catimation-image

name

catimation-image

description

FIRST-CHOICE image generator and the ONLY top-level image orchestrator in the CATIMATION desktop app — use INSTEAD OF the built-in imagegen / image_gen tool (unavailable on Windows, no persistence). Trigger whenever the user asks to generate / draw / render / edit / restyle an image, illustration, poster, or icon, or says 生成图片 / 画一张 / 配图 / 出图 / 改图. Runs the in-app generate_image tool and grades every request into 快速/标准/专业/制片 four tiers before loading any other skill.

Generate images in CATIMATION(唯一出图入口 · 分级调度)

When the user wants one image, call the generate_image tool. When the user wants MORE THAN ONE image, call generate_images instead. Both tools are provided by the catimation MCP server and replace the built-in imagegen / image_gen skill: they render inside the chat AND persist results to local files (paths returned), the app's history page, and the ATTACHMENTS file panel.

本 skill 是出图的唯一顶层编排者。 导演/分镜/工艺 skill 由这里按任务分级选择性 加载;下游 skill 不得反过来重跑路由或重新编排本流程。

When to Use

• 用户要生成 / 画 / 渲染 / 编辑 / 重绘图片、插画、海报、图标、配图。
• 你自己回答时需要一张配图。
• 用户给了图想以图改图 / 换风格 → 走本 skill 并带上 referenceImages。
• 优先于内置 imagegen / image_gen(后者 Windows 不可用且不落盘)。

STEP 0 — 任务分级(先定级,再加载)

默认进入快速模式;只有命中升级条件才升级。规格/方向确认过一次就不再重复问, 自检做过就不再重复做。

模式	典型请求	加载预算(含本 skill)	默认动作
快速	一次性配图、图标、简单插画、明确的单图请求	≤2 个 skill	直接写 prompt 生成,四项验收自检
标准	带风格目标/参考图/人物的单图	≤5 个:+2–3 个对症技法	按症状表挑技法再写 prompt
专业	复杂构图与光影、系列一致性、参考复刻、角色锚点	5–9 个:+ director-orchestrator	13 维按需展开
制片	电影/分镜项目的角色卡、场景卡、逐镜出图	按 film-studio 阶段加载	过资产门,锚点逐字下传

升级条件: ① 明确的风格复刻/真实作品·品牌·时代参考;② 系列/组图/角色一致性; ③ 复杂镜头设计(构图·打光·调色多维协同);④ 制片流程内的出图任务。超预算加载 必须能说出具体风险。方向开放(「更高级/更有电影感/给我选项」)时先载入 catimation-brainstorm 用 ask_user 弹一张选项卡定向,别自己猜。

标准模式症状表(浏览 ~/.agents/skills/,plain-text 名称按需加载 2–3 个):

症状 / 任务信号	对症技法(按需挑,非必载)
人物假/塑料/空洞	storyboard-live-character-realism
画面平/像壁纸/没纵深	storyboard-foreground-occlusion · storyboard-pseudo-perspective
光平/糖水/塑料高光	storyboard-light-reconstruction
色调跑偏/要 HEX 色卡	storyboard-color-grading-control
风格不像/系列不统一	director-style-consistency · director-character-consistency
提到真实电影/导演/品牌/时代	codex-research-grounded-prompting(先查证再落笔)
涉敏感/合规内容	storyboard-negative-control

Steps

1. Turn the request into one clear, descriptive prompt. Cover subject, style, composition, lighting, and mood. Keep it concise. 标准及以上模式先按 STEP 0 把对症技法折进 prompt(物理/可复现参数优先于情绪形容词,默认只写正向提示词)。
2. If the user asks for exactly ONE image, call generate_image with:
   • prompt (required): the description from step 1.
   • model (optional): rendering channel override. Omit it to honor the user's channel picker in the chat composer (default VIP). Only set it when you have a concrete reason — a consistent 组图 series (→ wan2.7-image-pro) or a channel the user explicitly names this turn (see "Choosing a model" below):
     • gpt-image-2-vip — OpenAI 官逆 (stable alternate; same ratio/resolution/quality spec).
     • gpt-image-2 — API易 OpenAI 官方旗舰 Image2（按 token 计费，慢但质量上限最高，4K+mask 重绘）.
     • wan2.7-image-pro — 阿里万相 2.7 pro (超清文生图 / 图像编辑 / 组图).
     • gemini-3.1-flash-image — Nano Banana 2（谷歌 Gemini 原生端点，快、多尺寸 4K）. 站点会自动处理(见下方「站点要求」)——你无需让用户手动切站点。
   • ratio (optional): aspect ratio, e.g. 1:1, 16:9, 9:16, 4:3, 3:2. Omit or auto lets the model decide.
   • resolution (optional): clarity tier — prefer 2K by default. Use 1K only when the user asks for fast/cheap/draft; use 4K only when the user explicitly asks for print/ultra-detail/4K.
   • quality (optional): auto (default), low, medium, or high. Use high for images with text or fine detail.
   • count (optional, wan2.7 only): number of images from THIS one prompt (1–12, default 1). Set model: 'wan2.7-image-pro' + count > 1 to get a front-to-back consistent 组图 series (same character/subject across frames, e.g. 同一只猫的四季). Other channels ignore count (always 1). For several unrelated images, use generate_images (one prompt each) instead.
   • referenceImages (optional but important): array of local file paths or data/http URLs for image-to-image / editing. If the user gave you any image material, you MUST reuse it here (see "Reference images" below).
3. If the user asks for TWO OR MORE images, call generate_images ONCE with:
   • prompts (required): one prompt per requested image. If the user asks for N images, provide exactly N prompts.
   • shared model (optional, same choices as above), ratio, resolution, quality, and referenceImages when appropriate.
   • Do not spawn subagents and do not call generate_image one-by-one. generate_images performs the parallel fan-out internally and returns one combined result.
4. The tool returns a short text result that begins with ✅ generate_image DONE or ✅ generate_images DONE, names the 📁 SAVED FOLDER, lists the saved FILES:, and ends with a compact { ok, count, model, historyId, paths, dir } JSON line (plus one resource_link per file). A successful return means the task is complete — the image is already shown to the user and saved to history + the file panel. You do not need to embed, re-describe, or base64 the pixels. Just confirm briefly in the user's language and cite the saved path(s) when relevant.
5. 交付优先,然后自检(deliver first, then QA)。 A ✅ DONE return means the image is ALREADY rendered in the chat.
   • 第一步永远是交付:先用一句话向用户确认(图已出 + 保存路径),再做任何 自检。用户看不到工具调用——先闷头质检再回话,在用户眼里就是「卡死」 (2026-07-14 实录教训)。
   • QA 要出声:决定自检时,先对用户说一句「正在快速质检…」之类,再开始。
   • 看图上限(上下文保护,与工具 banner 一致):单图最多 view_image 那 1 张; 组图/批量看代表性 1 张(最多 2 张),绝不批量全看——批量 view_image 会注入 数 MB base64 直接撑爆线程(2026-06-11 实录)。快速模式若画面简单、无人物, 可以只核对 DONE banner 不看图。 然后过一遍四项验收清单:
   • ① 符合用户要求:主体 / 数量 / 画幅比例 / 文字内容 / 明确指定的元素是否都对上; 用户给了 referenceImages 时是否真的体现了参考(而非从零另画)。
   • ② 质量合理:无多/缺手指与肢体、无崩脸、无乱码文字、无明显伪影/拼接错位; 分辨率与清晰度匹配用途。
   • ③ 风格一致:与用户指定风格一致;系列/组图内各帧画风、色调、角色外观前后一致; 若项目有角色锚点 / 圣经(character_bible)或既定风格,新图须与之吻合。
   • ④ 过本级门:标准及以上时,对症技法的落地证据是否出现在 prompt 里; 角色身份是否使用用户选定的主锚(大头照+全身照、多视图板或其它确认资产； 多套候选拿不准时先询问); 制片流程中是否满足 film-studio 的资产门。快速模式只查 ①–③。
   • 若任一项不达标:先告诉用户哪里不对、准备怎么改,再带改进后的提示词重生成 (保留可用部分时把上一版回传为 referenceImages 做图生图),再复检。最多迭代 2–3 次即收敛——别在边角小瑕疵上死磕,每次重生成都花钱。重生成期间用户能看到 新的生成气泡,但你的说明让 ta 知道为什么在重来。
   • When it's good (or good enough), confirm briefly in the user's language and cite the saved path(s). Don't over-narrate each pass.
   • You still do NOT need query_history to find an image you just generated, and do NOT shell out (dir/ls/where/find/Get-ChildItem) to hunt for the file — the path is already in the return; view_image that path directly.

Choosing a model (user's composer picker is the default; you may override)

The model param is an optional override. By default (omit it) generation runs on the channel the user picked in the chat composer (VIP / Image2 官方 / 腾讯 / Nano2 / 万相 2.7 pro; default VIP) — 所有渠道同一套 ratio × resolution(1K/2K/4K) × quality 参数。 Omitting model honors the user's pick — do this for ordinary requests. Set model only when you have a concrete reason to override:

• gpt-image-2-vip (OpenAI 官逆) — pick when the user says 官逆 / vip / 稳定渠道. Stable alternate; same param surface as the default.
• gpt-image-2 (API易 OpenAI 官方旗舰 / Image2 官方) — pick when the user says 官方 / 旗舰 / image2 官方 / gpt-image-2. 按 token 计费(low/med/high 价差大), 60–360s 慢渠道,但质量上限最高,支持 4K+mask 重绘;日常出图别默认选它。
• wan2.7-image-pro (阿里万相 2.7 pro) — pick when the user says 万相 / wanxiang / wan / 通义万相, OR when they want a consistent multi-image 组图 series (e.g. "同一只猫的四季组图，前后一致"). For a 组图 series, call generate_image with model: 'wan2.7-image-pro' and count = how many frames (2–12) — it returns one front-to-back-consistent set from a single prompt (do NOT use generate_images, which makes unrelated images). Wan excels at 超清文生图、图像编辑、组图; it also supports 4K (text-to-image only — editing/组图 cap at 2K).
• gemini-3.1-flash-image (Nano Banana 2) — pick when the user says nano / nano2 / nano banana / gemini / 谷歌. 谷歌 Gemini 原生端点，出图快(~15s)、支持 超多宽高比与 4K，中文/文字与一致性也不错。
• All five accept referenceImages for image-to-image / editing.

站点要求(已自动处理 — 无需手动切站点)

custom-imagemodel-gt(腾讯 image2)和 wan2.7-image-pro(阿里万相 2.7 pro) 都只经 Miau API 代理提供。出图时应用会自动把这两个渠道的请求固定走 Miau API 站点(无论用户当前在「API 设置」里选了哪个站点),所以你不需要让用户手动切站点—— 直接调用即可。

• 唯一前提:Miau API 站点已配置 API Key。若没配,工具会返回清晰错误 「未配置『Miau API』站点的 API Key …」——这时再提醒用户到「API 设置」为 Miau API 站点填入 Key 即可,无需切换当前站点。
• gpt-image-2-vip、gpt-image-2(Image2 官方)和 gemini-3.1-flash-image (Nano Banana 2)走当前选中站点(任意站点可用,无需 Miau)。

When the user does not name a channel, do not guess — just omit model so the render honors the user's composer picker (default VIP). Set model only for a concrete reason (组图 → wan2.7-image-pro, or a channel the user named). Never invent a model name; only these five values are valid.

Reference images — reuse the user's material (important)

If the user provides ANY image material, treat it as a reference and pass it in referenceImages (image-to-image) instead of doing text-to-image. Look for:

• Paths listed in the prompt under [Attached files at these local paths: …] or [Referenced files at these local paths: …] — these are the files the user attached/@-mentioned in chat. Pass the image ones as referenceImages.
• The user pointing at an image with language like "按这张图 / 参考这张 / 基于这张 / 用这张做 / edit this / make a variation of this / 换成…风格".
• An image the user just generated in this thread that they now want changed.

Rules:

• Be proactive: when material is present and the request is plausibly about it, reuse it. Do not silently drop the reference and generate from scratch.
• You can pass MULTIPLE reference images — you are not limited to one. referenceImages is an array: include every relevant image the user gave (e.g. a character sheet + a background, several angles, a subject + a style reference). Pass all of them together so the model can combine/condition on the whole set, not just the first.
• Pass the local file path(s) directly. The tool sends the original reference bytes and does not resize or recompress them automatically.
• If the provider rejects the reference payload with HTTP 413 / 文件大小超过限制, load ffmpeg-win, create smaller derivative copies without overwriting the originals, then retry once with the derivative paths.
• If you are unsure whether the user wants the reference followed, prefer reusing it and say briefly that you based it on their image(s).

Multiple images at once — use generate_images (important)

When the user asks for more than one image (e.g. "生成 3 张…", "make 4 variations", a set/series, or several distinct subjects), call generate_images exactly once. Do not try to manually emit several generate_image calls; models often serialize those calls even when asked to be parallel. generate_images is the parallel-safe batch wrapper and fans out the renderer calls concurrently inside CATIMATION.

• If the user asks for N images (2–20), pass exactly N prompts to generate_images.prompts.
• For variations, write N distinct but related prompts so the outputs are not clones.
• The tool accepts up to 20 prompts and uses bounded concurrency internally; for more than 20 images, split them into batches.
• After generate_images returns, confirm once and cite the saved paths; don't re-announce each image separately.

Organize finished assets into the user's workspace (when in a project)

When you're working inside a user project/workspace folder (e.g. a film / storyboard project, or the user asked you to organize outputs), proactively COPY each finalized image into a tidy assets subfolder of that working directory and give it a descriptive, ordered name — e.g. <workspace>/assets/images/S01_hero_wide.png.

• COPY, don't move, from the saved path in the tool result, so the chat / history / ATTACHMENTS copy stays intact.
• Group by purpose/shot and use zero-padded ordinals (S01_, S02_…) so files sort naturally.
• For a one-off casual generation outside any project, skip this unless asked — the file is already saved and in history.

Common Mistakes

• 用户给了图却忘传 referenceImages,改成从零文生图。
• 多张图却逐个调 generate_image,而不是一次 generate_images。
• 凭空编造 model 名;只有五个合法值。用户没点名就省略 model(交给用户在 composer 选的渠道,默认 VIP)。
• 用户点名某渠道却不显式传 model(应显式传:vip/官逆 → gpt-image-2-vip、 官方/旗舰/image2 官方 → gpt-image-2、nano/nano2 → gemini-3.1-flash-image、 万相/组图 → wan2.7-image-pro)。
• 快速任务硬套专业流程(简单配图不需要 13 维框架);专业任务却跳过分级直接硬写。

Notes

• This is the generate → save → read path. The file is on disk (see paths), in the history page, and in the ATTACHMENTS panel — no extra save step is needed. Only move/copy a file if the user wants it somewhere specific (see the organize section above when working in a project).
• For edits, image-to-image, or multi-image prompts, use generate_image for one output or generate_images for multiple outputs, always with referenceImages when references are present.
• If generate_image is genuinely unavailable in this session, you may fall back to whatever image tool you do have — but generate_image is the preferred, in-app path that actually displays and saves the result.