AI summary by TwinMind
AI-Generated Creatures That Stretch the Boundaries of Imagination | Sofia Crespo | TED
TED
9m 16s
Introduction: Boundaries of Human Imagination
- Challenge: Imagine color never seen before - impossible without drawing from experience
- Colors outside visible spectrum exist but can't be visualized
- Boundary to imagination - can only imagine shades of previously seen colors
- Human vision limitations:
- Humans have 3 photoreceptors
- Not all humans see same way (colorblind variations)
- Disagreement on simple things (blue/black vs white/gold dress)
- Peacock mantis shrimp: 12-16 photoreceptors, much more colorful world perception
AI's Role in Expanding Creative Boundaries
- AI can see within data it's fed, can't imagine truly new things
- Value in creating combinations of known elements to form new ones
- Artist's 5-year AI experience: AI augments creativity rather than replaces it
- Boundary of imagination based on experience = exploration area
Neural Zoo Series Development
- Origin: Jellyfish on aquarium screen with old 3D glasses (blue/red lens)
- Goal: Recreate textures + create new jellyfish never seen before
- Evolution: Jellyfish → sea creatures (anemoni, coral, fish) → amphibians → birds → insects
- Series name: "Neural Zoo"
- Key insight: No single creature in images - AI distills and recombines textures
- Process advantage: Months of hand drawing compressed, overcomes poor drawing skills
Historical Context: Medieval Illustration Parallels
- Pre-camera era: Expeditions → explorers describe to illustrators
- Illustrators never saw described creatures, drew based on known animals
- Result: Hybrid animals (beaver described → rodent head + dog body + fish tail)
- Contemporary parallel to AI-assisted creation process
Artificial Natural History Series
- Process: Thousands of natural history archive illustrations fed to neural network
- Output: New versions of historical illustrations generated
- Transition from 2D to 3D work
3D Beetle Generation Project
- Collaboration with studio partner Felican McCormick
- Training: Neural network on 3D scanned beetles dataset
- Initial results: Extremely blurry blobs
- Challenges:
- Limited openly available 3D insect datasets
- Repurposing image-generation networks for 3D
- Solutions: Data augmentation (added ants, beetle-like insects)
- Advanced techniques:
- 3D style transfer for texture mapping
- Natural language model for scientific names/anatomical descriptions
- Network architecture handling 3D meshes
- Final result: Speculative biology study of non-existent creatures
Endangered Species Visibility Project
- Problem: Data availability disparity (red panda vs Boreal for liking)
- Red panda: Easier to find online data (cuteness factor)
- Both species: Classified as endangered
- Goal: Bring visibility to less digitally represented endangered species
AI-Generated Endangered Species Examples
- Training: AI on millions natural world images
- Text prompting process:
- Peacock tarantula (critically endangered spider): AI generated vs real comparison
- Mangrove Finch (critically endangered bird): AI generated vs real comparison
- Key limitation: Generated image accuracy depends on available data
Future Vision and Conclusion
- AI-generated creatures: Different way of envisioning future
- Not literal prediction but imagination expansion about ecosystems
- Better equipped to recognize new opportunities/potential
- Boundary awareness as motivation rather than limitation
- Goal: Expand imagination boundary, seek unseen colors/things
- Result: Enriched imagination through exploration
