AI image-generation pipeline

Every property burns through images. Hero shots, social posts, seasonal refreshes, all of it constant. The default answer is a designer on retainer or an agency queue, and I wanted neither. What I actually wanted was to type a prompt and get back a graded, on-brand image built from our own photography.

How it works

I split the job along a line I learned the hard way. Gemini's image model (billing on, paid tier) does the photographic part: color grade, atmosphere, background restyle, the scrim. Pillow does the text layer, and only the text, because the image model misspells caption copy every single time. An orchestrator script hands the graded background to a Pillow text engine. I cache every Gemini background per source photo, so reworking the copy or the scrim re-composites for free with no new API call.

• Gemini for the light and the grade, Pillow for crisp type, never the other way around
• A brightness-aware legibility scrim owned in Pillow, because the model places its gradient inconsistently
• Cached backgrounds, so iterating on the words costs nothing
• Runs against a property's real photos, not an invented building

Gemini owns the light. Pillow owns the letters. Ask the image model to spell and it will let you down every time.

This is the whole marketing-production seat compressed into a repo. There is no creative queue to wait on, and when an exception shows up (a bad grade, a template-y layout) I fix the script once and it stops coming back. That is the shape of the bet in every seat. The owner keeps the taste, the agent keeps producing.