ueboxai/uebox/image-generation

Image generation and editing

Quick start

Use generate_image with the user's configured model. It accepts prompt, optional reference_images, count, aspect_ratio, size, seed, model and name. It returns images for inspection and saves outputs to the asset library under AIGC/图片. Do not invent extra tool parameters, use a different provider without authorization, or substitute a generated picture for actual engine output.

Decide the intent

• New image: describe the desired result. References may guide style, composition or subject.
• Edit: identify the target image and state exactly what changes and what stays unchanged.
• Inspect unseen local input images with an available image-viewing tool before generation.

Use attachment or tool-returned absolute paths; never guess image contents from filenames.

• For each reference, explicitly label its role in the prompt using input order:

“Image 1: edit target; Image 2: style only; Image 3: object to insert.”

• One prompt with count produces variants. Distinct assets require distinct prompts/calls.

Keep the requested number of deliverables; do not create an unsolicited batch.

Shape the prompt

If the user is specific, normalize their request without adding creative requirements. If generic, add only useful framing, lighting or scene detail. Do not invent extra characters, objects, brands, slogans, palettes or story beats. Do not choose arbitrary left/right placement unless the requested layout calls for it. Chinese and English are both valid.

Use only the useful lines of this scaffold; it is prompt text, not additional tool arguments:

Asset use: where this image will be used
Primary request: the user's objective
Input images: numbered roles, if present
Scene and subject: environment, identity, exact subject counts and actions
Style/medium: photo, illustration, pixel art, etc.
Composition: viewpoint, framing, placement and useful empty space
Lighting/materials: relevant light, mood and surface detail
Text (verbatim): "exact requested text", typography and placement
Change only: the requested edit
Keep unchanged: relevant identity, pose, layout, palette or object details
Avoid: unwanted elements and explicit prohibitions

Keep prompts concise. Preserve user-supplied text verbatim, including language and spelling; do not translate visible lettering when translating the rest of the prompt. Specify counts explicitly. For edits, repeat invariants every iteration and describe one targeted change. For transparent assets, request actual transparency; a drawn checkerboard is not transparency. Do not promise alpha or exact dimensions merely because the prompt requested them.

Examples

New asset: “Game character concept. Exactly one adventurer cook and exactly three friendly monsters at a fantasy campsite. Chibi illustration, wide composition. No text, logos or UI.” Use these subject/count/style choices only when requested; examples are not defaults.

Background edit: “Image 1 is the edit target. Replace only the background with a warm sunset sky. Keep the person's face, hair, clothing, pose and framing unchanged. Add no objects.”

Style reference: “Image 1 is the character to preserve. Image 2 provides brushwork only. Apply that brushwork to Image 1; retain its identity, outfit, pose, palette and composition. Do not copy subjects or lettering from Image 2.”

Inspect and deliver

1. Inspect the returned image against subject/identity, counts, composition, exact text,

style, invariants and avoid items. Tool success alone does not establish correctness.

1. Distinguish objective mismatches from taste. Describe the actual mismatch plainly.

Only judge images actually viewed; beyond the returned previews, inspect saved files.

1. Every generation spends the user's API credits. Default to one image unless more are

requested. Use ask_user before an unapproved paid retry; an explicit redo request or previously authorized iteration budget does not require repeated approval.

1. For authorized edits, reuse the approved output as the edit target and repeat invariants.

A fixed seed cannot guarantee identity or composition. The current OpenAI/GPT and Gemini integrations do not send seed; only SiliconFlow and Ark integrations forward it.

1. Show the result and saved path, actual model and final prompt. Project-bound assets must

reach the requested destination using available file/import tools. Do not overwrite the original unless replacement was requested. Use ue_content_import for UE project imports.

1. Report save failures and unverified requirements honestly. Verify alpha and dimensions

from the file with available inspection tools when those properties are required.

Failure handling

Missing model configuration: relay the tool's setup guidance. Do not guess a model. Invalid reference path: recover the real path from attachments or prior tool output. Unsupported multiple references: do not silently drop or merge images; explain the limitation and ask the user to choose a supported integration or explicitly select a single image. Reference/edit request failure: do not automatically remove references and spend on a different task to diagnose it. Report the actual error; a text-only success does not prove why editing failed. Ask only when a missing detail blocks execution, otherwise proceed within the user's request.