OpenClaw GPT Image Generation

Problem

When calling an image generation API from a local agent, common failure modes include scattered auth configuration, oversized reference images, multi-image requests being collapsed into one canvas, and long-running tasks losing progress after interruption. gpt-image-openclaw turns this into a scripted, resumable workflow.

How It Works

• Configuration and auth: Prefer config.json in the skill directory; OPENCLAW_IMAGE_ACCESS_KEY and OPENCLAW_IMAGE_BASE_URL also work. The key should not be echoed in replies, logs, or generated files.
• Generation modes: Supports text-to-image, plus local reference images via --reference-image. Files larger than 4M are re-encoded below 4M; transparent backgrounds are flattened to white before upload.
• Multi-image sets: Detects cues like “first image,” “套图,” or “product image set,” then splits the request into separate tasks while preserving product subject, brand style, and aspect ratio.
• Quality and credits: Defaults to medium, with optional low / high; prints estimated_credits before task creation, and final charges follow the OpenClaw API response.
• Task recovery: Maintains openclaw_tasks.json with task IDs, status, image URLs, and timestamps, then resumes pending tasks on the next run.

Boundaries and Notes

This is well suited for local generation of e-commerce assets, detail pages, hero images, and scene shots that need to be downloaded and reused individually. Treat the access key like a password; it is not stored in the runtime task file. Higher quality tiers consume more credits, and failures identify the stage: config check, reference upload, task creation, or polling.