Auramask-ensemble-poprocket Open-source Image Model - Supports Image Adversarial, Aesthetics and Quality Enhancement Processing

Auramask Ensemble Poprocket

Developed by logasja

This model uses an improved vnet architecture for image-to-image processing, supporting tasks such as adversarial, aesthetic, and quality enhancement

Image Generation Open Source License:Gpl-3.0 #Facial beautification filter #Adversarial image generation #Multi-feature fusion loss

Downloads 15

Release Time : 3/2/2025

Model Overview

A deep learning model based on an improved vnet architecture, primarily used for image processing tasks including image enhancement, filter effects, and facial feature preservation. The model combines the feature extraction capabilities of ArcFace and VGGFace, achieving high-quality output through multi-loss function optimization.

Model Features

Multi-feature fusion

Combines the feature extraction capabilities of ArcFace and VGGFace to maintain facial feature consistency

Hybrid loss function

Uses multiple metrics such as TopIQ-FR, ArcFace cosine distance, and VGGFace2 cosine distance for joint optimization

High-quality output

Optimized through aesthetic and quality evaluation metrics to generate high-quality image output

Adversarial enhancement

Supports adversarial processing to enhance specific image features

Model Capabilities

Image enhancement

Filter effect application

Facial feature preservation

Image quality improvement

Adversarial processing

Use Cases

Image processing

Portrait beautification

Enhances aesthetics and improves the quality of portraits

Improves overall image quality while preserving facial features

Filter effects

Applies specific style filter effects

Generates images with artistic styles

Security and privacy

Adversarial processing

Processes images adversarially to protect privacy

Alters key features while maintaining visual quality

🚀 Keras Image-to-Image Model

This is an image-to-image model based on Keras, which uses a modified vnet for 2D input/output and is suitable for various image processing tasks such as adversarial, aesthetic, and quality filtering.

🚀 Quick Start

The model uses a modified vnet for 2D input/output. You can find the implementation here.

✨ Features

Multiple Metrics: Utilizes metrics like TopIQ-FR, ArcFace Cosine Distance, and VGGFace2 Cosine Distance.
Diverse Datasets: Trained on the logasja/FDF dataset.
Specific Tags: Tags include adversarial, aesthetic, quality, and filter, indicating its application scenarios.

📦 Installation

No specific installation steps are provided in the original document.

💻 Usage Examples

No code examples are provided in the original document.

📚 Documentation

Model Configuration

The model has the following configuration:

{
    "activation": "ReLU",
    "batch_norm": false,
    "filter_num": [
        128,
        256,
        512,
        1024,
        1024
    ],
    "n_labels": 3,
    "output_activation": "tanh",
    "pool": false,
    "res_num_ini": 1,
    "res_num_max": 3,
    "unpool": false
}

Training Configuration

The training configuration is as follows:

{
    "alpha": 0.0001,
    "batch": 32,
    "epochs": 500,
    "epsilon": 1,
    "input": "(256, 256)",
    "losses": {
        "FEAT_ArcFace": {
            "d": "cosine_similarity",
            "f": "ArcFace",
            "name": "FEAT_ArcFace",
            "reduction": "sum_over_batch_size",
            "threshold": 0.68,
            "weight": 0.05
        },
        "FEAT_VGG-Face": {
            "d": "cosine_similarity",
            "f": "VGG-Face",
            "name": "FEAT_VGG-Face",
            "reduction": "sum_over_batch_size",
            "threshold": 0.68,
            "weight": 0.05
        },
        "IQASSIMC": {
            "lower_better": false,
            "name": "IQASSIMC",
            "reduction": "sum_over_batch_size",
            "weight": 0.5
        },
        "TopIQ": {
            "full_ref": true,
            "lower_better": false,
            "name": "TopIQ",
            "reduction": "sum_over_batch_size",
            "score_range": "~0, ~1",
            "weight": 0.5
        }
    },
    "mixed_precision": true,
    "optimizer": {
        "amsgrad": false,
        "beta_1": 0.9,
        "beta_2": 0.999,
        "clipnorm": null,
        "clipvalue": null,
        "ema_momentum": 0.99,
        "ema_overwrite_frequency": null,
        "epsilon": 1e-07,
        "global_clipnorm": null,
        "gradient_accumulation_steps": null,
        "learning_rate": 9.999999747378752e-05,
        "loss_scale_factor": null,
        "name": "adamw",
        "use_ema": false,
        "weight_decay": 0.004
    },
    "seed": "BIIIIIGSTRETCH",
    "testing": 0.01,
    "training": 0.99
}

Model Architecture Plot

🔧 Technical Details

The model uses a modified vnet for 2D input/output. The specific configuration details are shown in the above JSON code.

📄 License

This model is licensed under the GPL-3.0 license.

Information Table

Property	Details
Library Name	keras
Widgets	- text: input, output: input.png - text: target, output: target.png - text: output, output: output.png
Metrics	TopIQ-FR, ArcFace Cosine Distance, VGGFace2 Cosine Distance
Datasets	logasja/FDF
Pipeline Tag	image-to-image
Tags	adversarial, aesthetic, quality, filter
Base Model	vnet, logasja/ArcFace, logasja/VGGFace
License	gpl - 3.0

Training logs can be found here.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご