Stable Diffusion API POST请求中ADetailer参数配置问题求助

在调用Stable Diffusion API做图生图时，尝试在POST请求Payload中加入ADetailer参数，但多种写法都未被SD识别，生成的图片信息里也没有ADetailer相关记录。以下是试过的无效Payload代码：

payload = {
    "init_images": [encoded_string],
    "prompt": pr,
    "denoising_strength": 0.4,
    "negative_prompt": npr,
    "batch_size": 1,
    "steps": 22,
    "cfg_scale": 7,
    "width": np.array(img).shape[1],
    "height": np.array(img).shape[0],
    #"ADetailer": [{"model": "face_yolov8n.pt"}], 
    #"adetailer": [{"model": "face_yolov8n.pt"}], 
    #"ADetailer model": "face_yolov8n.pt",
    #"adetailer model": "face_yolov8n.pt",
    "alwayson_scripts": {
        "ADetailer": {
            "args": [{
                "model": "face_yolov8n.pt",
                #"denoising_strength": 0.2
                }
                ]
        },
        "adetailer": {
            "args": [{"model": "face_yolov8n.pt", 'enable': True}, ]
        },
        "adetailer":  {"model": "face_yolov8n.pt", 'enable': True},
        "controlnet": {
            "args": [
                {
                    "input_image": encoded_string,
                    "module": "canny",
                    #'preprocessor': 'canny',
                    "model": "control_v11p_sd15_canny_fp16 [b18e0966]",
                },
                {
                    "input_image": encoded_string,
                    #"module": "lineart_standard (from white bg & black line)",
                    'preprocessor': 'lineart_standard (from white bg & black line)',
                    "model": "control_v11p_sd15_lineart_fp16 [5c23b17d]",
                    #"weight": 0.5
                },
                {
                    #'preprocessor': 'reference_only',
                    "input_image": reference,
                    "module": "reference_only"
                }
            ]
        },
    }
}

错误原因分析 #

• Python字典中重复定义adetailer键，会导致前面的配置被最后一个覆盖，直接失效
• ADetailer脚本名称严格区分大小写，必须用"ADetailer"作为键名，小写的"adetailer"无法被识别
• args必须是数组格式，不能直接传递单个对象
• 部分配置缺少enable: true参数，默认可能未启用ADetailer功能

正确的Payload配置 #

payload = {
    "init_images": [encoded_string],
    "prompt": pr,
    "denoising_strength": 0.4,
    "negative_prompt": npr,
    "batch_size": 1,
    "steps": 22,
    "cfg_scale": 7,
    "width": np.array(img).shape[1],
    "height": np.array(img).shape[0],
    "alwayson_scripts": {
        # 仅保留一个ADetailer配置，键名严格使用大写开头
        "ADetailer": {
            "args": [
                {
                    "model": "face_yolov8n.pt",
                    "enable": True,  # 必须显式启用
                    "denoising_strength": 0.2,  # 可选：修复降噪强度
                    "confidence": 0.3,  # 可选：人脸检测置信度阈值
                    "dilate_erode": 4,  # 可选：Mask膨胀/腐蚀程度
                    "mask_blur": 4,  # 可选：Mask模糊度
                    "inpaint_only_masked": True,  # 可选：仅修复Mask覆盖区域
                    "inpaint_padding": 32  # 可选：修复区域扩展 padding 值
                }
            ]
        },
        "controlnet": {
            "args": [
                {
                    "input_image": encoded_string,
                    "module": "canny",
                    "model": "control_v11p_sd15_canny_fp16 [b18e0966]",
                },
                {
                    "input_image": encoded_string,
                    "preprocessor": "lineart_standard (from white bg & black line)",
                    "model": "control_v11p_sd15_lineart_fp16 [5c23b17d]",
                },
                {
                    "input_image": reference,
                    "module": "reference_only"
                }
            ]
        },
    }
}

配置完成后，生成的图片信息中会出现类似ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3...的记录，说明ADetailer已正常生效。

内容的提问来源于stack exchange，提问作者Roman Zakharov