The Burrito Test: one prompt stress-tests 25 image generation models side by side
generativist · x · 2026-11-17
Generativist updates the 'Generative Burrito Test,' a fun benchmark inspired by the horse-riding-astronaut meme and Simon's pelican benchmark. A single prompt — a partially eaten burrito with seven listed ingredients — is run through 25 models (SD 1.5/3.5, Flux family, Ideogram, Imagen 4, Qwen Image, Nano Banana Pro, Seedream V4, Hunyuan V3, etc.) via fal.ai defaults. He found the smushed-and-congealed-ingredients concept surprisingly hard despite abundant training data, and deliberately avoids heavy prompt engineering as 'cheating.'
More from Fun
- Nvidia board says no Jensen Huang return, odds crash to 6% on Kalshi — kevinnbass · 2026-11-17
- KOL predicts AI agent team meetings will replace human meetings within a year — kimmonismus · 2026-11-17
- Asking ChatGPT and Claude to play NeurIPS reviewers — they reject the papers — burny_tech · 2026-11-17
- "Everything happens all the time": AI circle's collective sigh — burny_tech · 2026-11-17
- 80s-style fantasy AI video dubbed better than movies — mark_k · 2026-11-17
- Neither has used Isabelle: a theorem-prover flame war in two lines — zetalyrae · 2026-11-17