photo_judgers/docs/research-siglip-v2-5.md

SigLIP-Based Aesthetic Predictor V2.5

Model Info

• GitHub: https://github.com/discus0434/aesthetic-predictor-v2-5
• PyPI: aesthetic-predictor-v2-5
• License: MIT
• Updated: December 2024
• Improvement over V2: Uses SigLIP (Google's CLIP alternative) instead of OpenAI CLIP; better at illustrations and diverse image domains

Architecture

• Backbone: SigLIP (Google) — superior to OpenAI CLIP on many benchmarks
• Head: Linear layer on SigLIP embeddings
• Output: Single float score (0-10 scale)
• Score threshold: 5.5+ is considered "great" (vs. 6+ for V2)

Hardware Requirements

• SigLIP ViT: ~2-3 GB VRAM (slightly larger than CLIP ViT-L/14)
• Precision: Requires BF16 for best results
• Batch size: ~16-32 on 24 GB GPU

Python Usage

from aesthetic_predictor_v2_5 import convert_v2_5_from_siglip
from PIL import Image
import torch

model, preprocessor = convert_v2_5_from_siglip(low_cpu_mem_usage=True)
model = model.to(torch.bfloat16).cuda()

image = Image.open("photo.jpg").convert("RGB")
pixel_values = preprocessor(images=image, return_tensors="pt").pixel_values.to(torch.bfloat16).cuda()

with torch.inference_mode():
    score = model(pixel_values).logits.squeeze().float().cpu().numpy()

print(f"Aesthetics score: {score:.2f}")

What It's Good At

• Better than V2 on illustrations/art: SigLIP trained on larger, more diverse dataset
• Same simplicity: One-line HuggingFace-style API
• Same speed: Linear head on embeddings
• Slightly better generalization: SigLIP's training data is more diverse than OpenAI CLIP's

What It's Bad At

• Same fundamental limitations as V2: No technical quality detection, no blur/exposure detection, no genre awareness
• Still not trained on real photography: Same bias toward AI-generated aesthetic patterns
• BF16 requirement: Needs GPU with BF16 support (RTX 30-series+); not ideal for older hardware
• Community model: Not an official LAION release; maintained by a third party