19 hours ago
- The author was initially amazed by DALL-E's outputs, feeling a sense of magic and discovery.
- Comparing old Midjourney v2 with a new model (Nano Banana Pro) shows that newer models are technically better at generating images but less capable of creating evocative, artistic work.
- Old models produced imperfect, vague, and contradictory images that sparked imagination and mystery, while new models are too precise and literal, resulting in dull, desaturated outputs.
- Specific examples (e.g., Electron Contours, Kowloon Walled City, Dream Garden) illustrate that older outputs, though less accurate, were more beautiful and compelling due to their coarse, impressionistic style.
- The new model often interprets prompts literally (e.g., 'from the British Museum' as a photograph) rather than evocatively, losing artistic nuance.
- The author concludes that new models' high precision and resolution prevent them from creating abstract, multi-interpretable art, and calls for making AI art weird again.