fix: load LAION V2 directly via transformers (trust_remote_code); drop stale wrapper; pin requirements

requirements.txt

@@ -1,13 +1,13 @@
 # Photo Judgers — Requirements
 # Install in conda env: pip install -r requirements.txt
 #
-# simple-aesthetics-predictor automatically installs torch + transformers.
-# We avoid pinning torchvision/torchaudio here to prevent version mismatch
-# with the torch that simple-aesthetics-predictor resolves. Pillow is needed
-# for image loading; tqdm for progress bars.
+# We load the LAION V2 model directly via transformers (trust_remote_code),
+# so we do NOT depend on the stale `simple-aesthetics-predictor` wrapper
+# (breaks with modern transformers). torch is required by transformers.
 
-simple-aesthetics-predictor
-transformers
+transformers>=4.40
+torch>=2.1
+torchvision
 Pillow
 tqdm
 numpy

src/scorers/laion.py

@@ -1,12 +1,23 @@
-"""LAION Aesthetic Predictor V2 scorer."""
+"""LAION Aesthetic Predictor V2 scorer.
+
+Loads the V2 model directly from HuggingFace via transformers'
+`trust_remote_code` mechanism. This avoids the `simple-aesthetics-predictor`
+wrapper package, which is stale (last release Dec 2024) and breaks with
+modern transformers (circular import in `dependency_versions_check`).
+"""
 
 from typing import Optional
+
+
 class LaionScorer:
     """LAION V2 aesthetic scorer — CLIP-based, 0-10 scale."""
 
     name = "LAION Aesthetic Predictor V2"
     description = "CLIP-based aesthetic scoring (0-10 scale)"
 
+    # V2 trained on SAC + LAION-Logos + AVA (current best per research)
+    MODEL_ID = "shunk031/aesthetics-predictor-v2-sac-logos-ava1-l14-linearMSE"
+
     def __init__(self, cache_dir: str = "models") -> None:
         self.model = None  # type: ignore[assignment]
         self.processor = None  # type: ignore[assignment]
@@ -17,24 +28,18 @@ class LaionScorer:
         """Load model weights (downloads ~2GB on first run)."""
         try:
             import torch  # noqa: F401
+            from transformers import AutoProcessor, AutoModel
 
-            # Package on PyPI is `simple-aesthetics-predictor`, but the
-            # importable module is `aesthetics_predictor` (no `simple_`).
-            from aesthetics_predictor import AestheticsPredictorV2Linear
-            from transformers import CLIPProcessor
-
-            print(f"  Loading LAION V2 model... (first run downloads ~2GB)")
+            print("  Loading LAION V2 model... (first run downloads ~2GB)")
             print(f"  Cache directory: {self._cache_dir}")
 
-            # V2 trained on SAC + LAION-Logos + AVA (current best per research)
-            model_id = "shunk031/aesthetics-predictor-v2-sac-logos-ava1-l14-linearMSE"
-
-            self.model = AestheticsPredictorV2Linear.from_pretrained(
-                model_id,
+            self.model = AutoModel.from_pretrained(
+                self.MODEL_ID,
                 cache_dir=self._cache_dir,
+                trust_remote_code=True,
             )
-            self.processor = CLIPProcessor.from_pretrained(
-                model_id,
+            self.processor = AutoProcessor.from_pretrained(
+                self.MODEL_ID,
                 cache_dir=self._cache_dir,
             )
             self.model.eval()