achiru/DiffSynth-Studio-ImageMetrics

图像质量评估模型

本仓库包含了接入 DiffSynth-Studio 的多款主流图像质量评估模型权重。支持图文语义对齐、人类视觉偏好、纯图像美学以及数据集分布等多个维度的评测。

评估效果效果展示

prompt: A cat is sitting on a stone.

评估指标
Pickscore	22.958	23.321
ImageReward	1.419	1.786
HPSv2	30.169	30.528
HPSv3	12.287	12.969
CLIP Score	40.271	39.065
Aesthetic	5.096	5.848
UnifiedReward	'alignment': 4.5, 'coherence': 4.0, 'style': 3.5	'alignment': 4.0, 'coherence': 4.0, 'style': 3.5

---

指标总览

指标名称	输入要求	输出结果
PickScore	prompt + PIL 图像	人类视觉偏好分数
ImageReward	prompt + PIL 图像	人类视觉偏好分数
HPSv2	prompt + PIL 图像	人类视觉偏好分数
HPSv3	prompt + PIL 图像	人类视觉偏好分数
CLIP Score	prompt + PIL 图像	图文相似度
Aesthetic	PIL 图像	美学分数
UnifiedReward	prompt + PIL 图像	多维评分
FID	参考图目录 + 生成图目录	分布距离

快速使用

• 安装 DiffSynth-Studio

git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio
pip install -e .

• 示例 1：使用 PickScore 评估图文偏好

from diffsynth.metrics import PickScoreMetric, ModelConfig
from modelscope import dataset_snapshot_download
from PIL import Image

dataset_snapshot_download(
    "DiffSynth-Studio/diffsynth_example_dataset",
    allow_file_pattern="flux/FLUX.1-dev/*",
    local_dir="./data/diffsynth_example_dataset",
)
image = Image.open("data/diffsynth_example_dataset/flux/FLUX.1-dev/1.jpg").convert("RGB")
prompt = "a dog"
metric = PickScoreMetric.from_pretrained(
    model_config=ModelConfig(model_id="DiffSynth-Studio/ImageMetrics", origin_file_pattern="PickScore/model.safetensors"),
    device="cuda"
)
score = metric.compute(prompt, image)[0]
print(f"PickScore score:: {score:.3f}")

• 示例 2：使用 Aesthetic 评估纯美学质量

from diffsynth.metrics import AestheticMetric, ModelConfig
from modelscope import dataset_snapshot_download
from PIL import Image

dataset_snapshot_download(
    "DiffSynth-Studio/diffsynth_example_dataset",
    allow_file_pattern="flux/FLUX.1-dev/*",
    local_dir="./data/diffsynth_example_dataset",
)
image = Image.open("data/diffsynth_example_dataset/flux/FLUX.1-dev/1.jpg").convert("RGB")
metric = AestheticMetric.from_pretrained(
    model_config=ModelConfig(model_id="DiffSynth-Studio/ImageMetrics", origin_file_pattern="Aesthetic/model.safetensors"),
    device="cuda"
)
score = metric.compute(image)[0]
print(f"Aesthetic score: {score:.3f}")

关于所有指标的详细使用方法和说明，请参考 DiffSynth-Studio 的相关文档。