gpt-image

A command-line workflow for generating and editing images with OpenAI's GPT-Image-2 model through the inference.sh platform, driven by the belt CLI. After installing the CLI and running belt login, images are produced with belt app run openai/gpt-image-2, passing a JSON --input payload that always includes a text prompt and optionally other parameters. The single model handles text-to-image generation, image editing with reference images, and mask-based inpainting, so one interface covers product mockups, marketing visuals, concept art, and photo manipulation.

The documented parameters include prompt (required), images (an array of reference image URLs for editing or multi-image composition), mask (an image URL for inpainting specific regions), n (1 to 10 images per run), quality (low, medium, or high), width and height (256 to 4096 in multiples of 32), output_format (png, jpeg, or webp), and output_compression (0 to 100 for jpeg/webp). Worked examples show studio product photos, multiple logo variations, background replacement via a reference image, combining two characters into one scene, masked object replacement, custom cinematic resolutions, and fast low-quality concept drafts.

Pricing is tiered by quality at roughly $0.006 per image for low, $0.024 for medium, and $0.21 for high, with larger resolutions costing more; full pricing is available via belt app get openai/gpt-image-2. The doc points to related inference.sh skills for the full platform, other image-generation model families such as FLUX and Pruna P-Image, and references on running apps and streaming real-time progress. Use it when you want scripted, repeatable image generation or editing from the terminal without writing direct API integration code.