Allow image generation to use reference image.

Venice has ability to analyze an uploaded image, and you have multiple image generation models, but they are only text to image, not text + reference image to image.

All of the venice text to image models have difficulty positioning bodies precisely (such as in the schoolgirl pin wrestling position the wrestler is always in squat position with knees in air instead of against the mat and no level of prompting seems able to get the knees pressed down. A reference image would be very useful in such scenarios.