeyC" class="key">C<span class="userText"></span></div>
             <div data-key="KeyV" class="key">V<span class="userText"></span></div>
-            <div data-key="KeyB" class="key key--colour--navy">B<span class="userText">Analyze Best Shots (Ctrl+B)</span></div>
+            <div data-key="KeyB" class="key key--colour--navy">B<span class="userText">Analyze Best Shots</span></div>
             <div data-key="KeyN" class="key">N<span class="userText"></span></div>
             <div data-key="KeyM" class="key">M<span class="userText"></span></div>
             <div data-key="Comma" class="key">, &lt;<span class="userText"></span></div>
@@ -946,7 +946,7 @@ <h3 style="float: left;" class="editable">Photosort Shortcuts (shift)</h3>
             <div data-key="KeyX" class="key">X<span class="userText"></span></div>
             <div data-key="KeyC" class="key">C<span class="userText"></span></div>
             <div data-key="KeyV" class="key">V<span class="userText"></span></div>
-            <div data-key="KeyB" class="key key--colour--navy">B<span class="userText">Analyze Best Shots (Selected)</span></div>
+            <div data-key="KeyB" class="key key--colour--navy">B<span class="userText">Analyze Best Shots (Images Selected)</span></div>
             <div data-key="KeyN" class="key key--colour--orange">N<span class="userText">Decline Rotation Suggestions</span></div>
             <div data-key="KeyM" class="key">M<span class="userText"></span></div>
             <div data-key="Comma" class="key">, &lt;<span class="userText"></span></div>

requirements-cuda.txt

@@ -11,4 +11,6 @@ opencv-python
 pyexiv2
 piexif
 onnxruntime-gpu
-torchvision
+torchvision
+pyiqa
+mediapipe

requirements.txt

@@ -14,3 +14,5 @@ pyexiv2
 piexif
 onnxruntime
 torchvision
+pyiqa
+mediapipe

src/core/ai/__init__.py

@@ -1,9 +1,7 @@
-"""
-AI helper utilities for advanced ranking/scoring pipelines.
+"""AI helper utilities for best-shot ranking and scoring."""
 
-Currently exposes the experimental best-photo selector which chains together
-multiple pre-trained models (face detection, eye-state classification, and
-image quality scoring) to rank similar shots.
-"""
-
-from .best_photo_selector import BestPhotoSelector, BestShotResult  # noqa: F401
+from .best_photo_selector import (  # noqa: F401
+    BestPhotoSelector,
+    BestShotResult,
+    MetricSpec,
+)