AI cosmetics bottle photo: glass, reflections and a clean label

For a cosmetics bottle still, send your real bottle photo as a reference and ask the model to restage it on a new surface with new light while keeping the label untouched. Glass, liquid and reflections are where generated stills look fake, and small label type is where they break, so run a draft at low quality, then render the keeper at high. On Sume the quality values for GPT Image 2.5 are auto, low, medium, high, xhigh and max (Image API).

Draft, then final

Draft the composition at quality: "low" and review framing, surface and light. Only when you like it, rerun the same prompt at high. Quality is billed by the model's output tokens, so the final is the expensive call; the draft pass keeps the number of expensive calls small (GPT Image quality tiers).

Prompt for glass

Describe the light, not just the object: "soft box from the upper left, a thin bright reflection along the left edge of the glass, a soft shadow with a slight caustic". Name the surface: wet black stone, pale travertine, mirrored acrylic. Keep the label sentence: "the label text must stay exactly as in Image 1".

If you want a brand name rendered fresh, quote it and spell it out once in the prompt. Do not expect small ingredient text to survive; place it later in layout.

• One bottle, centred, label facing camera.
• Surface and backdrop named explicitly.
• Reflection and shadow wording in the prompt.
• Label preserved by instruction, then checked by eye.

Output formats

Ask for output_format: "png" when you will retouch, and webp or jpeg for the web. Ratios on the GPT family include 1:1 and 4:5. For a transparent cutout of the bottle, GPT Image 2.5 supports background: "transparent"; see the OpenAI guide for the PNG and WebP requirement (OpenAI).