AIbase
EN
Home
Tecoa4 Convnext Base W Laion2b S13b B82k Augreg
T

Tecoa4 Convnext Base W Laion2b S13b B82k Augreg

Developed by chs20
A robust perceptual metrics model based on CLIP, adversarially fine-tuned to enhance performance in perceptual similarity tasks.
Text-to-Image
Safetensors
Open Source License:MIT#Adversarial Robustness CLIP#Zero-shot Image Classification#Perceptual Similarity Metrics
Downloads 24
Release Time : 8/14/2024

Model Overview

This model is a vision-language model based on the CLIP architecture, adversarially fine-tuned on ImageNet using the TeCoA method to improve robustness in perceptual similarity tasks.

Model Features

Adversarial Fine-tuning
Adversarially fine-tuned on ImageNet using the TeCoA method, enhancing the model's robustness under adversarial attacks.
High Robustness
Demonstrates strong performance retention under adversarial attacks in both L-infinity and L2 norms.
Excellent Perceptual Similarity Performance
Outstanding performance on perceptual similarity tasks in the NIGHTS dataset, achieving 92.3 on clean data.

Model Capabilities

Zero-shot Image Classification
Perceptual Similarity Metrics
Adversarial Robustness Evaluation

Use Cases

Computer Vision
Image Classification
Used for zero-shot image classification tasks, capable of classifying images without specific training.
Performance on clean data: 92.3.
Perceptual Similarity Metrics
Used to evaluate perceptual similarity between images, suitable for tasks like image quality assessment and image retrieval.
Excellent performance on the NIGHTS dataset.
Security and Robustness
Adversarial Robustness Evaluation
Used to evaluate the model's performance retention under adversarial attacks.
Performance under L-infinity norm: 81.9, under L2 norm: 78.5.
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
EnglishįŽ€äŊ“中文įšéĢ”中文ãĢãģんご
Š 2025AIbase