---
schema: 1
type: term
title: Pelican Bench
description: >-
  **Pelican Bench** is a benchmark for image generation models built around a
  deliberately absurd prompt — render a pelican riding a bicycle — whose
  rendered details (leg placement, neck thickness, bicycle frame) serve as a
  discriminating probe of model quality.
canonical: 'https://liquidgold.cc/channel/everyinc/explore/term/pelican-bench-c6e9b73beb2d'
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/pelican-bench-c6e9b73beb2d.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# Pelican Bench

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/pelican-bench-c6e9b73beb2d

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

**Pelican Bench** is a benchmark for image generation models built around a deliberately absurd prompt — render a pelican riding a bicycle — whose rendered details (leg placement, neck thickness, bicycle frame) serve as a discriminating probe of model quality.

**Pelican Bench** is a benchmark for evaluating image generation models, built around a deliberately strange prompt: render a *pelican riding a bicycle* and then inspect the details. Its creator, Simon, describes it as the strangest and most ridiculous reference point — and yet one that offers real value for exploring AI progress.Simon joined the DevDay 2026 live stream with the bench in hand, and it got a workout against the new **Sol 6.1** model:- The rendered pelican came out well enough, but the bicycle was, in Simon's words, “garbage” — a **surprisingly deformed frame**.- Comparing across versions suggested a pattern: **Sol 5.6** produced the right all-around bike shape, and quality dropped below that. But as Simon noted, you “roll the dice every time” — one bad render doesn't settle the comparison.- The test had run at the **default thinking level**, which hadn't been checked beforehand — a reminder that benchmark-style probes carry confounds worth exploring.The pelican test in miniature: ask a model to render a pelican on a bicycle, then look for legs on both sides of the frame, a neck that isn't too thick, and a coherent bicycle. Small, absurd, and surprisingly discriminating.- **Pelican Bench** is an image-model benchmark built on the absurd *pelican riding a bicycle* prompt.- It probes fine details — leg placement, neck thickness, bicycle frame coherence — that reveal quality differences between models.- Results are stochastic: every render is a fresh roll of the dice, so a single bad output doesn't prove a model is worse.