Coarse is Better: why new image models make better images but worse art than DALL-E era models
zetalyrae · x · 2026-09-10
A widely shared essay compares the same prompts run through 2022-era DALL-E/Midjourney v2 and the new Nano Banana Pro, arguing models got far better at making images but 'infinitely worse at making art.'
- Old models' imperfections, vagueness and contradictions created 'gaps through which your imagination can breathe life into the art' — MJ's Kowloon Walled City alley misses the subject but is coarse, impressionistic and full of mystery.
- New models nail the style references yet produce muted, desaturated outputs that 'look like every modern movie.'
- Even explicitly prompting for coarse, evocative, contradictory qualities fails to restore the magic.
The essay asks what AI art is really optimizing for — and whether correctness is killing what made early outputs compelling.
More from Multimodal
- Hands-on with H3 Director: AI video that auto-fits your prompts to the music — jfischoff · 2026-09-10
- MiniMax H3 Director on fal streams realtime video you direct live with prompts, at $0.02/s promo pricing — jfischoff · 2026-09-10
- Gradium TTS lands on LiveKit Inference with voice cloning, sub-250ms latency, free until Oct 9 — alexcovo_eth · 2026-09-10
- Runway adds GPT-Image 2.5 with fast Flare mode and high-fidelity Sunburst editing — runwayml · 2026-09-10
- Generic AI music is a skill gap: expert-level prompts unlock pro results — _Stocko_ · 2026-09-10
- SimWorld Teases Code4Scene Benchmark: Prompt-to-Explorable-3D Container Port Demo — Lianhuiq · 2026-09-10