Table of contents
Open Table of contents
The Situation
I’ve always had a soft spot for public domain literature – those wonderful books from the 19th and early 20th centuries that are freely available but often overlooked. Many contain fantastic stories with illustrations that were remarkable for their time but feel dated now, or worse, have been lost entirely over the decades.
I wondered: could modern AI image generation tools breathe new life into these texts? Not to replace the originals, but to create fresh interpretations that might make these stories more accessible to contemporary readers.
The Challenge
The hype around AI image generation has been… considerable. Marketing promises aside, the practical reality of using these tools for a specific creative purpose – consistent illustration across an entire book – is more nuanced than “type a prompt, get art.”
The challenges I encountered:
- Consistency: Getting AI to maintain a consistent visual style across 20+ illustrations is harder than it sounds
- Accuracy: AI tools often misinterpret period-appropriate details – Victorian clothing, architectural styles, historical objects
- Resolution and quality: Raw AI output rarely meets print-quality standards without significant post-processing
- The “AI look”: There’s often a distinctive, slightly uncanny quality to AI-generated images that needs addressing
The Lesson
After experimenting with several approaches, I settled on a hybrid workflow that treats AI as a starting point rather than a finished product.
The Tools
DALL-E 3 became my primary generator. Its improved prompt-following compared to earlier versions meant I could describe specific scenes from the text and get results that at least resembled what I needed. The key was being extremely specific:
“Victorian gentleman in formal morning dress, standing in a London parlour, gaslight illumination, 1890s interior decoration, illustrated book style, fine line work”
Rather than:
“Old-fashioned man in a room”
Outpainting (extending images beyond their original boundaries) proved invaluable for creating full-page illustrations from initially smaller compositions. DALL-E can extend images in semantically consistent ways – adding appropriate backgrounds, extending scenes, filling in context.
Adobe Illustrator and Photoshop handled the refinement. This is where the real work happened:
- Correcting anatomical oddities (AI still struggles with hands)
- Adjusting colour palettes for print
- Adding consistent line weights
- Ensuring style coherence across illustrations
- Removing that telltale “AI shimmer”
The Workflow
- Extract key scenes from the source text
- Write detailed prompts including period-appropriate details, artistic style, and composition notes
- Generate multiple variations – typically 8-12 per scene – and select the best starting point
- Extend with outpainting where needed for composition
- Refine in Illustrator/Photoshop – typically 1-2 hours per illustration
- Style consistency pass across all illustrations to ensure cohesion
The Tactic
If you’re interested in trying something similar, here’s what I’d recommend:
Embrace the hybrid approach. Pure AI output isn’t publication-ready. Budget significant time for post-processing. In my workflow, AI generates maybe 30% of the final result; human refinement delivers the remaining 70%.
Build a style guide early. Before generating anything, document the visual style you’re aiming for: line weights, colour palette, level of detail, artistic influences. Reference this in every prompt.
Generate in batches. Work on similar scenes together. If you’re illustrating three parlour scenes, do them consecutively – you’ll develop prompt patterns that work and can maintain better consistency.
Save your prompts. Build a library of effective prompts for common elements: Victorian interiors, period clothing, natural scenes, character types. Reuse and refine these rather than starting from scratch each time.
Accept the limitations. AI tools struggle with: hands, complex spatial relationships, accurate historical details, consistent character faces across multiple images. Know these limitations and plan your post-processing accordingly.
Results and Reflections
I’ve now completed illustration refreshes for three public domain books using this approach. Is it faster than commissioning an illustrator? Marginally – perhaps 40% time savings, primarily in the initial concepting phase.
Is it cheaper? Yes, significantly, assuming your time has no value (it does, but side projects get special accounting treatment).
Is it creatively satisfying? Surprisingly, yes. The AI generates unexpected interpretations that I wouldn’t have conceived, and the refinement process involves genuine creative decision-making.
Is the output as good as skilled human illustration? No, let’s be honest. A talented illustrator working without AI constraints would produce superior results. But for side projects with limited budgets, the hybrid approach produces results that are genuinely usable and, I think, respectful to the source material.
Conclusion
AI image generation isn’t magic, and the “type prompt, receive masterpiece” narrative is marketing fiction. But as one tool in a creative workflow – particularly for projects that wouldn’t otherwise receive illustration at all – it has genuine utility.
The key insight: treat AI as a collaborator that provides raw material, not as a replacement for human creative judgment. The interesting work happens in the refinement, not the generation.
