AAL Plus Image Quality Assessment

提供商LoliRimuru
分类image-quality-assessment
许可证apache-2.0
下载量0
星标0

简介

AAL Plus 是一款专注于图像质量客观评估的轻量化模型。与传统的视觉模型不同，它不负责生成图像，而是充当“评审员”，通过量化指标对图片的清晰度、噪点及伪影进行打分。对于经常使用 Stable Diffusion 或 Midjourney 的创作者来说，它可以替代主观的肉眼筛选，帮助开发者在自动化管线中快速剔除低质量生成图，或用于对比不同采样器、模型对画质的影响。该模型上手简单，适合集成到图像处理工作流中作为质量把关环节。

核心亮点

  • 客观量化画质，替代主观肉眼筛选图片
  • 高效识别图像伪影与噪点，把控生成质量
  • 轻量级部署，易于集成至 AI 绘画工作流
  • Apache-2.0 协议，支持商业化灵活部署

使用方法

安装依赖
# 安装 Hugging Face transformers
pip install transformers torch
SDK 使用
# 使用 transformers 加载模型
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("LoliRimuru/AAL-Plus_Image_Quality_Assessment")
tokenizer = AutoTokenizer.from_pretrained("LoliRimuru/AAL-Plus_Image_Quality_Assessment")

Hugging Face 下载

我们推荐使用命令行或者 Hugging Face Hub SDK 来进行模型的下载。

操作指引：在下载前，请先通过如下命令安装 huggingface_hub：

操作指引 
pip install -U huggingface_hub

命令行下载

下载完整模型库

下载完整模型库 
huggingface-cli download LoliRimuru/AAL-Plus_Image_Quality_Assessment

下载单个文件到指定本地文件夹（以下载 config.json 到当前路径下 ./dir 目录为例）

下载单个文件到指定本地文件夹（以下载 config.json 到当前路径下 ./dir 目录为例） 
huggingface-cli download LoliRimuru/AAL-Plus_Image_Quality_Assessment config.json --local-dir ./dir

更多命令行下载选项，可参见官方文档

SDK 下载

SDK 下载 
# 模型下载
from huggingface_hub import snapshot_download
model_dir = snapshot_download('LoliRimuru/AAL-Plus_Image_Quality_Assessment')

Git 下载

请确保 lfs 已经被正确安装

Git 下载 
git lfs install
git clone https://huggingface.co/LoliRimuru/AAL-Plus_Image_Quality_Assessment

如果您希望跳过 lfs 大文件下载，可以使用如下命令

跳过 LFS 
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/LoliRimuru/AAL-Plus_Image_Quality_Assessment

模型文件托管在 Hugging Face Hub，使用 HF CLI / SDK / Git 直接下载，不经过本站。

PyTorch / Transformers 使用

安装 Transformers

安装 Transformers 
pip install -U transformers torch

模型加载和推理

模型加载和推理 
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained('LoliRimuru/AAL-Plus_Image_Quality_Assessment')
tokenizer = AutoTokenizer.from_pretrained('LoliRimuru/AAL-Plus_Image_Quality_Assessment')

完整文档

来源: HuggingFace

---
license: apache-2.0
tags:

  • image-classification

  • qualits-assessment

  • compression-artifacts

  • art

pipeline_tag: image-classification
---

AAL-Plus Image Quality Assessment

A lightweight model that predicts compression quality levels and detects artifacts across multiple image formats (JPEG, WebP, AVIF, JXL). Trained to identify subtle compression artifacts that indicate the quality level used during image encoding.

📊 Model Performance

Based on validation results from the final training epochs:

| Metric | Value |
|--------|-------|
| Overall Validation Accuracy | 97.1% |
| Overall Validation Loss | 0.0006 |
| Training Accuracy | 95.6% |

Per-Format Accuracy

| Format | Validation Acc | Training Acc | Quality Range |
|--------|----------------|--------------|---------------|
| JPEG | 99.4% | 98.9% | 0-100 |
| WebP | 97.0% | 95.5% | 0-100 |
| AVIF | 97.1% | 95.3% | 0-100 |
| JXL | 94.8% | 92.6% | 0-100 |

*Accuracy measured as predictions within ±5% range of actual quality values*

🎯 Key Features

  • Multi-format Support: Detects artifacts in JPEG, WebP, AVIF, and JXL formats
  • Lightweight Architecture: Only ~2M parameters (~8MB model size)
  • High Precision: 99.4% on JPEG
  • Fast Inference: Optimized for real-time processing even on very weak HW using either CPU or GPU
  • Format-agnostic: Single model handles all compression types

Limitations

1. Resolution Dependency: Model trained on 512x512 crops; very small images may yield reduced accuracy
2. Quality Range: Model assumes standard quality ranges; non-standard encoders may produce different results
3. Image Resizing: Trying to detect compression artifacs in images which were compressed and than converted to lossless format and resized may fail or be much less accurate as these artifacts are washed by the resizing algorithms.

Environmental Impact

  • Training: ~12 GPU hours on RTX 5090
  • Model Size: 8MB (reduces storage and bandwidth costs)
查看官方来源