Manual testing in ComfyUI surfaced two defects in the gpt-image-2.5 work.
background=transparent was silently useless: the API returned a correct RGBA
PNG, but pil_to_tensor ran convert("RGB") on every response and composited the
alpha onto black. Both nodes now output (IMAGE, MASK), with the alpha carried on
the mask as 1 - alpha to match the LoadImage convention, so Join Image with
Alpha consumes it directly. The three response decode helpers collapsed into
response_item_to_pil, since the old path already decoded each item to a PIL
image and then discarded the alpha -- routing through a single decode was
smaller than adding a second one.
CUSTOM_DIM_MIN was 1024 while validate_custom_dimensions() accepts a short edge
down to 480 (1440x480 at the 3:1 aspect limit). The widget was a second, stricter
gate that disagreed with the validator, so legal sizes like 1536x864 snapped up
to 1024 and were unreachable from the UI. The floor is now derived from what the
validator accepts.
The existing tests missed both: they assert on request kwargs and never checked
that alpha survives the response, and they called run_generate() directly,
bypassing the widget bounds entirely.
Verified against the live API (transparent output recomposes to a fully
transparent corner pixel) and by manual testing in ComfyUI.
Alexander Garzon
e8fdd92f21