kie-image
Generate images, video, music, speech or chat with ANY model on KIE.ai (GPT Image, Nano Banana, Seedream, Flux, Kling, Veo, Wan, Runway, Suno, ElevenLabs...). Optimizes the prompt, calls the right endpoint, polls, downloads and shows the result inline. Use when the user asks to generate/make an image, logo, video, music or voice "with kie" or names a KIE model.
mohamedbenhandil · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add mohamedbenhandil/kie-claude-skill --skill kie-image技能说明
name: kie-image description: Generate images, video, music, speech or chat with ANY model on KIE.ai (GPT Image, Nano Banana, Seedream, Flux, Kling, Veo, Wan, Runway, Suno, ElevenLabs...). Optimizes the prompt, calls the right endpoint, polls, downloads and shows the result inline. Use when the user asks to generate/make an image, logo, video, music or voice "with kie" or names a KIE model.
KIE generation (any model)
Setup
- API key from env
KIE_API_KEY(user sets once:setx KIE_API_KEY "<key>"on Windows, then reopens the app;exporton macOS/Linux). Never write the key into files, prompts or replies. If unset, ask the user to set it. - Header on every KIE call:
Authorization: Bearer $KIE_API_KEY. Base:https://api.kie.ai. - Balance:
GET /api/v1/chat/credit.
Find ANY model (do not rely on memory)
curl -s https://docs.kie.ai/llms.txtis the full index. Entries look likeImage Models > Google [Title](url.md); categories: Image, Video, Music/Audio, Chat. Grep it for the model name the user said (e.g.grep -i "veo\|kling").- Fetch the model's doc page (
<url>.md, English, not/cn/). Read the OpenAPI YAML: themodelenum value,inputproperties (required fields, enums, limits), and the endpoint path. Never guess parameter names or values; a model'saspect_ratio/resolutionenums differ per model. - Most models live under
https://docs.kie.ai/market/...and share one flow (below). A few families have dedicated docs/endpoints (Sunosuno-api, Veo3veo3-api, Runwayrunway-api, Flux Kontext, 4o Image, Chat models): follow that page's own endpoint and its own status/result endpoint. - Several size/quality or text-to-X vs image-to-X variants have different model ids. Pick the one that matches the request (text-to-image vs image-to-image, etc.).
Market flow (unified, verified for image models)
POST /api/v1/jobs/createTaskwith{"model":"<id>","input":{...}}->data.taskId. OptionalcallBackUrl(not needed; poll instead).- Poll
GET /api/v1/jobs/recordInfo?taskId=IDevery 5s (video/music: longer, 10-15s, up to several minutes) untildata.state=success(orfail: readfailMsg). Result URLs:resultJson->resultUrls.creditsConsumedshows the cost. - Download to the project's
output/folder and display (images:SendUserFilewithdisplay: "render", several files in one call; video/audio: attach and give the path). - Run several jobs in parallel with background
&+wait.
Verified models so far: gpt-image-2-5-flare-text-to-image (prompt, aspect_ratio, resolution 1K/2K/4K) and nano-banana-2-1 (prompt, image_input[], aspect_ratio, resolution, output_format). ~3 credits per 1K image. Pricing for other models: check the doc page or tell the user it is unknown and ask before expensive jobs (video, 4K, music).
Reference images / files
- Upload a local file:
POST https://kieai.redpandaai.co/api/file-stream-uploadmultipartfile=@path,uploadPath=images/ref,fileName=x.png->data.downloadUrl(temporary). Pass that URL where the model expects a URL (image_input,image_urls, ...; check the doc page). Also available: base64 and URL upload (docs: file-upload-api). - KNOWN BUG: Nano Banana 2.1 can return the reference image UNCHANGED when
image_inputis given with a "recreate/variation" prompt. After generating, compare file hashes or look at the image; if identical, regenerate text-only (omitimage_input) with a detailed description. Photo edits that preserve a brand logo worked withimage_input; still verify the text.
Claude intelligence layer (what makes this better than raw API calls)
- Pick the model for the user. If no model is named, choose from the KIE list by need and say why in one line: photoreal/editing, text-in-image or logos, fast drafts, text-to-video vs image-to-video, music, voice. Prefer the cheapest model that fits; ask before expensive jobs (video, 4K, music).
- Rewrite the prompt like a pro. Never send the raw sentence. Build: subject, setting, style/medium, lighting, lens or art direction, palette (hex), composition (e.g. text space on the left), mood, explicit exclusions ("no text, no watermark"). Video: add camera move, duration, subject motion, start/end frame. Music: genre, tempo, mood, instruments, structure, vocals or instrumental. Voice: language, tone, pacing. Follow the prompt style the model's doc page recommends.
- Read the reference. When the user sends an image, describe its mood, colors, composition and subject first, then reuse those words in the prompt so the result matches.
- Automate the loop. Launch variants in parallel, poll until done, download, display, then self-check outputs (spelling, cropping, forbidden elements, identical-to-reference) and regenerate failures once without asking. Offer one concrete next step (more variants, upscale, other ratio) instead of a menu.
- Keep session context. Reuse the user's chosen model, ratio, palette and uploaded reference URLs for follow-ups ("other style", "darker") without re-asking.
- Be economical. Show model, resolution, ratio, cost in credits and one line per result.
Quality rules learned
- Optimize the user's prompt (subject, style, palette with hex colors, composition, "no text/logos" when relevant) and use the user's exact model, resolution and ratio.
- Look at at least one output when text, logos or "no X" constraints matter, and say plainly when something was not checked.
- Text inside images (names, Arabic) is unreliable: tell the user to verify spelling; Arabic needs a native speaker.
- Logos: "flat vector, thin monoline, solid background", hex colors; "bold" reads as not luxurious; one concept per variant unless asked to combine. Output is PNG, not vector (needs SVG vectorizing for web/app).
- Web hero: subjects on the right, empty left for titles, "nothing cropped".
- Keep replies terse (user saves tokens): model, resolution, ratio, one