Drafting publication figures: a one-week tool sweep
Scientific figures eat more of my paper-writing time than any other single task. I spent the last week pulling out tools to see what actually moves the needle and what is just hype.
The baseline I am benchmarking against
Before any tool changes I time-tracked a single methods figure for a small biophysics paper:
- 18 min sketching on paper
- 35 min in Illustrator (arrows, labels, grouping, hairlines)
- 22 min fixing alignment and colour after coauthor feedback
- 11 min exporting + checking 300 DPI and CMYK behaviour
That is ~90 min for one panel. A typical paper has 4–6 multi-panel figures. Easy to lose a full week.
What I tried this week
I ran the same methods figure through five tools and graded each on three axes:
| Tool | Time to first usable draft | Vector output | Realistic editing afterwards |
|---|---|---|---|
| Illustrator (control) | 90 min | yes | painful but possible |
| Inkscape | 110 min | yes | clunky but free |
| BioRender | 25 min | yes (paid) | locked into their library |
| ChatGPT image (gpt-image-1) | 4 min | no, raster only | impossible — pixel art |
| A research-focused research figure maker I had not seen before | 9 min | yes (SVG) | yes, components are real shapes |
The two that surprised me are the bottom two — and not for the reason you would expect.
Why text-to-image still fails for scientific figures
Image diffusion models are great for vibes, terrible for figures. I asked gpt-image-1 for "a methods schematic showing a microfluidic chip with three inlet channels feeding into a serpentine mixer, then into a detection chamber with a fluorescence probe". What I got:
- vaguely chip-shaped pixels
- three things that look like channels but two are connected wrong
- text labels that are unreadable hallucinations like "Inllt" and "Detctn Chambr"
- nothing is editable; it is one PNG
That last point kills it for academic use. Reviewers ask for changes. If your figure is a raster you cannot re-label, you have to redraw.
What worked for me
The thing that actually saved time was treating figure generation as a vector composition problem, not an image problem. The tool I linked above takes a prose description and emits structured SVG where each domain element (channel, inlet, label, scalebar) is a real grouped object you can edit afterwards. Same prompt as above, output was a clean 4-component schematic in ~9 minutes, fully editable in Illustrator or Inkscape afterwards.
This is the workflow I converged on:
- Write the figure caption first, in prose, the way I would in the paper.
- Feed that caption into the figure tool.
- Open the SVG in Illustrator only to nudge layout and fix the parts where the model misunderstood domain conventions.
- Export at 300 DPI.
Step 3 is still ~20 min of manual work but the starting point is the difference. Going from blank canvas to "almost right schematic with editable components" in under 10 minutes is a structural change to my paper workflow.
What I would not use it for
- Photorealistic biology illustrations — BioRender is still better for tissue-level art
- Network/graph layouts with thousands of nodes — Gephi or Cytoscape territory
- Anything where domain-specific notation matters (e.g. circuit schematics — KiCad, music notation — Lilypond)
What I am tracking next week
I want to see whether the same approach holds for two harder cases:
- Multi-panel figures with shared legends across 3+ subplots (a/b/c style)
- Methods schematics with quantitative inputs, like dose-response curves embedded inside a schematic
If either of those holds, I think a non-trivial chunk of methods-section figure work is just... gone. Will report back.
If you have a figure workflow that already works for you, I would love to hear it. Especially: how do you handle round 2 of reviewer feedback when the senior author wants to "just move the arrow"?
