Skip to content

Latest commit

 

History

History
40 lines (30 loc) · 3.41 KB

results.md

File metadata and controls

40 lines (30 loc) · 3.41 KB
Raw

Default Diffusion Parameters:

Inference Steps Guidance Strength
20 7.5 1.0

Classification Results

  1. Prompts Ablations

Prompt Classification Models CLIP Models Adversarially Trained Models Stylised ImageNet Models
"This is a picture of class name" 99.86% to 95.84% 82.23% to 82.92% 91.21% to 88.32% 75.42% to 74.67%
blip-2 Caption 99.86% to 92.13% 82.23% to 72.22% 91.21% to 83.27% 75.42% to 68.80%
"This is a photo of textures background" 99.86% to 88.96% 82.23% to 66.67% 91.21% to 75.76% 75.42% to 58.39%
"This is a photo of intricately rich textures background" 99.86% to 86.57% 82.23% to 65.69% 91.21% to 72.90% 75.42% to 52.54%
"This is a photo of colorful textures background" 99.86% to 83.84% 82.23% to 60.78% 91.21% to 65.35% 75.42% to 44.13%
"This is a photo of distorted textures background" 99.86% to 80.23% 82.23% to 59.79% 91.21% to 63.29% 75.42% to 41.90%
"class Name" + against a vivid red background 99.86% to 90.98% 82.23% to 70.07% 91.21% to 79.40% 75.42% to 62.97%
"class Name" + against a vivid green background 99.86% to 91.69% 82.23% to 71.38% 91.21% to 82.27% 75.42% to 65.28%
"class Name" + against a vivid blue background 99.86% to 91.43% 82.23% to 71.05% 91.21% to 82.19% 75.42% to 63.83%
"class Name" + against a vivid colorful background 99.86% to 90.96% 82.23% to 69.79% 91.21% to 79.82% 75.42% to 61.76%
This is a photo of a vivid red background 99.86% to 87.72% 82.23% to 64.85% 91.21% to 70.15% 75.42% to 54.19%
This is a photo of a vivid green background 99.86% to 87.94% 82.23% to 65.49% 91.21% to 74.24% 75.42% to 54.25%
This is a photo of a vivid blue background 99.86% to 87.81% 82.23% to 65.03% 91.21% to 72.01% 75.42% to 49.33%
This is a photo of a vivid colorful background 99.86% to 85.19% 82.23% to 61.18% 91.21% to 63.89% 75.42% to 45.08%

  1. Adversarial Attack

Attack Classification Models CLIP Models Adversarially Trained Models Stylised ImageNet Models
Latent Attack 99.86% to 35.1% 82.23% to 44.68% 91.21% to 30.64% 75.42% to 22.62%
Prompt Attack 99.86% to 37.17% 82.23% to 38.23% 91.21% to 30.86% 75.42% to 20.05
Ensemble Attack 99.86% to 21.65% 82.23% to 32.87% 91.21% to 18.6% 75.42% to 11.9%

BLIP-2 captions are used as prompts for the adversarial setting

Object Detection Results