codex-gpt-image-2-5-flare

Write a prompt specifying one isolated asset with genuine alpha transparency, describing materials, silhouette, padding, and transparent openings. Call the built-in image_gen.imagegen tool without reference parameters for new images, then verify the output before use.

Run the bundled inspect_png.py script with --require-transparency and the image path. It uses Pillow to report alpha range and transparent, partial, and opaque pixel percentages, failing if the PNG is fully opaque or entirely invisible.

Yes. Inspect each local target with view_image first, then pass referenced_image_paths to the image generation tool. Use num_last_images_to_include only for targets without local paths, and never combine both reference mechanisms.

No. The current built-in tool exposes no model selector, and GPT 2.5 Flare is only the workflow's requested name. Never invent a model argument or report an unverified model as fact.

RGBA mode alone does not prove transparency; the alpha channel can be fully opaque, and painted checkerboards mimic transparency visually. The inspection script measures actual alpha pixel distribution to catch these cases.

It cannot select models, sizes, or output paths through tool parameters, and alpha statistics do not confirm intended holes are transparent. If the built-in tool is unavailable, the workflow stops rather than substituting external providers or hand-drawn SVGs.