Configure a subagent in your coding harness to spin up a new sub-session with any vision model for those tasks and feed the result back to the main model. No need for "one model that does everything"

Are you suggesting it should summarize the image in text or generate it in HTML or something else?