LLMs can be multimodal now. You can chuck a screenshot into claude code or whatever you use and the image generation process can also be done by the LLM itself rather than an external tool call as far as I know