Which AI model fools people most?

Live ranking from real quiz answers. Genuine photographs are fixed at 1500 — anything above that line gets picked as “the real one” more often than reality does.

Limited to the models built into this quiz — not every model on the market, and not always the latest version of each. A model's absence here says nothing about how good it is.

Image models

29 models listed · 849 AI images in the quiz · click a column to sort

			Confidence
—	Real image	1500	—	4485	2455	55%	—
1	Nano Banana 2 with Web SearchGoogle	1500	1422–1578	80	31	39%	10
2	Ideogram 4.0 QualityIdeogram	1463	1398–1529	125	40	32%	10
3	GPT Image 2 MediumOpenAI	1417	1349–1484	125	35	28%	10
4	Nano Banana 2 LiteGoogle	1382	1302–1462	105	23	22%	10
5	Grok Imagine Image QualityxAI	1371	1260–1481	57	12	21%	10
6	Reve 2.1Reve AI	1345	1221–1468	52	9	17%	10
7	Qwen Image 2.0 ProAlibaba	1337	1250–1423	96	19	20%	10
8	Juggernaut Pro FluxRunDiffusion	1334	1298–1371	583	116	22%	33
9	Z-ImageAlibaba	1330	1308–1352	1579	307	20%	133
10	Gemini 2.5 Flash Image / Nano BananaGoogle	1325	1280–1371	347	70	20%	14
11	Imagen 2Google	1315	1279–1351	937	160	41%	29
12	FLUX.2 [klein] 9BBlack Forest Labs	1308	1208–1408	93	16	23%	14
13	Stable Diffusion 3.5 LargeStability AI	1296	1232–1360	199	34	17%	10
14	Seedream 4.5ByteDance	1295	1264–1326	871	145	17%	33
15	FLUX.2 [pro]Black Forest Labs	1282	1172–1392	75	11	15%	10
16	Kling Image V3Kuaishou	1281	1201–1362	138	21	15%	10
17	FLUX.1 [schnell]Black Forest Labs	1272	1250–1295	1920	273	14%	157
18	Stable DiffusionStability AI	1261	1210–1311	422	57	18%	37
19	GPT Image 1.5OpenAI	1257	1217–1296	651	87	13%	23
20	Juggernaut Flux LightningRunDiffusion	1253	1177–1329	173	23	13%	10
21	SDXL-LightningByteDance	1241	1197–1285	582	68	12%	72
22	Seedream 4ByteDance	1241	1162–1320	165	21	13%	13
23	FLUX.1 [dev]Black Forest Labs	1221	1180–1261	724	79	11%	41
24	DALL·E 3OpenAI	1203	1137–1269	299	29	10%	27
25	FLUX.2 [max]Black Forest Labs	1195	1125–1266	256	26	10%	10
26	Kling IMAGE 3.0 OmniKuaishou	1191	1137–1244	458	45	10%	24
27	Imagen 4 UltraGoogle	1186	1075–1297	113	10	9%	10
28	MAI-Image-2.5Microsoft AI	1156	988–1325	61	4	7%	10
29	Qwen-Image-MaxAlibaba	1101	997–1206	180	11	6%	12

Dimmed rows have fewer than 60 appearances — early evidence, not yet a settled rating.

A model joins this table once it has at least 10 images in the quiz and has appeared in at least 30 answers — below that, the number would say more about one lucky (or unlucky) image than about the model itself. New models and images are added regularly, so more entries — and firmer ratings — are on the way.

In the quiz but not yet listed: FLUX.2 LoRA Gallery: Realism (fal), FLUX1.1 [pro] (Black Forest Labs), Imagen 4 (Google), Lucid Realism (Leonardo AI), Titan Image Generator G1 v1 (Amazon) — fewer than 10 images in the quiz so far; Imagen 3 Fast (Google), Recraft V3 (Recraft), Seedream 5.0 Pro (ByteDance) — shown too rarely so far, under 30 answers.

Last recomputed 10 Aug 2026, 03:15 · refreshed nightly

Add your own answers

Every round you play feeds straight into these numbers.

Architecture and geometry

Paintings

Brushwork and composition

Cars

Reflections and badges

🎬 Video quizzes

Not in the ranking above yet — too few answers so far to rate video models.

Gravity, fire, liquids

How the rating works

Bradley-Terry, not a win count. Each round counts as a choice among the options actually on screen, so beating strong rivals is worth more than beating weak ones. Two-option rounds and four-option rounds therefore combine without their different odds distorting anything.

Read the range, not the position. Where two intervals overlap, the order between those models is not established. Dimmed rows have fewer than 20 appearances and are listed for completeness only.

Pool size is context. A model represented by three images is judged on those three images, however often they were shown.

Replays don't get extra weight. Only your first look at a round counts — after that you'd recognise it, and it stops being a fresh judgment.

Limits. Only answers from players who accepted analytics are counted, and the method assumes all players judge alike. This measures our image pool, not a model's ceiling.