A 7-phase professional workflow for creating publication-quality scientific illustrations using AI generation tools. Covers discovery & scoping, reference assembly, structural definition, component generation, post-processing, compositing, animation, and QA. Incorporates a 17-expert panel audit spanning diffusion models, prompt engineering, scientific accuracy, accessibility, and academic publishing.
Token counts approximate (OpenAI cl100k_base encoding, within ~15% of Claude). Pricing last verified 2026-04-12.
| Phase \ Failure | Single-shot generation | Style drift across components | Baked-in text labels | Unlicensed reference use | Cognitive overload | Pre-multiplied alpha fringing |
|---|---|---|---|---|---|---|
| Phase 1: Phase 0 — Discovery & Scoping | ○ | ○ | ○ | ○ | ○ | ○ |
| Phase 2: Phase 1 — Reference Library Assembly | ○ | ○ |
Define exactly what the illustration must communicate before any generation begins. Write the brief, set cognitive load budget, specify output constraints.
**Goal:** Define exactly what the illustration must communicate before any generation begins. **Steps:** 1. Write a one-sentence **illustration brief**: *"This diagram must show [subject] doing [proc…
Establish three distinct reference categories — Style DNA, Anatomy DNA, and Component References — before any generation runs.
**Goal:** Establish three distinct reference categories before any generation runs. | Reference Type | Purpose | How to Find | |---|---|---| | **Style DNA** | Locks visual language — colors, flatness…
Define the precise geometry/structure of the primary subject before any styled generation. Use wireframe, reference-constrained generation, SVG template, or open-source structural tools.
**Goal:** Define the precise geometry/structure of the primary subject before any styled generation. **Option A — Lineart/Wireframe (highest precision):** - Sketch or generate a simple wireframe (sha…
Generate each distinct element of the illustration in complete isolation. One generation pass per named component. Lock seed before generating.
**Goal:** Generate each distinct element of the illustration in complete isolation. > *If you can name a part of the diagram separately, it gets its own generation pass.* **Standard component list:*…
Prepare all components for compositing: background removal, alpha channel audit, edge cleanup, upscale, color correction, size normalization.
**Goal:** Prepare all components for compositing. **Standard pipeline per component:** 1. **Background removal** → transparent PNG (always required — do not skip) 2. **Alpha channel audit** → determi…
Assemble components into the final illustration using CSS Grid stacking (not absolute positioning). Labels always in HTML/CSS — never baked into image layers.
**Goal:** Assemble components into the final illustration. **Layer stack (bottom to top):** ``` Layer 8: Labels / annotations (always CSS/HTML, never baked into image) Layer 7: Interaction effects (g…
Add motion to communicate process or state change. Tier-based: CSS keyframes → GSAP → Lottie → Rive → AI image-to-video. Always implement prefers-reduced-motion.
**Goal:** Add motion to communicate process, change-of-state, or emphasis. **Animation tier selection:** | Tier | Technique | Best for | Effort | |---|---|---|---| | 1 | CSS Keyframes | Simple drift…
Validate against scientific accuracy, visual quality, accessibility, reproducibility, and publication/licensing standards before deployment.
**Goal:** Validate illustration against accuracy, accessibility, and publication standards. **Scientific Accuracy:** - [ ] Primary structural components verified against peer-reviewed source (cite th…
Treating AI image generation as a one-step process. Generates a "complete" illustration in one pass instead of isolating components. Result: style inconsistency, structural inaccuracies, and no component-level control.
Not establishing Style DNA before generation begins. Each component is generated independently without a locked visual language, producing illustrations where colors, stroke weights, and flatness levels vary across elements.
Including labels or annotations directly in the generation prompt. Labels baked into generated images cannot be corrected, localized, resized for publication, or made accessible. Always add labels in post-processing as HTML/CSS.
Using textbook figures or journal diagrams as anatomy DNA references without verifying the license is CC0 or CC-BY. "Educational fair use" does not protect AI-assisted derivative works. License verification must happen at reference collection time, not in QA.
Exceeding the cognitive load budget (≤5 elements for novice audiences; ≤9 for expert) without an information design justification. Adding animation does not exempt the diagram — animation adds cognitive load, it does not reduce it.
Displaying pre-multiplied alpha PNGs (common output from Pixa, Adobe tools) against light backgrounds causes color fringing at component edges. Must be caught at post-processing and fixed at source — not compensated for in CSS.
| ○ |
| ○ |
| ○ |
| ○ |
| Phase 3: Phase 2 — Structural Definition | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 4: Phase 3 — Component Generation | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 5: Phase 4 — Post-Processing Pipeline | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 6: Phase 5 — Compositing | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 7: Phase 6 — Animation (Optional) | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 8: Phase 7 — Quality Assurance | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
Write a structured illustration brief for the following: **Subject:** [REPLACE:subject] **Target audience:** [REPLACE:audience] **Publication context:** [REPLACE:publication_context] **Output format:** [REPLACE:output_format] Based on the above, produce: 1. **One-sentence brief:** "This diagram must show [subject] doing [process] for [audience] to understand [key insight]." 2. **Information hierarchy:** - Primary element (the most important visual focus): - Secondary elements (supporting structures): - Tertiary elements (context, labels, indicators): 3. **Cognitive load assessment:** Count all distinct visual elements planned. Budget: ≤5 for novice audiences, ≤9 for expert audiences. Total planned elements: [count] Decision: [ ] Within budget — proceed [ ] Over budget — split into sequence of [N] diagrams 4. **Output specification:** - Dimensions: [width] × [height] (or journal column spec) - Color mode: [ ] RGB (screen) [ ] CMYK (print) - DPI: [300 for print / 72–144 for screen] - Theme: [ ] Dark background [ ] Light background 5. **Scientific accuracy standard:** [ ] Publication-grade (peer-reviewed sources required for every structural claim) [ ] Educational (accurate but simplified for pedagogy) [ ] Conceptual/schematic (accuracy secondary to communication clarity) 6. **Reference collection checklist:** [ ] Style DNA reference identified (source: ___) [ ] Anatomy/Structure DNA identified and license verified: CC0 / CC-BY (source: ___) [ ] Component references identified (one per named element) [ ] All licenses logged in refs/attribution.md
Generate a positive + negative prompt pair for isolated component generation. **Component to generate:** [REPLACE:component_name] **Subject:** [REPLACE:subject] **Shape description:** [REPLACE:shape_description] **Color:** [REPLACE:color_description] **Style DNA reference:** [REPLACE:style_dna_description] **Anatomy DNA reference:** [REPLACE:anatomy_dna_description] --- **POSITIVE PROMPT:** [Style DNA reference image] flat style, dark background, color palette. [Anatomy DNA reference image] proportions and structure. Generate ONLY: [REPLACE:component_name] Shape: [REPLACE:shape_description] Color: [REPLACE:color_description] Background: solid dark navy Style: flat 2D vector, zero gloss, zero sheen, zero 3D highlights Isolation: [REPLACE:component_name] only, nothing else in frame Crop: element fills 80% of canvas **NEGATIVE PROMPT:** text, labels, watermarks, photorealistic, 3D render, glossy, specular, drop shadow, other biological structures, [REPLACE:elements_to_suppress], blurry, noisy, gradient (if unwanted) --- **Generation settings:** - Attach: attachments[0] = Style DNA reference, attachments[1] = Anatomy DNA reference - CFG scale: 7–9 - Lock seed BEFORE generating — record in generation log - Generate 2–3 variations, select best - Save sidecar JSON with prompt, seed, model, version, CFG, references, date
Refinement prompt for iterating on an existing component with Flux Kontext (or equivalent reference-preserving model). **Existing component to refine:** [REPLACE:component_name] **What to preserve:** [REPLACE:what_to_preserve] --- **REFINEMENT PROMPT:** Keeping [REPLACE:what_to_preserve], change ONLY: 1. [REPLACE:change_1] 2. [REPLACE:change_2] 3. [REPLACE:change_3] Do not modify: [REPLACE:locked_elements] --- **Notes for Flux Kontext:** - Flux Kontext is better at local edits than full style transformations - Do NOT use for initial generation from references — use for iterative refinement of an existing image - If more than 3 simultaneous changes are needed, split into sequential refinement passes - Re-lock seed per pass if reproducibility is required
Scientific animation prompt for AI image-to-video generation (Kling, Veo, Sora, etc.). **Remove all labels from the source still image before submitting to the video model.** --- **MOTION PROMPT (F.O.R.M.S. structure):** **Focus** (what object is moving): [REPLACE:moving_element] **Outcome** (what the motion communicates scientifically): [REPLACE:scientific_process_depicted] **Realism** (motion style): [REPLACE:realism_style] Options: organic / schematic / cinematic **Motion** (specific motion type): [REPLACE:motion_type] Examples: drift, expand, rotate, pulse, flow, divide, contract **Setting** (camera): [REPLACE:camera_setting] Options: static / slow zoom in / slow zoom out / gentle pan --- **Full prompt:** "Scientific animation of [REPLACE:subject]. [REPLACE:moving_element] [REPLACE:motion_type] [REPLACE:direction_and_speed]. Background and [REPLACE:static_elements] remain completely still. Style: [REPLACE:realism_style]. Loop seamlessly. Duration: 5–6 seconds. Camera: [REPLACE:camera_setting]." --- **Expect and budget for:** - Label drift (reason: labels were removed before generation — re-composite in CSS after) - Color saturation creep (especially Kling/Veo) - Edge ghosting on elements with transparent backgrounds **Export:** WebM (preferred) + MP4 fallback. Add `<track kind="descriptions">` if motion communicates scientific information.
Conduct a pre-publication QA review of this scientific illustration. **Illustration:** [REPLACE:illustration_name] **Subject:** [REPLACE:subject] **Target publication:** [REPLACE:publication_or_venue] **Scientific domain:** [REPLACE:domain] **Anatomy DNA reference used:** [REPLACE:anatomy_source_and_license] I will describe the illustration or provide the composite image. Review against these criteria: --- **1. SCIENTIFIC ACCURACY** - Does the information hierarchy (primary → secondary → tertiary elements) match the stated brief? - Are all spatial relationships (containment, direction, adjacency) scientifically accurate for this subject? - Are there any "plausibly wrong" elements — things that look correct but may misrepresent the actual structure? - Does the diagram avoid misleading visual metaphors? **2. COGNITIVE LOAD** - Count all distinct simultaneously visible elements. Is the total ≤9 (expert) or ≤5 (novice)? - Is there a clear primary focal point, or does the diagram compete for attention across multiple elements? **3. ACCESSIBILITY** - Is color the sole differentiator between any two elements? If so, flag which pairs. - Is there a meaningful alt text or accessible description for the composite? - Would labels remain legible at 7–8pt for print publication? **4. PUBLICATION READINESS** - Identify any elements that may require license verification before submission - Flag any missing AI generation disclosure information - Note any venue-specific requirements for [REPLACE:publication_or_venue] **5. RECOMMENDED CHANGES** List any required changes (blocking) and suggested improvements (non-blocking) before submission.