ML 모델
BurstPick은 6개 카테고리에 걸쳐 20개 이상의 온디바이스 모델 제공자를 지원합니다. 기본 모델은 별도 설정 없이 동작하고, 더 큰 모델 패밀리는 필요할 때 다운로드하며 Pro가 필요합니다. 사진 분석을 위해 원본을 업로드하지 않습니다.
이미지 품질 평가
선명도, 노이즈, 노출, 지각적 선명함 등 기술적 품질을 기준으로 각 사진을 평가합니다. 버스트 클러스터 내에서 사진 순위를 매기는 데 사용됩니다.
Heuristic (Laplacian + Luma)
번들Instant scoring using Accelerate vDSP/vImage. Measures sharpness, exposure, noise, and eye closure. Cannot judge composition or semantic quality. Best for fast initial culling passes.
TOPIQ NR
No-reference IQA using ResNet50 backbone. General-purpose technical quality scoring.
MUSIQ (KonIQ)
Multi-scale transformer trained on KonIQ-10k real-world distortions. Used for natural-photo quality assessment.
MANIQA
NTIRE 2022 IQA Challenge winner. Multi-dimension attention captures fine perceptual differences. Larger than the other bundled IQA options.
NIMA (MobileNet)
Neural Image Assessment trained on AVA (250K aesthetic ratings). Outputs a 10-class probability distribution with a compact MobileNet backbone.
미적 점수
구도, 색상 조화, 시각적 균형을 기준으로 미적 매력을 평가합니다. 대규모 선호도 데이터로 학습했습니다.
LAION Aesthetic v1
번들Lightweight linear probe on CLIP embeddings. Trained on LAION aesthetic ratings.
ViT-B/16 Aesthetic
Standalone ViT-B/16 fine-tuned on AVA dataset (250K human aesthetic ratings). Runs independently of CLIP.
이미지 임베딩
사진을 벡터로 변환해 유사도 기반 클러스터링에 사용합니다. 버스트 그룹화와 중복 감지에 쓰입니다.
Apple Vision FeaturePrint
번들Built into macOS. No download required. Used for general-purpose scene similarity.
DINOv2 ViT-S/14
Self-supervised features from Meta's LVD-142M training. Used for visual similarity and scene structure.
CLIP ViT-B/32
Multimodal embeddings for grouping photos by content. Required by the LAION Aesthetic scorer.
얼굴 임베딩
얼굴 ID 벡터를 만들어 인물별로 클러스터링합니다.
EdgeFace-XS
Fastest option — lightweight 4 MB download. Good face grouping for most photos (LFW 99.73%). Best when speed is the priority.
EdgeFace-S
Good balance of speed and accuracy (LFW 99.82%, IJB-B 94.38%). Small download. Handles varied lighting well. Recommended balanced choice.
AdaFace IR-18
번들Strong on low-quality and challenging face crops via adaptive margin (CVPR 2022). LFW 99.82%. Good mid-tier choice.
AdaFace IR-50
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.
AuraFace v1
Large ResNet-100 backbone with permissive Apache 2.0 license. Choose mainly for licensing requirements.
GhostFaceNets
SOTA 2025 lightweight face recognition model. High performance with minimal computational overhead.
비전 언어 모델 (VLM)
자연어로 사진 내용을 파악합니다. 장면 설명이나 품질 판단 등 숫자 점수로는 안 되는 분석을 합니다.
Heuristic Estimate
번들Built-in fallback using heuristic image analysis (sharpness, exposure, noise, faces). No download required. Replace with a real VLM for improved results.
SmolVLM2 256M
Smallest VLM — fastest inference with minimal memory. Basic scene recognition and quality commentary. Best for quick screening on constrained hardware.
SmolVLM2 2.2B
Full-size SmolVLM for scene description and quality reasoning. Slower than the 256M variant.
FastVLM 0.5B
Apple FastVLM with FastViTHD hybrid encoder. Tuned for on-device scene recognition.
FastVLM 1.5B
Largest VLM option in the list. Used for detailed scene descriptions and quality reasoning.
이미지 분류
장면, 객체 레이블을 태깅합니다. Apple Vision 프레임워크 기반.
Apple Vision Classification
번들Built-in macOS image classification using VNClassifyImageRequest. Fast, no download required. Provides scene and object tags for filtering.