Add gpt-image-2.5-flare (new default) and gpt-image-2.5-sunburst, released
2026-09-08. Both support custom WxH sizes, transparent backgrounds and the
new xhigh/max quality tiers; gpt-image-2 falls back to high when an extended
tier is selected.
Retire gpt-image-1 and gpt-image-1.5 ahead of their 2026-10-23 and
2026-12-01 shutdown dates. That drops the input_fidelity parameter with
them: gpt-image-2 and 2.5 always use high input fidelity and error if it is
sent, so no shipped model can use it.
With every remaining model accepting custom sizes, the custom_size
capability flag became constant-true; removing it collapses resolve_size to
a validation call and drops its unused model argument.
BREAKING CHANGE: saved workflows selecting gpt-image-1 or gpt-image-1.5 must
have their model widget re-picked.
BYOK OpenAI gpt-image generation and editing nodes for ComfyUI — calls go
directly to the OpenAI API with your own key (no proxy, no credit markup).
- Two nodes: "Imagent: OpenAI Image" (text-to-image) and "Imagent: OpenAI
Image Edit" (edit / inpaint / multi-reference compose).
- Models gpt-image-2 / gpt-image-1.5 / gpt-image-1 with a per-model dynamic UI
(comfy_api DynamicCombo): size, background, custom WxH, and input_fidelity
are revealed only for the models that support them.
- Edit node accepts an auto-growing set of reference images (up to 16) plus an
optional inpaint mask; generate supports batches up to 8.
- API key via OPENAI_API_KEY or a local config.json.
Requires ComfyUI >= 0.23.0.
Alexander Garzon