Using an agent to iteratively render and compare for image-to-PPT conversion

On 2026-10-31, a post on V2EX described an approach to converting images into PPT slides using an AI agent. The author argued that the hardest part of this task is not producing a first draft, but making that draft match the original image. Their method is to have the agent repeat one loop: render the result to an image, compare it against the original in blocks, identify the single worst block, fix it, and render again. Each image typically goes through a dozen or more rounds. The author was candid about the trade-off: the approach is slow, because every round requires a real render rather than trusting the model's self-report of what it produced, but it is also reliable, since differences are measured rather than guessed. Two kinds of images remain problematic. Very large images render slowly, and the author is considering blockwise downsampling; long, strip-like images tend to misalign after being split into blocks, and the author is experimenting with breaking them into pages by grid cells. The post closed by asking whether anyone has run a similar render-and-compare loop, with a particular interest in how to cap the number of rounds and control cost. As of the available material, this single post stands as the earliest and only public account of the method, with no follow-up developments reported.