EPIC! ChatGPT vision does INSANE things for research!

ChatGPT’s image understanding is becoming a practical shortcut for science writing: upload figures, ask it to interpret what’s happening, then iterate until the key trends and details match the paper—turning hours of figure comprehension and draft text into a fast first pass. The most striking results come from feeding it peer-reviewed-style schematics, complex multi-axis graphs, and photo panels, often without providing captions. In those tests, it reliably identified the overall process steps (including the order of operations), extracted meaningful relationships from cluttered plots, and summarized what the figure is communicating—then converted that understanding into paragraph-ready text suitable for a thesis or manuscript.

A clear example involves a schematic describing a transparent electrode made from carbon nanotubes and silver nanowires. Without the figure caption, ChatGPT still recognized the process structure: combining two solutions, passing them through filter paper, and describing what remains on the filter paper (including thickness/pattern formation). It made at least one mistake—stacking filter papers rather than the intended setup—but the error was described as fixable with a short follow-up correction. The same pattern showed up with a highly complex graph featuring many arrows, shared axes, and multiple data elements. Even without the caption, it produced an explanation that matched the major trends, such as the relationship between weight fraction of single-wall carbon nanotubes and properties like sheet resistance and average spectral transparency. While minor inaccuracies appeared, the “important trends” were treated as correct, which is exactly what researchers need when drafting figure narratives.

Photographic figures also worked well, especially when captions were included. For a structural stability experiment using acetone exposure across different film compositions, ChatGPT generated a paragraph that captured degradation behavior across columns—rapid degradation for the pure silver nanowire film and more limited structural changes for films with 20% single-wall carbon nanotube weight fraction. The transcript emphasizes that writing or supplying accurate figure captions boosts performance, likely because captions encode the experimental intent and column labels that the model might otherwise infer imperfectly.

Tables were the weak spot. A low-resolution table with mixed text and numbers was summarized correctly at a high level, but it struggled with specific numeric comparisons—such as which RS or RP values increased versus decreased. The workaround proposed is iterative prompting: correct the model by pointing out what’s wrong, then ask for a revised paragraph. Because the workflow is conversational rather than one-shot, the model can be guided toward the correct interpretation.

Finally, the transcript draws a comparison between paid and free tiers. ChatGPT Plus handled the same schematic prompts far better than a free/open model (“lava”), which missed key details in the schematic—details that matter most for research accuracy. The practical takeaway is a workflow: (1) screenshot someone else’s confusing figure and ask for an explanation, (2) upload your own figures and request trend extraction, (3) ask for paragraph-ready text, then (4) correct any mistakes and re-run the request. Used this way, image-capable ChatGPT becomes a self-editing tool for figure clarity and a time-saver for the “boring parts” of drafting scientific writing.