Generate or edit images on the user's own GPU through Guaardvark: single images, instruction edits, background cut-outs, inpaint and outpaint, consistent characters from the Cast Library, and batch runs of many prompts. Use when the user asks to create, draw, render, visualize, edit, or batch-generate images locally.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Read setup first if the backend or the comfyui plugin state is unknown.
generate_imageprompt is scene, pose, lighting, setting. Plain prose. Do not paste JSON or tag soup; the
default model (Z-Image Turbo) reads prompts as language, and SD-era tag lists hurt it.model default auto picks the best downloaded model. Only override when the user names one:
zimage-turbo, krea2-turbo, krea2-raw, flux-dev, sd-xl, sdxl-turbo,
realistic-vision, epic-realism.width / height: 512, 768 or 1024. style: realistic, artistic, anime, photographic, digital-art.subject_ids=[<cast id>] as its own array. Never put the
trigger word alone in the prompt and expect the LoRA to load. Find ids with
GET /api/cast-library (see the cast skill)./api/outputs/generated_images/<file>.png, relative to the
backend), the model that ran, steps, seed and whether a Cast LoRA was applied. Show the URL
and the prompt you used. Measured: 768x768 on Z-Image Turbo in ~20 s on a free 16 GB card.wait_for_result defaults to false there) and
returns Image queued as batch ImageBatch_... at once. Poll
get_generation_status(batch_id=...) every few seconds until completed; it returns the
file URL. Pass wait_for_result: true to block for the render instead (allowed up to 30
minutes). A call that exceeds the server's timeout answers with an error that says the
render is still running; it is not lost.inspect_gpu and GET /api/plugins/status are the two checks that resolve most.edit_image, inpaint_image, outpaint_image and remove_background answer at once over
MCP with ... queued as job tooljob_...: the work runs in the Guaardvark backend and keeps
going if the client disconnects. Poll get_generation_status(batch_id="tooljob_...") every few
seconds until done (it gives the file URL) or failed (it gives the error). Pass
wait_for_result: true to wait up to 60 s (half the server's MCP timeout when that is
shorter) for the result; a longer edit still comes back as the job id. A Qwen-Image-Edit takes
about three minutes on a 16 GB card, and a job waits its turn while another render holds the GPU.
edit_imageinstruction is the change ("put a cowboy hat on him", "make the shirt red"). The image the
user just attached is used automatically; otherwise pass image as a path or URL.
Optional reference_image_2 / reference_image_3 (Qwen-Image-Edit only) for extra people or
style; the call is refused when a reference cannot be read or the edit would run on another backend.model auto uses Qwen-Image-Edit when installed, else FLUX.1 Kontext. With neither
installed the call is refused and names the pack; relay that instead of retrying.
Override with qwen-image-edit or kontext. Install those packs from Manage Image Models →
Image editing (qwen-image-edit, flux-kontext-dev) — do not Install unless the user asked.
Naming another downloaded image model runs a light img2img pass that keeps most of the picture.generate_image. For a new scene
that keeps a face use generate_identity.generate_identity (off by default)GUAARDVARK_IDENTITY_TOOL=1: the likeness it keeps
has not passed verification yet. If the tool is absent, say so and offer edit_image instead.consented must be true. Refuse if they have not confirmed that.prompt is the new scene. Needs the PuLID identity pack (pulid-flux; Manage Image Models →
Image editing installs it with its face files, EVA02-CLIP and flux-dev). Comfy must have
been restarted after the PuLID-Flux custom node was added.remove_backgroundedit_image / generate_identity, or describe the
new scene in edit_image if Qwen-Image-Edit is installed.inpaint_image: change or remove something ("remove the coffee cup"). Runs on Qwen-Image-Edit
or FLUX.1 Kontext; refused when neither is installed.outpaint_image: extend the canvas (left/right/top/bottom pixels) and fill. Prefers Qwen.B=${GUAARDVARK_URL:-http://localhost:5000}
curl -s -X POST $B/api/batch-image/generate/prompts -H 'Content-Type: application/json' -d '{
"prompts": ["prompt one", "prompt two"],
"model": "auto",
"subject_ids": []
}'
prompts may be strings or {"prompt": "..."} objects. There is a per-batch maximum; if the
server answers 400 "Too many prompts", split the list.adapters (user LoRAs from the models skill) and subject_ids (Cast Library).data.batch_id (ImageBatch_<date>_<n>). Poll
GET $B/api/batch-image/status/<batch_id>?include_results=true: status goes
running → completed, with completed_images / total_images, output_dir, and one
results[] entry per prompt (success, image_path, thumbnail_path, generation_time,
metadata.model_used). A contact sheet: GET $B/api/batch-image/preview/<batch_id>; one file:
GET $B/api/batch-image/image/<batch_id>/<image_name> (the basename of image_path).
Cancel with POST $B/api/batch-image/cancel/<batch_id>. Measured: one 1024x1024 prompt
completed in ~30 s.POST /api/batch-image/enhance-prompt, /analyze-prompt, /expand-concept (JSON
body with the prompt) when the user wants prompt help before spending GPU time.まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Build consistent characters, environments and props in Guaardvark's Cast Library and train LoRAs for them locally (reference photos → vision bible → sample plan → approved samples → training). Use when the user wants the same face or object across images, videos, a music video, or a Film Crew production, or asks to "train a LoRA".
日本語の概要は準備中です。原文の説明を表示しています。
Code intelligence over repositories Guaardvark has indexed: semantic code search, exact-text search, file listing, PageRank repository map, import dependency graph, reading one class or function, analysis and generated rewrites, plus the System Mapper and self-improvement status. Use when the user asks where something is implemented, how modules depend on each other, or wants an architecture map of an indexed repo.
日本語の概要は準備中です。原文の説明を表示しています。
Run Guaardvark's five-role Film Crew locally: screenwriter, casting (Cast Library LoRAs), cinematographer storyboards, keyframe renders, editor. Turns a logline or screenplay into a finished short video with human gates at casting and storyboard. Use when the user says "film this", "make a short film", "storyboard this script", or "start the film crew".
日本語の概要は準備中です。原文の説明を表示しています。
Answer from the user's own indexed documents with Guaardvark's local RAG, browse and read indexed files section by section, remember facts across sessions, and process or fetch new material. Use when the user asks what their documents say, wants a source-cited answer from their files, says "remember this", or asks to index a PDF, DOCX, spreadsheet, image or URL.
日本語の概要は準備中です。原文の説明を表示しています。
Add any Hugging Face image or video model, checkpoint or LoRA to Guaardvark from a URL, list what is installed, and download registry models on request. Use when the user pastes a huggingface.co link or a .safetensors URL, asks "can it run <model>", or wants a new LoRA or checkpoint available in the Studio.
日本語の概要は準備中です。原文の説明を表示しています。
Generate full songs with vocals or instrumentals (ACE-Step) and sound effects or ambience (Stable Audio Open) on the user's GPU through Guaardvark's Audio Foundry. Use when the user asks for a song, a beat, a jingle, background music, or a sound effect, made locally with no subscription.
日本語の概要は準備中です。原文の説明を表示しています。