教程:文生图

用 generate_image 工具从一段文本提示词生成图像。这是最简单的工作流 —— 没有输入图像、没有蒙版、没有 vibe。

1. 生成你的第一张图像

result = await ctx.session.call_tool("generate_image", {
    "prompt": "1girl, fox ears, autumn leaves, masterpiece, best quality",
    "negative_prompt": "lowres, bad anatomy, watermark, signature",
    "width": 832,
    "height": 1216,
    "steps": 28,
    "scale": 5.0,
    "sampler": "k_euler_ancestral",
    "seed": 42,
})

结果是一个列表,其中包含一个 Image 块(智能体可以 “看到” 图像)和一个带有 保存路径的文本块:

[
    Image(data=b"...png bytes...", format="png"),
    "Saved 1 image(s): ['outputs/generate-20260725-133702-001.png']"
]

2. 生成多个变体

传入 n_samples 可以在一次 API 调用中产出多张图像(比分开调用更便宜):

result = await ctx.session.call_tool("generate_image", {
    "prompt": "1girl, fox ears, autumn leaves",
    "n_samples": 4,
    "seed": 42,
})
# 4 images saved; the first is returned inline.

3. 换一个模型

# V3 model
await ctx.session.call_tool("generate_image", {
    "prompt": "1girl, fox ears",
    "model": "nai-diffusion-3",
})

# V4.5 full (current default)
await ctx.session.call_tool("generate_image", {
    "prompt": "1girl, fox ears",
    "model": "nai-diffusion-4-5-full",
})

# NovelAI Diffusion V5 — natural-language prompts, EN/JP, true alpha
await ctx.session.call_tool("generate_image", {
    "prompt": "a fox girl sitting in autumn leaves, soft watercolor",
    "model": "nai-diffusion-5-full",
    "straight_alpha": True,  # pair with "transparent background" in the prompt
})

4. 复现一次生成

要复现同一张图像,设置相同的 seed 与相同的参数:

await ctx.session.call_tool("generate_image", {
    "prompt": "1girl, fox ears, autumn leaves",
    "seed": 42,
    "width": 832, "height": 1216, "steps": 28,
    "scale": 5.0, "sampler": "k_euler_ancestral",
})

输出逐位相同(前提是 NovelAI 的模型权重没有变化)。

5. 等价的 CLI 命令

uv run python -m novelai_image_mcp generate \
  --prompt "1girl, fox ears, autumn leaves, masterpiece, best quality" \
  --negative "lowres, bad anatomy, watermark" \
  --width 832 --height 1216 --steps 28 --scale 5.0 \
  --seed 42

下一步