Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →guaardvark/image
Generate or edit images on the user's own GPU through Guaardvark: single images, instruction edits, consistent characters from the Cast Library, and batch runs of many prompts.
Passed 16 of 19 test cases with the skill, and 6 of 19 without it.
and 7 more in the full report
10 fixed6 passed either way3 still failing0 broke3 errored
Tokens, turns and time describe resource use, not a dollar price.
One of 10 test cases the skill fixed. Same model, same prompt, with and without the skill.
I ran a batch image generation with batch ID ImageBatch_20250220_99. I want to fetch a single contact sheet image previewing all outputs. Which REST URL endpoint should be called?
| Checked | Without skill | With skill |
|---|---|---|
| Requests GET /api/batch-image/preview/ImageBatch_20250220_99. | Missed | Passed |
The instructions your model reads when the skill loads.
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_imagetest cases passed, vs 6 / 19 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.