LLMs are essentially doing probabilistic text generation; they’re not real search engines, and they can’t perform real-time, web-wide reverse image matching. When they “look at images,” it’s more about recognizing style and features, not checking them one by one against a database.
Finding high-resolution sources relies on image indexing, feature hashing, and stock library matching at a fundamental level. That’s a capability of search systems, not an LLM’s strength. If the image itself isn’t well-known—no celebrity, no classic magazine shot—there may not be many high-res versions indexed online in the first place, so GPT basically can’t conjure one up. Put plainly, at the level of underlying principles, this kind of application is hard to do well with an LLM.
For obscure images, you still have to use traditional image search and sort by size—no way around it.