---
license: other
license_name: unknown-community-meme-benchmark-use-only
task_categories:
- text-to-image
- image-to-image
tags:
- benchmark
- image-generation-evaluation
- vision-language-judge
- frog
- meme
- bufo
pretty_name: bufo-bench
configs:
- config_name: default
  data_files:
  - split: tasks
    path: data/tasks.jsonl
  - split: evaluations
    path: data/evaluations.jsonl
---

# 🐸 bufo-bench

Can your image model draw a **real bufo** — or does it draw a generic cute frog
and call it a day?

Bufo is a riff on the original **"worry frog" / froge** meme: simplex smooth
construction, fresh leaf-green, LARGE pale-yellow eyes with huge near-black
pupils in a worried *averted* gaze, and an iconic little nose (protruding snout,
tiny nostrils, **no mouth**). Image models reliably fail at this — they draw
bright-green tree-frogs with stalk eyes and big smiles instead.

## What's here

- **`data/tasks.jsonl`** (5 rows) — the eval set. Each row: `task_id`, `title`,
  `prompt` (task instruction), `checks` (verifiable checklist), `style_block`
  (shared bufo spec prepended to every generation prompt), `reference_images`,
  `notes`. Run a model on all five with the `style_block` + references, then
  judge.
- **`references/froge-original.png`** — ground truth: the original worry-frog
  (via [KYM photo #2561104](https://knowyourmeme.com/photos/2561104-froge-worry-frog)).
- **`references/negative-*.png`** — documented off-model riffs (what "close but
  wrong" looks like: olive-drab, deadpan, MS-paint).
- **`data/evaluations.jsonl`** (15 rows + `evaluations/<run>/<task>.png`) —
  every judged output so far: model, provider, track, per-axis scores
  (`shape`/`color`/`face`/`style`, 0–2 each), `bufoness` (0–8), `task_score`,
  judge verdict, judge model, rubric version, and `expert_verdict` where a human
  expert graded it (null = judge only).

## Scoring (see the bench repo for the full rubric)

`bufoness` = shape + color + face + style. **PASS** iff bufoness ≥ 6 AND all
task checks hold AND no axis is 0. Model score = pass rate over the 5 tasks.
The load-bearing discriminator is the **gaze test**: a frog that looks AT you
is not worried — worry means averted, half-lidded, or closed eyes with no mouth.

## Provenance & license

Reference art is community meme culture of unclear authorship (see SOURCES in the
bench repo); model outputs belong to whoever generated them. Provided for
benchmark/evaluation use. If you hold rights to the original froge and want it
handled differently, open an issue on the bench repo.

Bench code, runner, judge prompts, and eval queue:
local `bufo-bench` project (being open-sourced separately).