Rather than writing image-generation prompts from scratch, screenshot a chosen reference image and send it to a custom-trained GPT (trained on instructions for how to prompt image models) to have it broken down into a correctly formatted, model-ready prompt. The same GPT can also take product photos plus a separate style reference together and describe "the product rendered in that style" as the target image.
The GPT is pre-instructed on prompting conventions for different downstream image models, so its output is already formatted for whichever generator (e.g. Higgsfield) will consume it — removing the manual step of translating a visual reference into model-specific prompt syntax by hand.
Before reaching for the custom GPT, the base technique is simpler: screenshot a Pinterest reference image and ask plain ChatGPT to break it down into an image-generation prompt. A variant combines two images in one request — a product render plus a separate style-reference image — asking ChatGPT to describe a single final image of the product rendered in that style. The custom GPT is a trained shortcut for this same conversion step, pre-loaded with model-specific prompt formatting so the phrasing matches what a given image model (e.g. Higgsfield) expects.