DOCSEARCHbyAlgolia
Submit Docs

genimage-dataset/genimage

repository·main·Indexed 20 days ago

https://github.com/genimage-dataset/genimage

A repository providing implementations and scripts for GAN detection and image synthesis. It includes tools for AutoGAN, CycleGAN, and CNNDetection, as well as implementations for data-efficient architectures (DeiT, CaiT, ResMLP), F3Net for face forgery detection, and Global Texture Enhancement for fake face detection. The project covers training and evaluating classifiers using pixel-based and spectrum-based features to distinguish between real and synthesized images.

Tokens
155.6K
Snippets
447
Records
589
Agent score
69%

What's inside genimage

  1. Overview of Co-training 2L Submodels for Visual Recognition (Cosub)

    main

    The detector_codes/deit-main repository provides PyTorch evaluation code, training code, and pretrained models for several vision transformer and MLP-based projects, including the Cosub (Co-training 2L Submodels for Visual Recognition) training recipes. These recipes are designed to improve previous training strategies across various architectures.

    Supported architectures/projects in this directory include:

    • DeiT (Data-Efficient Image Transformers)
    • CaiT (Going deeper with Image Transformers)
    • ResMLP (Feedforward networks for image classification)
    • PatchConvnet (Augmenting Convolutional networks with attention-based aggregation)
    • 3Things (Three things everyone should know about Vision Transformers)
    • DeiT III (DeiT III: Revenge of the ViT)
    • Cosub (Co-training 2L Submodels for Visual Recognition)

  2. Overview of Data-Efficient architectures and training for Image classification

    main

    This repository provides PyTorch evaluation code, training code, and pretrained models for several research papers focused on data-efficient architectures for image classification.

    Supported architectures include:

    • DeiT: Data-Efficient Image Transformers (ICML 2021)
    • CaiT: Going deeper with Image Transformers (ICCV 2021)
    • ResMLP: Feedforward networks for image classification (TPAMI 2022)
    • PatchConvnet: Augmenting Convolutional networks with attention-based aggregation
    • 3Things: Three things everyone should know about Vision Transformers (ECCV 2022)
    • DeiT III: Revenge of the ViT (ECCV 2022)
    • Cosub: Co-training 2L Submodels for Visual Recognition (CVPR 2023)

  3. Overview of F3Net for Face Forgery Detection

    main

    F3Net is a face forgery detection implementation based on the ECCV 2020 paper "Thinking in frequency: Face forgery detection by mining frequency-aware clues". It builds upon the XceptionNet architecture and utilizes three core components to mine frequency-aware clues for detecting manipulated faces:

    1. MixBlock: A Two-stream Collaborative Framework implemented at block 7 and block 12.
    2. FAD: Frequency-aware Decomposition.
    3. LFS: Local Frequency Statistics.

  4. Overview of Guided Diffusion models

    main

    This repository contains diffusion models and noised image classifiers based on the paper "Diffusion Models Beat GANs on Image Synthesis".

    Available model types include:

    • Noisy ImageNet classifiers: Available at resolutions 64x64, 128x128, 256x256, and 512x512.
    • Class-unconditional ImageNet diffusion models: Resolution 256x256.
    • Class-conditional ImageNet diffusion models: Available at 64x64, 128x128, 256x256, and 512x512 resolutions.
    • Class-conditional ImageNet upsampling diffusion models: Performs upsampling from 64x64 to 256x256 or 128x128 to 512x512.
    • LSUN diffusion models: Trained on three specific classes (cat, horse, bedroom) at 256x256 resolution.

  6. Overview of Stable Diffusion v1

    main
    Stable Diffusion v1 is a diffusion-based text-to-image generation model. It is a Latent Diffusion Model that utilizes a fixed, pretrained CLIP ViT-L/14 text encoder to generate or modify images based on English text prompts. The model operates by encoding images into a latent space using an autoencoder with a relative downsampling factor of 8.

  7. Overview of Swin Transformer implementations

    main

    The Swin Transformer repository provides the official implementation of the 'Swin Transformer: Hierarchical Vision Transformer using Shifted Windows' paper and its follow-ups. It serves as a general-purpose backbone for various computer vision tasks.

    Available task implementations include:

    • Image Classification: Core functionality included in this repository. See get_started.md for setup.
    • Object Detection and Instance Segmentation: Managed in a separate repository.
    • Semantic Segmentation: Managed in a separate repository.
    • Video Action Recognition: Managed in the Video Swin Transformer repository.
    • Self-Supervised Learning (SSL): Includes Masked Image Modeling (SimMIM) and Contrastive Learning (Transformer-SSL).
    • Mixture-of-Experts (MoE): Includes the Swin-MoE variant.
    • Feature Distillation: Includes models and code for improving fine-tuning performance.

  8. Understand the validation and benchmark datasets

    main

    This repository contains validation and benchmark results for models in the pytorch-image-models-0.6.12 collection.

    Validation Results

    Validation scores are provided for models with pretrained weights and ImageNet-1k heads. The results are categorized into several datasets:

    • ImageNet Validation (results-imagenet.csv): The standard 50,000 image ImageNet-1k validation set.
    • ImageNet-'Real Labels' (results-imagenet-real.csv): The standard validation set using reassessed labels to correct original annotation mistakes.
    • ImageNetV2 Matched Frequency (results-imagenetv2-matched-frequency.csv): A 10,000 image test set sampled to replicate the original ImageNet curation process.
    • ImageNet-Sketch (results-sketch.csv): 50,000 non-photographic images (sketches/doodles) covering 1000 classes.
    • ImageNet-Adversarial (results-imagenet-a.csv): 7,500 naturally occurring adversarial images covering 200 classes. (Note: Compare against results-imagenet-a-clean.csv for clean runs on the same 200 classes).
    • ImageNet-Rendition (results-imagenet-r.csv): 30,000 images representing renditions of 200 ImageNet classes. (Note: Compare against results-imagenet-r-clean.csv for clean runs on the same 200 classes).

    Benchmark and Metadata Files

    • Benchmarks: CSV files prefixed with model_benchmark contain inference performance numbers (e.g., on RTX 3090 with AMP).
    • Metadata: CSV files prefixed with model_metadata contain information regarding the source training, such as the pretraining dataset and technique (e.g., distillation, SSL, WSL).

  9. Explore available model architectures in timm

    main

    The timm (pytorch-image-models) library provides implementations of various computer vision architectures. This documentation segment lists several supported model families, including:

    • BiT: Big Transfer ResNetV2
    • CSPNet: Cross-Stage Partial Networks
    • DenseNet: Densely Connected Convolutional Networks
    • DLA: Deformable Lightweight Attention
    • DPN: Dual-Path Networks
    • EfficientNet/BYOB: GPU-Efficient Networks
    • HRNet: High-Resolution Networks
    • Inception family: Inception-V3, Inception-V4, and Inception-ResNet-V2
    • NASNet: Neural Architecture Search Networks
    • Xception: Modified Aligned version (based on the Encoder-Decoder with Atrous Separable Convolution paper)

  10. What is Noisy Student Training?

    main

    Noisy Student Training is a semi-supervised learning approach that extends self-training and distillation. It involves three main steps:

    1. Teacher Training: Train a teacher model on labeled images.
    2. Pseudo-labeling: Use the teacher to generate pseudo labels on unlabeled images.
    3. Student Training: Train a student model on a combination of labeled images and pseudo-labeled images.

    Key characteristics:

    • Scaling: The student model is typically equal-or-larger than the teacher.
    • Noise Injection: To force the student to learn harder from pseudo labels, noise is added during training via input noise (e.g., RandAugment) and model noise (e.g., dropout and stochastic depth).

  11. What is Swin Transformer?

    main

    Swin Transformer (standing for Shifted window) is a hierarchical Transformer designed to serve as a general-purpose backbone for computer vision.

    Its core innovation is a shifted windowing scheme that improves efficiency by limiting self-attention computation to non-overlapping local windows. This scheme also enables cross-window connections, allowing the model to capture both local and global context. It has demonstrated state-of-the-art performance on benchmarks like COCO object detection and ADE20K semantic segmentation.

  12. Limitations and biases of GLIDE (filtered)

    main

    Users should be aware of several technical and social limitations when using GLIDE (filtered):

    • Compositional Weakness: The model has difficulty binding attributes to objects and performing complex compositional tasks (e.g., composing multiple objects in complex ways).
    • Lack of Human Representation: Because the training dataset was filtered to remove images of people, the model has limited world knowledge regarding concepts involving humans.
    • Dataset Biases: The model exhibits biases, including:
      • Gendered biases (e.g., different outputs for 'toys for boys' vs 'toys for girls').
      • Cultural/Religious biases (e.g., gravitating toward churches for 'a religious place').
      • Western-centric biases.
      • Potential for generating hate symbols not explicitly covered by the training filters.
    • Prompt Complexity: The model may fail to produce realistic outputs for highly complex prompts or concepts underrepresented in its training data.