我用火山方舟的Doubao-Seedream-5.0-lite,想接入sillytavern中扩展功...

Creado el: 25 de marzo de 2026

Respondido usando GPT-5.4 Thinking por Chat01

Pregunta

我用火山方舟的Doubao-Seedream-5.0-lite,想接入sillytavern中扩展功能的图像生成,该怎么办

Pensamiento

Considering OpenAI image source and reverse proxy

I need to confirm whether the OpenAI image settings allow reverse proxies along with manual models. Searching for updates on the custom endpoint might help, or referencing a comment from issue 1381. I'll make sure the response is clear and cite any claims from internet sources. Here's a possible structure for my response:

  • Best setup: Latest version of SillyTavern + built-in image generation + OpenAI source + Ark reverse proxy URL.
  • Explanation: Ark is OpenAI-compatible; ST image generation supports it, but the custom endpoint is only for Chat Completion, not image generation.

Guidance for setting up image generation

To set up image generation, use "Extensions" -> "Image Generation" rather than "API Connections' Custom (OpenAI-compatible)." Here are the steps:

  1. Prepare Ark: Enable model, generate API Key, and use Ark's console for the endpoint/model ID. Ark’s docs suggest simply switching base_url/model/api_key for OpenAI compatibility.
  2. Update to ST’s latest version, preferably 1.16.0, to avoid a Seedream URL bug in 1.15.0.
  3. In ST: Enable "Image Generation" under Extensions, choose OpenAI as the source, and enter your Ark API key, proxy URL (https://ark.cn-beijing.volces.com/api/v3), and endpoint. Apply Ark’s model ID from the console. You might use either URL or b64_json for response format, depending on support.

Thought for 3m 20s

Respuesta

最省事的接法是:

不要把火山方舟的 Seedream 当成「聊天模型的 Custom(OpenAI-compatible)」去接图像生成,
而是用 SillyTavern 内置的 Image Generation 扩展,把来源选成 OpenAI,再把它的请求转到火山方舟的 OpenAI 兼容地址。原因是:火山方舟官方明确说明它的模型调用尽量兼容 OpenAI API,只需要改 base_urlmodelapi_key 这几个关键项;而 SillyTavern 的图像生成扩展本身也把 OpenAI 列为受支持的图像来源。相反,SillyTavern 文档里那个 Custom (OpenAI-compatible) 主要是给 Chat Completion 用的,不是图像生成入口。(VolcEngine)

按这个顺序配:

  1. 先在火山方舟里确认三样东西:
    已开通 Doubao-Seedream-5.0-lite、拿到 API Key、并记下你在控制台里实际可调用的 模型 ID / Endpoint ID。火山方舟官方图像文档和示例教程都说明了 Seedream 5.0 lite 走的是图片生成 API,且这套接口是持续更新中的。(VolcEngine)

  2. 把 SillyTavern 升到新版本再配。
    截至 2026-03-25,SillyTavern 的最新 release 是 1.16.0(发布于 2026-02-14)。另外,SillyTavern 上个月还有用户报过 1.15.0 在 Seedream 图像生成里使用了不对的官方 URL;报错里提到官方正确地址应为 https://ark.cn-beijing.volces.com/api/v3/images/generations。所以你如果现在还是旧版,先升级,能少踩很多坑。(GitHub)

  3. 在 SillyTavern 里这样进:
    Extensions → Image Generation,启用图像生成扩展;
    然后把 Source / 生成源 选成 OpenAI。SillyTavern 官方文档列出的图像来源里,OpenAI 是受支持的。生成后你可以用魔杖菜单,或者直接用 /sd 你的提示词 测试。(SillyTavern Documentation)

  4. OpenAI 图像源里这样填:

    • API Key:填你的火山方舟 API Key
    • Proxy URL / Reverse Proxy / Endpoint:填
      https://ark.cn-beijing.volces.com/api/v3
    • Model:不要手敲“Doubao-Seedream-5.0-lite”这种展示名,直接填你在火山方舟控制台里能调用的实际模型 ID 或 Endpoint ID
      这样做的依据是:火山方舟官方给出的 OpenAI 兼容接入方式,就是把 base_url 指向 https://ark.cn-beijing.volces.com/api/v3;而 SillyTavern 维护者也明确说过,OpenAI 图像生成可以通过生成设置里的 proxy URL 来改目标地址。(VolcEngine)
  5. 先做最小测试。
    进任意聊天窗口后直接发:

    text
    /sd a girl in hanfu, cinematic lighting

    如果这一步能出图,说明图像后端已经打通了。SillyTavern 官方文档说明 /sd ... 就是图像生成扩展的直接调用方式。(SillyTavern Documentation)

  6. 你想要“聊天里自动触发出图”的话,再开下面两个功能之一:

    • Use function tool:让模型判断何时该调用图像生成
    • Use interactive mode:按关键词触发“发图”
      SillyTavern 文档写得很清楚:图像生成扩展支持这两种自动触发方式;而 function calling 可用于包括 Custom API 在内的一批 Chat Completion 来源。(SillyTavern Documentation)

你最容易踩的坑有三个:

  • 坑 1:配错地方
    把火山方舟只配在 API Connections → Chat Completion → Custom(OpenAI-compatible) 里,这样只是聊天能走,不代表图像生成扩展会走那里。图像生成要在 Extensions → Image Generation 里单独配。(SillyTavern Documentation)

  • 坑 2:模型名填错
    火山方舟文档里经常同时出现“模型展示名”和“实际请求用的模型 ID/版本串”。SillyTavern 这里应该优先填你控制台里实际可调用的 Endpoint/Model ID,不要猜。火山方舟的图片流式文档也说明返回里的 model 是一个实际模型 ID 格式,而不是纯展示名。(VolcEngine)

  • 坑 3:SillyTavern 版本太旧
    如果你用的是 1.15.0 一类版本,刚好可能踩到 Seedream 官方地址变化这个坑。先升级再试。(GitHub)

如果你在你的 SillyTavern 版本里 看不到 OpenAI 图像源的 Proxy/Reverse Proxy 输入框,那就别硬接,直接上一个很薄的本地中转层,把 ST 的 OpenAI 图像请求转发到 Ark。最小 FastAPI 代理可以这样写:

python
from fastapi import FastAPI, Request, Response import httpx import uvicorn ARK_BASE = "https://ark.cn-beijing.volces.com/api/v3" ARK_API_KEY = "你的火山方舟API_KEY" app = FastAPI() @app.post("/v1/images/generations") async def images_generations(req: Request): body = await req.body() headers = { "Authorization": f"Bearer {ARK_API_KEY}", "Content-Type": "application/json", } async with httpx.AsyncClient(timeout=120) as client: r = await client.post(f"{ARK_BASE}/images/generations", content=body, headers=headers) return Response(content=r.content, status_code=r.status_code, media_type=r.headers.get("content-type", "application/json")) if __name__ == "__main__": uvicorn.run(app, host="127.0.0.1", port=8000)

然后在 SillyTavern 里把 OpenAI 图像源的代理地址改成:

text
http://127.0.0.1:8000

这时候 Model 仍然填你火山方舟控制台里的实际模型/endpoint ID。

你要是愿意,我建议你下一条直接把这几个信息贴出来:
SillyTavern 版本号、你现在填的 URL、你填的 model、报错全文。我可以按你的界面给你逐项改成能用的配置。

Compartir este Q&A