[Qwen2.5-VL-72B-Instruct] PluginInvokeError: ImagePromptMessageContent is not JSON serializable when processing uploaded image in workflow

[wanghq032]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:)
• Please do not modify this template :) and fill in all the required fields.

1. Is this request related to a challenge you're experiencing? Tell me about your story.

我试图在 Dify 工作流中使用 硅基流动（SiliconFlow） 的 Qwen/Qwen2.5-VL-72B-Instruct 模型来识别用户上传的发票图片内容。我已在工作流中配置了图像上传节点，并将图像输出正确连接到 SiliconFlow 节点的图像输入端口。然而，当我运行工作流时，系统返回以下错误：

text
[siliconflow] Error: req_id: 295833e6c5 PluginInvokeError: {"args":{"description":"[models] Error: Object of type ImagePromptMessageContent is not JSON serializable"},"error_type":"InvokeError","message":"[models] Error: Object of type ImagePromptMessageContent is not JSON serializable"}
这让人沮丧，因为 Qwen2.5-VL-72B-Instruct 是一个明确支持视觉的多模态模型，理论上应该能处理图像输入。我不确定是哪个环节的参数配置错误，或是图像连接方式有误。

2. Additional context or comments

模型：Qwen/Qwen2.5-VL-72B-Instruct

平台：Dify Cloud + SiliconFlow

任务类型：视觉理解（发票识别）

图像上传方式：通过 Dify 文件上传节点，输出连接到 SiliconFlow 节点的图像输入端口

错误发生节点：SiliconFlow 节点

期望行为：模型应正确接收图像并返回文字识别结果

相关文件信息：

json
{
"filename": "发票.jpg",
"mime_type": "image/jpeg",
"transfer_method": "local_file"
}
我想知道：

如何在 Dify 工作流中确保上传的图片已正确连接到 SiliconFlow 节点的图像输入端口？

为了让模型正确“看懂”图像，哪些参数必须设置正确？

这是否是一个已知的兼容性或序列化问题？

[dosubot[bot]]

Hi @wanghq032! I'm Dosu and I’m helping the dify team.

This is a known serialization issue in Dify when using multimodal models like Qwen2.5-VL-72B-Instruct with SiliconFlow. The error happens when ImagePromptMessageContent objects (which represent images in the workflow) are not properly converted to a JSON-serializable format before being sent to the plugin or model. Dify normally handles this with its jsonable_encoder, but the error can occur if image variables are included in the System Prompt, or if the plugin/integration skips this conversion step.

To ensure images are correctly connected and processed:

• Make sure your image upload node outputs are connected to the LLM node’s image input, and that the “vision” feature is enabled in the node settings.
• The model expects image input as either a valid external URL or a base64 string prefixed with data:image/TYPE;base64,. You can control this with the MULTIMODAL_SEND_IMAGE_FORMAT config—set it to base64 for inline images, or url if your images are accessible via a public/signed URL. If using url, the image must be externally reachable by the model backend. If using base64, ensure the string is complete and correctly formatted, including the prefix source.
• Avoid referencing image variables in the System Prompt; only use them in User Prompts. This is a temporary workaround for a known bug where Dify only serializes User Prompts, not System Prompts, leading to this error source.
• If you are using a custom plugin or integration, ensure that any ImagePromptMessageContent objects are passed through .model_dump(), .dict(), or Dify’s jsonable_encoder before serialization. This converts them into a JSON-serializable format source.
• If you recently updated the dify_plugin SDK or related plugins and started seeing this error, try pinning the SDK to version 0.0.1b73 as a temporary fix. Newer versions introduced stricter type validation that can cause this issue with multimodal inputs source.
• Watch for updates to the dify-plugin-sdks and related plugins, as maintainers are actively working on a permanent fix for multimodal prompt serialization.

If you still see the error, double-check that your image data is correctly formatted and that your workflow does not include image variables in the System Prompt. If you need to debug further, log the serialized prompt messages before they are sent to the plugin to ensure all image objects are properly converted to dicts.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}