Rapidata/vlm-vs-crowd

Can VLMs judge animal cuteness? People vs. gemini-3.8-flash on 325 image pairs

We showed people and a VLM the same 325 pairs of pet images and asked both the same question: "Which animal is cuter?" Thirty people answered each pair through Rapidata, and google/gemini-3.8-flash answered each pair 30 times through OpenRouter. This dataset is the complete raw material of that comparison: the image pairs, all 9,750 human votes, and all 9,750 VLM votes with the VLM's confidence and stated reason for each one.

The write-up is on our blog: Can VLMs Judge Animal Cuteness?

The setup

Six image models made pictures for 22 prompts about cats, dogs and rabbits, like "a cute kitten curled up in a basket": gpt-image-2, seedream-5-lite, p-image, nano-banana-2, flux-2-flex and krea-2-medium. Pairing up each prompt's images gives 325 pairs. Neither people nor the VLM saw the prompt an image was made from.

People answered in the Rapidata app, which shows each person the two images in a random order:

The VLM got these system and user messages, with the two images in place of the placeholders:

[system]
You are shown two images, A and B. Answer the question by picking one of them. You must always pick A or B; ties are not an allowed answer.

[user]
Image A:
<first image>
Image B:
<second image>
Which animal is cuter?

Answer with the label of the image you prefer, a confidence between 0 and 1, and a reason of at most 15 words. Reply with JSON only.

Each image of a pair came first in 15 of the VLM's 30 calls, so a lean toward the first image cancels out.

How the two compare

	People	gemini-3.8-flash
Who answered	4,448 annotator sessions, 103 countries	one model
Votes per pair	30	30
Total votes	9,750	9,750
Pairs where all 30 votes agreed	0	216
Pairs split 16–14 or closer	53	8

People split 15–15 on the two kittens below, and the VLM picked the sleeping one all 30 times:

The VLM voted 30–0 on 216 pairs. On 163 of them most people picked the same image, and on 23 most people picked the other one. The other 30 were close to even for people, 16–14 or closer. Five of those pairs:

The VLM picked the image it was shown first in 52.7% of its calls. Flipping the order changed its majority pick on 31 of the 325 pairs.

Schema

One row per image pair.

column	description
matchup_id	The pair's id, e.g. cat-01|flux-2-flex|krea-2-medium
prompt_id	The generation prompt's id, e.g. cat-01
generation_prompt	The prompt both images were made from. Neither people nor the VLM saw it.
model1, model2	The image models that made image1 and image2
image1, image2	The two images, embedded. The numbers are labels, not the order anyone saw them in.
votes_image1_people, votes_image2_people	People's votes per image (30 per pair)
votes_image1_vlm, votes_image2_vlm	The VLM's votes per image (30 per pair)
detailed_results_people	JSON list of every human vote: votedFor (image1/image2), annotatorId, userScore, country, language
detailed_results_vlm	JSON list of every VLM call: votedFor, shownFirst (image1/image2), replicate (0–29), confidence, reason, inputTokens, outputTokens, costUsd

import json

from datasets import load_dataset

ds = load_dataset("Rapidata/vlm-vs-crowd", split="train").remove_columns(["image1", "image2"])

# pairs where all 30 votes went to the same image
for side in ["people", "vlm"]:
    print(side, sum(min(r[f"votes_image1_{side}"], r[f"votes_image2_{side}"]) == 0 for r in ds))
# people 0, vlm 216

# how often the VLM picked the image it was shown first
calls = [c for r in ds for c in json.loads(r["detailed_results_vlm"])]
print(sum(c["votedFor"] == c["shownFirst"] for c in calls) / len(calls))  # 0.527

Notes on the data

Annotator ids are anonymized sequential ids (h0001, h0002, …). Votes with the same id came from the same annotator session, and the ids can't be linked back to Rapidata accounts. Each pair's 30 votes come from 30 different sessions. Three votes have no country. userScore is the annotator's Rapidata reliability score (the global userScore, from 0 to 1) on that vote.

The Rapidata app shows each person the two images in a random order, but which order a person saw isn't in this data. For the VLM, shownFirst records it. model1 and model2 come from the benchmark's pairing and aren't random: gpt-image-2 is model1 in 109 pairs and nano-banana-2 in none. So use shownFirst, not the image number, to study order effects.

The VLM run cost $21.54 through OpenRouter, about 0.2 cents per call.