🤗 Simple Aesthetics Predictor

CLIP-based aesthetics predictor inspired by the interface of 🤗 huggingface transformers. This library provides a simple wrapper that can load the predictor using the from_pretrained method.

Install

pip install simple-aesthetics-predictor

How to Use

import requests
import torch
from PIL import Image
from transformers import CLIPProcessor

from aesthetics_predictor import AestheticsPredictorV1

#
# Load the aesthetics predictor
#
model_id = "shunk031/aesthetics-predictor-v1-vit-large-patch14"

model = AestheticsPredictorV1.from_pretrained(model_id)
processor = CLIPProcessor.from_pretrained(model_id)

#
# Download sample image
#
url = "https://github.com/shunk031/simple-aesthetics-predictor/blob/master/assets/a-photo-of-an-astronaut-riding-a-horse.png?raw=true"
image = Image.open(requests.get(url, stream=True).raw)

#
# Preprocess the image
#
inputs = processor(images=image, return_tensor="pt")

#
# Inference for the image
#
with torch.no_grad():
    outputs = model(**inputs)
prediction = outputs.logits

print(f"Aesthetics score: {prediction}")

The Predictors found in 🤗 Huggingface Hub

• 🤗 aesthetics-predictor-v1
• 🤗 aesthetics-predictor-v2

Acknowledgements

• LAION-AI/aesthetic-predictor: A linear estimator on top of clip to predict the aesthetic quality of pictures https://github.com/LAION-AI/aesthetic-predictor
• christophschuhmann/improved-aesthetic-predictor: CLIP+MLP Aesthetic Score Predictor https://github.com/christophschuhmann/improved-aesthetic-predictor