Wonky Cute Doodles

Provide a subject or one-sentence scene. Optional placement/ratio, palette, mood, props, relationships and authorized subject references refine it. Make reasonable choices for minor gaps rather than asking for a full art-direction questionnaire. Default to no text; request exact short wording only when essential.

Choose one useful action and an original cast built from simple shapes. Keep one focal subject, limited supporting props and readable silhouettes. Compose flat/front-facing on clean paper with generous margins; vary scale, eye size, limb posture and spacing without adding unrelated characters or tracing a source design.

Write visible style facts into the prompt: wax-crayon grain, doubled pencil contours, uneven pressure, scratchy looping fills, gaps and small overshoots. Use roughly four to seven bright colors and no gradients/airbrush realism. For a series repeat the locked paper, marks, face and palette/space contract; change scene content, not the whole visual system.

Use an available native image tool when the user chooses it. Otherwise the bundled route requires Node.js 20+ and server-side FAL_KEY and uses paid GPT Image generation/edit credits. Obtain authorization for a paid task and save the prompt, output and task paths. API quality defaults to low as a cost setting, not a definition of rough drawing quality.

Choose the intended size explicitly. Fresh-text generation defaults to landscape 1024×576; reference editing defaults to auto. Fixed presets also include 512/1024 squares, 768×1024, 1024×768 and 576×1024. The helper checks exact preset pixels and can reject a provider result that differs—even if larger. Use a supported setting and report actual dimensions; do not silently stretch or label a rejected output complete.

For editing, supply only permitted subject references. The helper accepts up to sixteen local PNG/JPEG/WebP, image data URIs or public HTTPS URLs, with local/inline size limits. Local/inline content is fingerprinted after normalization; remote references bind to URL strings rather than fetched pixels. Keep the role as subject guidance, not permission to copy artwork, branding or signature props.

The helper exclusively saves a submission record before the paid POST and locks task processing. Resume with the same prompt, references, endpoint, settings and output. If submission is uncertain, inspect provider history and adopt only the matching existing request ID; adoption attaches that task rather than generating again, but the script cannot prove the ID matches your brief. Never delete an uncertain record or automatically replace a failed/unsatisfactory paid task.

Preserve approved outputs and records. The helper checks returned raster format/container/dimensions and digest metadata, stages the bytes and finalizes without overwriting another image. Run the output validator with format/minimum-width/exact-16:9 options when relevant and scan supplied prompt/text directories for forbidden icon glyphs. Container checks are not full pixel decoding, OCR or semantic QA.

Inspect full-size and thumbnail images for crisp output, visible paper, doubled contours, real stroke direction, small gaps, cheerful readable faces, original silhouettes and complete crops. Short exact text must be checked character by character; incorrect lettering is not a style feature. Use an approved new generation or a separate typography layer for corrections, and preserve transparency only when actually required/verified.

Name one defect and one focused change if needed: restore sharpness without smoothing marks, reduce props for clutter, vary posture for uniformity or replace cast/action/arrangement if too close to a reference. A new paid task requires approval and a new record; keep prior outputs. Deliver WebP for ordinary web use, or appropriate PNG when required, with actual paths/counts, prompt, task/request details and remaining limitations.