关于图片生成的一些问题

[zgojin]

在一开始生成图片时 使用了 接口 /v1/chat/completions 出现了 和 #76 一样的问题，一个错误的网址 ，后面根据 #70 修改了一下 使用 /v1/responses 成功获取到图片，但是不清楚如何将已有图片上传 实现 图生图 尝试了messages 列表包含 text 和 image_url 的方式 但好像并没有成功， 能否在redeme中简单的描述一下接口 或者 写一个简单的demo

"Initially, when using the /v1/chat/completions endpoint to generate images, I encountered the same issue as #76 (an incorrect URL). Following the advice in #70, I switched to /v1/responses and successfully retrieved the images.
However, I am not sure how to upload an existing image to implement Image-to-Image (img2img). I tried including text and image_url in the messages list, but it didn't seem to work. Could you briefly describe the endpoint in the README or provide a simple demo?"

[luuquangvu]

You can check out the document here: OpenAI.

image_url should be either a URL or base64-encoded image data, such as data:image/png;base64,xxxxxx.

Here's an example:

curl --location 'http://localhost:8000/v1/responses' \
--header 'Authorization: Bearer xxxxxx' \
--header 'Content-Type: application/json' \
--data '{
    "model": "gemini-2.5-flash",
    "input": [
        {
            "type": "message",
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "Change her hair color to a maroon shade."
                },
                {
                    "type": "input_image",
                    "image_url": "https://i.pinimg.com/1200x/e0/21/86/e0218632f5ca4b4ea7adc2666a71ac69.jpg"
                }
            ]
        }
    ],
    "stream": false,
    "tool_choice": {
        "type": "image_generation"
    },
    "tools": [
        {
            "type": "image_generation"
        }
    ]
}'

Update: Ensure that a public image link remains active indefinitely.

[zgojin]

非常感谢，已经成功修改图片，以下是我过程中遇到的问题
由于我之前是使用的 /v1/chat/completions 接口 为了获取流式中的base64返回，使用正则进行了暴力匹配，因此最开始得到图片是我发送的原图 这是因为/v1/responses会在响应包的 input 字段中完整保留用户发送的请求数据，正则扫描时，优先匹配到了 JSON 顶部的 input 区域中的 Base64 数据。实际上需要的图片信息存在output_image中，图片为具体的返回结构
Thanks! . I have successfully modified the image. I would like to share the issue I encountered during the process:
Since I was using the /v1/chat/completions endpoint, I used a regex pattern to brute-force the Base64 data from the streaming response. As a result, the image I initially extracted turned out to be the original image I sent.
This happened because the /v1/responses payload retains the complete user request data within the input field. My regex matched the Base64 data in that input section (at the top of the JSON) first. I realized later that the actual image data I needed was located in output_image.