ChatGPT and Midjourney can produce a convincing-looking signalling pathway in forty seconds. Neither can produce one you can submit. The blocker is not artistic quality — it is file structure and publisher policy. Both tools emit flat raster images: OpenAI's image API returns PNG, WebP or JPEG at a maximum of 1536 × 1536 px, and Midjourney's V7 upscalers roughly double a 1024 px square to 2048 px. Nature's final-submission guide asks for the opposite — "editable, unflattened vector artwork" with live, un-outlined text. Separately, Springer Nature journals do not permit AI-generated images in publications at all, and Elsevier's June 2026 journal policy forbids general-purpose generative AI image tools for graphical abstracts. This piece covers what those rules actually say, the resolution arithmetic, and the parts of the figure workflow where generative tools are legitimately useful.
The failure is structural, not aesthetic
Most "AI can't do science figures" articles argue about whether the mitochondria look right. That is the wrong argument, and it is getting weaker every model release.
The argument that does not weaken is about what comes out of the pipe. A general-purpose image model produces a grid of pixels. There is no text object in it, no arrow object, no layer. The word "phosphorylation" in that image is not a word — it is a smear of coloured pixels that happens to resemble one, which is also why these models so often render it "phosphorlyation" and cannot be told to fix just that.
Nature's guidance on this is unusually blunt. From the final-submission guide: "Please do not rasterize line art or text in submitted figures and wherever possible supply editable, unflattened vector artwork." And a line above it: "Do not rasterize or covert text to outlines." The research figure guide repeats it — "For figures to be accepted all text needs to be legible and editable" — and explains why production needs it: they have to re-set your labels in house style and rescale artwork without re-typesetting your figure by hand.
A PNG cannot satisfy that requirement. Not a better PNG, not a bigger PNG. The requirement is about structure.
:::figure{size=large align=center}
BLOGFIGUREMARKER0END
:::
The resolution arithmetic, done honestly
Resolution is the secondary problem, but it is worth doing the multiplication because people assume "upscale it" solves everything.
Nature's standard figure widths are 89 mm single-column and 183 mm double-column, with a full page depth of 247 mm. Photographic images must arrive at a minimum of 300 dpi at the size they will be printed, and the research figure guide notes that online proofs top out at 450 dpi, so 450 dpi and above is the safe target.
Run the numbers on a native ChatGPT image at 1536 px on its long edge:
- At 183 mm (7.2 in) double-column width: 213 dpi — below the 300 dpi floor.
- At 89 mm (3.5 in) single-column width: 438 dpi — clears 300, misses the 450 dpi preference.
Midjourney's 2048 px upscale does slightly better: 284 dpi at double-column, 585 dpi at single-column. Its widescreen 2912 × 1632 output reaches about 404 dpi across a double column.
So the picture is not "AI images are always too small." A single-column raster panel can genuinely clear the resolution bar. What it cannot clear is the format bar, and Nature closes the escape hatch explicitly: "Artificially increasing an image's resolution in an artwork program will not improve its quality." Upscaling a 1536 px schematic to 4000 px produces a large file containing the same amount of information and the same non-editable labels.
What publishers actually say about AI-generated images in 2026
This is where the conversation has moved since 2024, and where most advice online is out of date. Two of the largest publishers now have explicit, differing positions.
Springer Nature: not permitted, with narrow exceptions
Springer Nature's generative AI guidance, as applied across its journals and books, states that while the legal questions around AI-generated images remain unresolved, its journals "are unable to permit its use for publication." The listed exceptions are narrow: images from agencies it has contractual relationships with; images in pieces specifically about AI, case by case; and generative tools built on attributable, verifiable scientific datasets. All exceptions must be labelled as AI-generated in the image field.
Elsevier: a three-category framework
Elsevier's journal policy (updated June 2026) is more permissive and more specific. It sorts images into three buckets:
- Explanatory images — flow charts, decision trees, timelines, schematic conceptual illustrations, experimental workflow diagrams. AI tools may be used, with disclosure in the image caption naming the specific tool and version, plus the general AI declaration.
- Research and data images — data visualisations must be reproducibly derived from real data; primary research images (microscopy, blots, scans) must not be created or altered by AI at all.
- Graphical abstracts and cover art — "General-purpose generative AI image tools must not be used to create graphical abstracts." Elsevier instead encourages "dedicated scientific illustration tools," on the grounds that they offer well-defined licensing terms.
That third clause is the sharpest sentence in the 2026 policy landscape for anyone considering Midjourney for a graphical abstract. It does not say AI is forbidden — it says general-purpose image generators are the wrong instrument.
Elsevier also imposes a requirement most researchers skip: check the tool's terms so that you "only grant limited rights to the AI Tool", that it does not train on your inputs, and that its terms do not "restrict the subsequent publication" of your work. Free consumer chat tiers frequently fail at least one of those tests.
Comparison: what each tool can hand you
Checked as of September 2026. Policies and product specs change; we re-check monthly.
| ChatGPT (image generation) | Midjourney V7 | Dedicated figure tool (e.g. FigPad) | |
|---|---|---|---|
| Native output | PNG / WebP / JPEG raster | Raster (JPG/PNG via web app) | Layered SVG, PPTX, PNG |
| Max native size | 1024×1024, 1024×1536, 1536×1024 | ~2048×2048 after built-in upscale | 2K / 4K / 8K PNG by tier; SVG is resolution-independent |
| Text after export | Pixels — not selectable, not spell-checkable | Pixels | Live text objects in SVG and PPTX |
| Fix one label | Re-prompt, get a different figure | Re-prompt or vary | Edit the label in place |
| Nature "editable, unflattened vector" | ✗ | ✗ | ✓ via SVG/PDF route |
| Elsevier graphical abstracts | ✗ general-purpose tool | ✗ general-purpose tool | Permitted category, still requires disclosure |
| Springer Nature journals | Not permitted | Not permitted | Not permitted either — see below |
| Trains on your inputs? | Depends on plan and settings | Check current terms | FigPad ToS: "We do not use Your Content to train our own models" |
| Genuinely good at | Drafting plotting code, literature synthesis, critiquing a draft figure | Concept art, talk openers, non-manuscript visuals | Editable manuscript figures, graphical abstracts, posters |
One honest note on the last column: no tool exempts you from Springer Nature's position. A figure generated by a dedicated scientific illustration tool is still a generated figure. What a dedicated tool changes is the format problem, the licensing clarity, and your ability to comply with Elsevier-style disclosure — not the existence of stricter publishers.
Why "just ask ChatGPT for SVG code" is a real answer, and its ceiling
There is a genuine workaround worth knowing, because it is the strongest thing ChatGPT can do here.
ChatGPT does not have to generate an image. It can write code that draws one: raw SVG markup, TikZ, matplotlib, D3, Mermaid. Code output is vector by construction, so the text stays text and every element is addressable. For a two-box-and-an-arrow flowchart, a CONSORT diagram, or a plot from your own data, this route is often better than any image generator and costs nothing extra.
Its ceiling is complexity. Ask for a hand-drawn-style kidney nephron with eight labelled cell types and you will get either something schematic and crude or a hundred lines of SVG that renders as overlapping rectangles. Anatomical and cellular illustration is exactly where hand-written path data collapses, and iterating on it in a chat window is slower than drawing it yourself. Use the code route for structural diagrams and data plots; do not use it for illustration.
How FigPad fits
FigPad exists because of the specific gap this article describes: the generation step is solved, and the revision step is not.
The workflow is prompt, sketch or reference image in, editable figure out. Exports are layered SVG — meaning text is still text and arrows are still arrows, not a stroked-and-welded blob — plus PPTX where the graphics land in the slide as editable objects, and PNG at 2K, 4K or 8K depending on tier. The layered SVG export is available on the Plus and Pro plans. If your entry point is a summary figure for a paper, the graphical abstract generator is the closest landing page; for manuscript panels and schematics, start from the AI scientific figure generator.
Two details that matter for the policy questions above. First, editing an element on the canvas costs nothing — as of September 2026 the published pricing FAQ lists one generation at 2 credits, a text correction at 2, on-canvas element edits free, and an editable SVG export at 4 credits, with repeat exports of the same version free. A full generate-revise-export cycle lands around 16 credits, which is the number that matters when a reviewer asks for three rounds of label changes. Second, FigPad's terms state plainly: "We do not use Your Content to train our own models", and "you own the outputs generated for you", with publication rights included in every paid tier. Those are the two clauses Elsevier's policy tells you to go and verify — see pricing for the current tiers.
None of that overrides a journal's own rules. Check your target journal's instructions for authors, disclose what you used, and never present a generated figure as primary experimental data.
:::figure{size=large align=center}
BLOGFIGUREMARKER1END
:::
When you should stay with ChatGPT or Midjourney
There is a real job for both, and pretending otherwise would be silly.
Stay with ChatGPT when you want to draft matplotlib, TikZ or SVG code for a structural diagram or a data plot; when you want a second opinion on whether a draft figure reads clearly; when you need to write or tighten a figure legend; or when you are sketching layout options before committing to a tool. Its weakness at reading colour and quantitative detail in figures is documented — a 2024 study in npj Precision Oncology found it strong at plot recognition and notably weaker on colour perception — so treat its visual critique as a prompt for your own eyes, not a verdict.
Stay with Midjourney when the deliverable is not a manuscript figure: a talk opening slide, a lab website header, conference social graphics, a book-chapter concept illustration where you control the licensing. It is better at atmosphere than any scientific tool, and atmosphere is occasionally what you need.
Stay with neither for anything that goes into a submission package. That is not a quality judgement; it is a format and policy judgement.
Key Takeaways
- The blocker is structure, not beauty: general-purpose image models output pixels, and Nature requires "editable, unflattened vector artwork" with text left un-outlined.
- Resolution is borderline, not fatal — 1536 px gives 213 dpi at Nature's 183 mm double-column width and 438 dpi at 89 mm — but upscaling adds no information and no editability.
- Springer Nature journals do not permit AI-generated images for publication, with narrow labelled exceptions.
- Elsevier (policy updated June 2026) permits AI for explanatory diagrams with per-caption disclosure, bans it for primary research images, and specifically excludes general-purpose AI image tools from graphical abstracts.
- ChatGPT's best contribution to figures is code, not images — and no tool, dedicated or not, removes your obligation to check your target journal's policy and disclose.
