AI Image Generation
SCIDraw's most powerful — and most misunderstood — feature. Read this page once before your first generation. It will save you time and prevent frustration.
Honest expectation setting: SCIDraw AI generates structured sketches, not pixel-perfect reproductions of paper figures. Think of it as a fast first draft — you'll always need to refine on the canvas. If you expect one-click paper-ready figures, you will be disappointed.
What AI generation actually does
When you type a description, the AI returns a structured drawing plan — a list of shapes (rectangles, polygons, paths, labels) — that SCIDraw renders as editable SVG on your canvas. This means:
- Every element is editable — move, recolor, resize, relabel. You're not stuck with a flat image.
- Style is "clean schematic" — think Nature summary figures, not photographically detailed art.
- Text labels are real text — searchable, copy-able, and ready for journal copyeditors.
What AI generation does NOT do well
Known limitations — do NOT expect these to work:
- Photorealistic biological illustrations — detailed cell renderings, anatomical cross-sections, realistic protein structures. Use the icon library instead.
- Perfect reproductions of existing paper figures — even with a reference image upload, AI interprets rather than copies. It will rearrange labels and simplify.
- Accurate molecular structures — SMILES rendering uses a dedicated chem tool (in the editor), not AI. AI-drawn molecules will have wrong bond angles.
- Complex multi-panel figures in one shot — generate each panel separately and compose them on the canvas.
- Chinese / non-English labels — the model is trained primarily on English. Use English labels in prompts, then translate text on canvas if needed.
Two generation modes
1. Text-to-figure
Type a description, get a figure. Best for: signaling pathways, experimental workflows, mechanism diagrams, schematic overviews.
2. Image-to-figure
Upload a reference image (paper figure, photo, hand sketch) and ask the AI to redraw it as an editable schematic. Best for: digitizing published figures you want to adapt, or cleaning up hand-drawn drafts.
Image-to-figure is interpretive, not photocopy. It captures layout and key elements but will simplify details. Always check the result against the original.
Effective prompt patterns
Real examples, tested against SCIDraw's current model (GLM-4.6V with thinking disabled for speed).
✅ Pattern 1 — Signaling pathway
Prompt
Draw a schematic of the PI3K/AKT/mTOR signaling pathway. Show receptor at top (EGFR), then PI3K, then AKT, then mTORC1 at bottom. Use arrows for activation (→) and T-bars for inhibition (⊣). Include feedback loop from mTORC1 back to AKT. Add protein names as labels and phosphorylation sites as small circles on proteins.
Verdict: ✅ Works well. Clean schematic output, all 4 proteins labeled, activation/inhibition arrows correct.
✅ Pattern 2 — Experimental workflow
Prompt
Create a horizontal flowchart of a Western blot protocol: 1) Cell lysis, 2) Protein quantification (BCA assay), 3) SDS-PAGE, 4) Transfer to PVDF, 5) Blocking, 6) Primary antibody, 7) Secondary antibody, 8) Detection. Each step in a rounded rectangle with the step name and a small icon. Connect with arrows.
Verdict: ✅ Works well. All 8 steps appear in sequence. Icons are schematic but recognizable.
⚠️ Pattern 3 — Mechanism diagram with realistic cells
Prompt
Draw a detailed realistic illustration of a neuron with myelin sheath, showing action potential propagation. Include Schwann cells, nodes of Ranvier, ion channels opening and closing.
Verdict: ⚠️ Partial. Layout is correct but the neuron will look like a schematic (rounded rectangles, simple shapes), not a realistic illustration. Use BioRender or Procreate for the realistic version.
❌ Pattern 4 — Bad prompt (too vague)
Prompt
draw a cell
Verdict: ❌ Will return a generic circle with maybe a nucleus. You haven't told the AI what kind of cell, what view, what organelles, what labels. Be specific.
Anatomy of a great prompt
- Subject: "PI3K/AKT signaling pathway" not just "pathway"
- Layout: "top-to-bottom" or "left-to-right" or "circular"
- Components: list every element you want by name
- Connections: specify activation (arrow), inhibition (T-bar), or modification (dashed)
- Labels: explicitly request labels — otherwise the AI may draw without any text
- Style hints (optional): "clean schematic", "minimal color", "black and white" — these do help
Generation time
Typical response time is 30–60 seconds. Complex prompts with many elements can take up to 2 minutes. If you don't see a result after 3 minutes, refresh the page and try again — long timeouts usually mean the model is overloaded.
Quota
- Free users: 3 AI generations per day
- VIP (any plan): unlimited generations, subject to fair-use rate limiting (no burst-abuse, no automated scripting)
Best practice: combine AI with the icon library. Use AI to generate the overall layout and rough shapes, then drag in specific high-quality icons (a detailed mitochondrion, a labeled ribosome) from the 27,000+ library to replace the AI's schematic versions. This is how professional figures are built.
When AI fails — common causes
- "SVG fallback" or "cache fallback" in your generation log: the model timed out or returned an invalid response. Retry with a simpler prompt.
- Labels missing or gibberish: the prompt didn't explicitly request labels. Add "with English labels for every component" to your prompt.
- Wrong biological relationship: AI doesn't verify biology. Always cross-check activation/inhibition direction against your source.
- Icons missing in the result: you asked for icons the AI can't draw (e.g., specific protein structures). Add them manually from the icon library.
Found a generation that's clearly broken? Email
support@scidrawio.com with your prompt and the resulting figure. We treat quality issues case-by-case — first-time subscribers within 14 days may request a refund (see
Terms §3).