ML 모델

Instant scoring using Accelerate vDSP/vImage. Measures sharpness, exposure, noise, and eye closure. Cannot judge composition or semantic quality. Best for fast initial culling passes.

No-reference IQA using ResNet50 backbone. Good general-purpose technical quality scores. Good balanced option for quality assessment.

Multi-scale transformer trained on KonIQ-10k real-world distortions. Strong on natural photos with better perceptual alignment than TOPIQ.

NTIRE 2022 IQA Challenge winner. Multi-dimension attention captures fine perceptual differences. Most accurate but largest in category.

Neural Image Assessment trained on AVA (250K aesthetic ratings). Outputs 10-class probability distribution. Compact MobileNet backbone — fastest aesthetic quality model.

Lightweight linear probe on CLIP embeddings — near-zero overhead if CLIP is loaded. Trained on LAION aesthetic ratings. Good default aesthetic scorer.

Standalone ViT-B/16 fine-tuned on AVA dataset (250K human aesthetic ratings). More nuanced aesthetic judgment than LAION probe. Independent of CLIP.

Built into macOS — zero download, instant availability. Good general-purpose scene similarity. Best for speed-first workflows.

State-of-the-art self-supervised features (Meta, LVD-142M). Excellent visual similarity and scene structure. Recommended balanced choice.

Rich semantic understanding from multimodal training. Groups photos by content meaning. Required by LAION Aesthetic scorer. Best for diverse libraries.

Fastest option — lightweight 4 MB download. Good face grouping for most photos (LFW 99.73%). Best when speed is the priority.

Good balance of speed and accuracy (LFW 99.82%, IJB-B 94.38%). Small download. Handles varied lighting well. Recommended balanced choice.

Strong on low-quality and challenging face crops via adaptive margin (CVPR 2022). LFW 99.82%. Good mid-tier choice.

Top-tier accuracy (LFW 99.82%, IJB-B 95.67%). Excels on difficult poses and low-quality crops. Best when face grouping precision is critical.

Large ResNet-100 backbone with permissive Apache 2.0 license. Choose mainly for licensing requirements.

SOTA 2025 lightweight face recognition model. High performance with minimal computational overhead.

Built-in fallback using heuristic image analysis (sharpness, exposure, noise, faces). No download required. Replace with a real VLM for improved results.

Smallest VLM — fastest inference with minimal memory. Basic scene recognition and quality commentary. Best for quick screening on constrained hardware.

Full-size SmolVLM with strong scene understanding and quality reasoning. More capable but slower than 256M variant.

Apple FastVLM with FastViTHD hybrid encoder. Optimized for on-device speed with solid scene recognition. Recommended balanced VLM choice.

Largest and most capable VLM. Deep scene understanding, nuanced quality reasoning, and detailed descriptions. Best when VLM quality is the top priority.

Built-in macOS image classification using VNClassifyImageRequest. Fast, no download required. Provides scene and object tags for filtering.