Rubik's Cube benchmark results

Prompt: “A valid scrambled standard 3x3 Rubik's Cube on a #7f7f7f background”

Backend fal, runware · 10 images per model · seeds random · extractor vision · evaluator 0.7.0 · run started 2026-10-20T20:51:20.876532+00:00 · scored 2026-10-20 21:03:30 UTC

Made with RubiksBench 0.7.0: the benchmark's code, method and how to run it.

Scores

Validity says how close an image comes to a valid, scrambled, standard 3x3 cube: the fewest stickers you would have to recolor to get one (plus any sticker no recoloring fixes, such as one with letters or two colors). 100% is a valid cube, and every 2.2 stickers to fix halve the score, so a cube that still fools someone never reaches 0%. Coloring the stickers at random lands at 25% (it needs about 4.4 recolorings, because the three hidden faces leave a lot of freedom). A barely scrambled cube or a designed pattern is held in slightly less regard: its validity is multiplied by a scramble factor between 95% and 100%. 0% is for images that fool no one: not read as a 3x3 cube, not scrambled at all, or with a flaw anyone sees at a glance.

Who it fools says who would be taken in by each image: every failed check is one that anyone, at a glance, anyone who has held a cube, a cuber looking carefully or only a solver would notice, and an image fools everyone below the first who notices. An image that passes every check fools everyone.

The rubric explains what is wrong: up to 17 points, 1 per check passed, ½ when inconclusive, 0 when failed or blocked (see Checks). An image passes only if it passes all checks. Scores are over generated images; requests that returned no image are counted separately as generation errors.

ModelValidityWho it foolsRubricImagesOutcomesCost
Reference photos (real cubes) reference
real photo
100%
100%–100%
validity 100%, 95% CI 100%–100%fools everyone: 3 of 317.0/173
3 passed
Valid (standard): 3 of 3–
0s each
Qwen-Image-3.0
alibaba:qwen-image@3.0
48%
36%–60%
validity 48%, 95% CI 36%–60%fools a casual viewer: 7 of 10fools a cube owner: 3 of 1012.4/1710
0 passed
Invalid: 10 of 10$0.300
66s each
Grok Imagine Image 2.0
xai:grok-imagine@image-2.0
42%
30%–55%
validity 42%, 95% CI 30%–55%fools a casual viewer: 5 of 10fools a cube owner: 5 of 1011.6/1710
0 passed
Invalid: 10 of 10$0.600
76s each
Qwen-Image-3.0-Pro
alibaba:qwen-image@3.0-pro
41%
31%–52%
validity 41%, 95% CI 31%–52%fools a casual viewer: 8 of 10fools a cube owner: 2 of 1012.4/1710
0 passed
Invalid: 10 of 10$0.400
53s each
Krea 3
krea/v3/text-to-image
41%
31%–51%
validity 41%, 95% CI 31%–51%fools a casual viewer: 6 of 10fools a cube owner: 3 of 10fools a cuber: 1 of 1012.4/1710
0 passed
Valid (non-standard): 1 of 10Invalid: 9 of 10–
20s each
Nano Banana 2.1
google:nano-banana@2.1
39%
33%–45%
validity 39%, 95% CI 33%–45%fools a casual viewer: 9 of 10fools a cube owner: 1 of 1011.4/1710
0 passed
Valid (non-standard): 1 of 10Invalid: 9 of 10$0.336
13s each
GPT Image 2
openai:gpt-image@2
38%
32%–44%
validity 38%, 95% CI 32%–44%fools a casual viewer: 7 of 10fools a cube owner: 3 of 1011.3/1710
0 passed
Invalid: 10 of 10$0.060
19s each
FLUX 3 Image
bfl:flux@3-image
37%
17%–57%
validity 37%, 95% CI 17%–57%fools no one: 3 of 10fools a casual viewer: 3 of 10fools a cube owner: 3 of 10fools everyone: 1 of 109.8/1710
0 passed
Valid (standard): 1 of 10Valid (non-standard): 1 of 10Invalid: 8 of 10$0.480
55s each
FLUX.2 [pro]
bfl:5@1
36%
29%–44%
validity 36%, 95% CI 29%–44%fools a casual viewer: 8 of 10fools a cube owner: 2 of 1010.8/1710
0 passed
Invalid: 10 of 10$0.300
11s each
Nano Banana 2 (Gemini 3.1 Flash Image)
google:4@3
36%
27%–44%
validity 36%, 95% CI 27%–44%fools a casual viewer: 5 of 10fools a cube owner: 5 of 1011.5/1710
0 passed
Valid (non-standard): 1 of 10Invalid: 9 of 10$0.690
14s each
Seedream 5.0 Pro
bytedance:seedream@5.0-pro
32%
22%–41%
validity 32%, 95% CI 22%–41%fools a casual viewer: 4 of 10fools a cube owner: 6 of 1011.1/1710
0 passed
Invalid: 10 of 10$0.481
39s each
GPT-Image-2.5 Flare
openai:gpt-image@2.5-flare
29%
24%–35%
validity 29%, 95% CI 24%–35%fools a casual viewer: 9 of 10fools a cube owner: 1 of 1011.2/1710
0 passed
Valid (non-standard): 1 of 10Invalid: 9 of 10$0.060
13s each
Nano Banana Pro
google:4@2
29%
17%–41%
validity 29%, 95% CI 17%–41%fools a casual viewer: 6 of 10fools a cube owner: 2 of 10not fully checked: 2 of 109.5/1710
0 passed
Invalid: 8 of 10Inconclusive: 2 of 10$1.380
32s each
Luma UNI-1
luma:uni@1
29%
22%–36%
validity 29%, 95% CI 22%–36%fools a casual viewer: 7 of 10fools a cube owner: 3 of 1011.2/1710
0 passed
Invalid: 10 of 10$0.404
93s each
Krea 2 Large
krea:krea@2-large
29%
20%–37%
validity 29%, 95% CI 20%–37%fools a casual viewer: 9 of 10fools a cube owner: 1 of 1010.7/1710
0 passed
Invalid: 10 of 10$0.600
42s each
Ideogram 4.5
ideogram:4.5@0
28%
18%–37%
validity 28%, 95% CI 18%–37%fools no one: 1 of 10fools a casual viewer: 8 of 10fools a cube owner: 1 of 109.8/1710
0 passed
Invalid: 10 of 10$2.000
13s each
Seedream 5.0 Flash
bytedance:seedream@5.0-flash
26%
18%–35%
validity 26%, 95% CI 18%–35%fools a casual viewer: 10 of 1010.7/1710
0 passed
Invalid: 10 of 10$0.193
19s each
GPT-Image-2.5 Sunburst
openai:gpt-image@2.5-sunburst
25%
20%–30%
validity 25%, 95% CI 20%–30%fools a casual viewer: 8 of 10fools a cube owner: 2 of 1011.0/1710
0 passed
Invalid: 10 of 10$0.060
18s each
HiDream O1
fal-ai/hidream-o1-image
22%
19%–25%
validity 22%, 95% CI 19%–25%fools a casual viewer: 9 of 10fools a cube owner: 1 of 1010.4/1710
0 passed
Invalid: 10 of 10–
22s each
FLUX.2 [dev]
runware:400@1
21%
16%–27%
validity 21%, 95% CI 16%–27%fools a casual viewer: 10 of 1010.2/1710
0 passed
Invalid: 10 of 10$0.120
7s each
Kling IMAGE O3
klingai:kling-image@o3
21%
12%–30%
validity 21%, 95% CI 12%–30%fools no one: 1 of 10fools a casual viewer: 4 of 10fools a cube owner: 3 of 10not fully checked: 2 of 108.2/1710
0 passed
Invalid: 8 of 10Inconclusive: 2 of 10$0.280
29s each
Kling IMAGE 3.0
klingai:kling-image@3
17%
8.1%–27%
validity 17%, 95% CI 8.1%–27%fools no one: 1 of 10fools a casual viewer: 5 of 10fools a cube owner: 2 of 10not fully checked: 2 of 107.8/1710
0 passed
Invalid: 8 of 10Inconclusive: 2 of 10$0.280
35s each
MOSS-Image
fal-ai/moss-image
17%
9.5%–25%
validity 17%, 95% CI 9.5%–25%fools no one: 1 of 10fools a casual viewer: 7 of 10fools a cube owner: 1 of 10not fully checked: 1 of 109.1/1710
0 passed
Invalid: 9 of 10Inconclusive: 1 of 10–
20s each
ERNIE-Image-Turbo
baidu:ernie-image@turbo
13%
7.2%–20%
validity 13%, 95% CI 7.2%–20%fools no one: 2 of 10fools a casual viewer: 7 of 10fools a cube owner: 1 of 109.4/1710
0 passed
Invalid: 10 of 10$0.043
28s each
ERNIE-Image
baidu:ernie-image@0
12%
4.0%–20%
validity 12%, 95% CI 4.0%–20%fools no one: 5 of 10fools a casual viewer: 5 of 106.1/1710
0 passed
Invalid: 10 of 10$0.206
38s each
Z-Image
runware:z-image@0
12%
0.8%–23%
validity 12%, 95% CI 0.8%–23%fools no one: 4 of 10fools a casual viewer: 5 of 10not fully checked: 1 of 109.2/1710
0 passed
Invalid: 9 of 10Inconclusive: 1 of 10$0.099
21s each
Recraft V4.1
recraft:v4.1@0
11%
1.3%–21%
validity 11%, 95% CI 1.3%–21%fools no one: 4 of 10fools a casual viewer: 6 of 109.2/1710
0 passed
Invalid: 10 of 10$0.350
8s each
MiniMax image-01
fal-ai/minimax/image-01
10%
1.7%–19%
validity 10%, 95% CI 1.7%–19%fools no one: 3 of 10fools a casual viewer: 5 of 10not fully checked: 2 of 106.1/1710
0 passed
Invalid: 8 of 10Inconclusive: 2 of 10–
28s each
Seedream 5.0 Lite
bytedance:seedream@5.0-lite
10%
4.4%–16%
validity 10%, 95% CI 4.4%–16%fools no one: 4 of 10fools a casual viewer: 6 of 107.2/1710
0 passed
Invalid: 10 of 10$0.350
52s each
Ideogram 4.0
ideogram:4@0
9.4%
1.5%–17%
validity 9.4%, 95% CI 1.5%–17%fools no one: 4 of 10fools a casual viewer: 4 of 10not fully checked: 2 of 105.7/1710
0 passed
Invalid: 9 of 10Inconclusive: 1 of 10$0.600
18s each
Luma Photon
fal-ai/luma-photon
8.9%
0%–19%
validity 8.9%, 95% CI 0%–19%fools no one: 6 of 10fools a casual viewer: 2 of 10fools a cube owner: 2 of 106.4/1710
0 passed
Valid (non-standard): 1 of 10Invalid: 9 of 10–
19s each
Hunyuan Image 3.0
fal-ai/hunyuan-image/v3/text-to-image
6.9%
0%–18%
validity 6.9%, 95% CI 0%–18%fools no one: 5 of 10fools a casual viewer: 1 of 10fools a cube owner: 1 of 10not fully checked: 3 of 103.9/1710
0 passed
Invalid: 7 of 10Inconclusive: 3 of 10–
20s each
Stable Diffusion 3.5 Large
fal-ai/stable-diffusion-v35-large
6.2%
0%–17%
validity 6.2%, 95% CI 0%–17%fools no one: 6 of 10fools a casual viewer: 3 of 10not fully checked: 1 of 104.3/1710
0 passed
Invalid: 9 of 10Inconclusive: 1 of 10–
12s each
Recraft V4.1 Flash
recraft:v4.1@flash
5.1%
0%–10%
validity 5.1%, 95% CI 0%–10%fools no one: 4 of 10fools a casual viewer: 4 of 10not fully checked: 2 of 106.9/1710
0 passed
Invalid: 9 of 10Inconclusive: 1 of 10$0.070
3s each
Bria FIBO
bria:20@1
4.2%
0%–10%
validity 4.2%, 95% CI 0%–10%fools no one: 8 of 10fools a casual viewer: 2 of 103.8/1710
0 passed
Invalid: 10 of 10$0.400
14s each
Z-Image-Turbo
runware:z-image@turbo
2.7%
0.8%–4.6%
validity 2.7%, 95% CI 0.8%–4.6%fools no one: 1 of 10fools a casual viewer: 8 of 10not fully checked: 1 of 108.2/1710
0 passed
Invalid: 9 of 10Inconclusive: 1 of 10$0.022
5s each
HiDream-I1 Full
runware:97@1
2.7%
0%–6.2%
validity 2.7%, 95% CI 0%–6.2%fools no one: 4 of 10fools a casual viewer: 2 of 10not fully checked: 4 of 103.5/1710
0 passed
Invalid: 6 of 10Inconclusive: 4 of 10$0.238
16s each
FLUX 3 Image Klein
bfl:flux@3-image-klein
0%
0%–0%
validity 0%, 95% CI 0%–0%fools no one: 9 of 10fools a casual viewer: 1 of 104.2/1710
0 passed
Invalid: 10 of 10$0.060
6s each
fools no onefools a casual viewerfools a cube ownerfools a cuberfools everyonenot fully checked
Valid (standard)Valid (non-standard)InvalidInconclusiveGeneration failed

Points per check

Share of the available points each model earned on each check. A check that cannot be decided (e.g. sticker colors on a face that is not 3x3) is blocked and scores 0.

Model1. Cube detected2. Cube shape3. 3x3 grid4. Cubies in place5. Plain stickers6. Single-colored stickers7. At most six colors8. One shade per color9. Standard palette10. Standard color pairs11. Scrambled12. Distinct centers13. Standard center order14. Possible pieces15. Unique pieces16. Solvable17. Solvable (standard)
Reference photos (real cubes)100%100%100%100%100%100%100%100%100%100%100%100%100%100%100%100%100%
Qwen-Image-3.0100%100%100%100%100%100%100%100%100%40%90%100%90%20%0%0%0%
Grok Imagine Image 2.0100%100%100%100%100%100%100%100%100%30%100%90%20%20%0%0%0%
Qwen-Image-3.0-Pro100%100%100%100%100%100%100%100%100%40%100%100%80%20%0%0%0%
Krea 3100%100%100%100%100%100%100%100%100%50%100%100%60%10%10%10%0%
Nano Banana 2.1100%100%100%90%100%100%100%100%100%40%100%60%20%10%10%10%0%
GPT Image 2100%100%100%100%100%100%100%100%100%20%100%70%40%0%0%0%0%
FLUX 3 Image100%100%80%80%80%80%80%80%70%30%60%60%10%20%20%20%10%
FLUX.2 [pro]100%100%100%100%100%100%100%100%100%10%100%40%30%0%0%0%0%
Nano Banana 2 (Gemini 3.1 Flash Image)100%100%100%100%100%100%100%100%100%10%90%70%20%40%10%10%0%
Seedream 5.0 Pro100%100%100%100%100%100%100%80%100%20%100%80%10%20%0%0%0%
GPT-Image-2.5 Flare100%100%100%100%100%100%100%100%100%20%100%70%0%10%10%10%0%
Nano Banana Pro100%90%90%80%80%80%80%80%80%10%80%70%30%0%0%0%0%
Luma UNI-1100%100%100%100%100%100%95%100%95%30%100%70%10%25%0%0%0%
Krea 2 Large100%100%100%100%100%100%100%90%100%20%90%60%10%0%0%0%0%
Ideogram 4.5100%100%90%90%90%90%90%90%90%0%90%40%20%0%0%0%0%
Seedream 5.0 Flash100%100%100%100%100%100%100%100%100%10%80%60%20%0%0%0%0%
GPT-Image-2.5 Sunburst100%100%100%100%100%100%100%100%100%10%100%60%20%10%0%0%0%
HiDream O1100%100%100%100%100%100%100%80%100%20%100%40%0%0%0%0%0%
FLUX.2 [dev]100%100%100%100%100%100%100%80%100%30%100%10%0%0%0%0%0%
Kling IMAGE O3100%90%85%70%70%70%70%70%70%10%70%40%0%10%0%0%0%
Kling IMAGE 3.0100%90%80%70%70%70%70%40%70%10%70%40%0%0%0%0%0%
MOSS-Image100%95%80%80%80%80%80%80%80%20%80%40%0%10%0%0%0%
ERNIE-Image-Turbo100%100%100%100%100%100%80%30%80%20%70%50%0%10%0%0%0%
ERNIE-Image100%100%50%50%50%50%50%40%50%0%30%30%10%0%0%0%0%
Z-Image100%100%100%90%100%90%95%80%60%40%5.0%50%5.0%0%5.0%0%0%
Recraft V4.1100%100%90%90%90%90%55%90%45%50%70%50%0%0%0%0%0%
MiniMax image-01100%90%60%50%50%50%50%20%50%10%40%30%10%0%0%0%0%
Seedream 5.0 Lite100%100%60%60%60%60%60%40%60%0%60%50%10%0%5.0%0%0%
Ideogram 4.0100%95%50%40%40%40%40%40%40%20%40%20%0%0%0%0%0%
Luma Photon100%100%60%50%40%40%40%30%40%20%40%30%0%30%10%10%0%
Hunyuan Image 3.0100%85%35%20%20%20%20%10%20%10%20%20%0%10%0%0%0%
Stable Diffusion 3.5 Large100%100%35%30%30%30%30%20%30%0%10%10%0%0%5.0%0%0%
Recraft V4.1 Flash100%100%70%60%60%60%55%40%35%10%40%50%10%0%0%0%0%
Bria FIBO100%100%20%20%20%20%20%20%20%15%20%0%0%0%0%0%0%
Z-Image-Turbo95%100%90%80%80%80%80%80%80%30%10%10%0%0%0%0%0%
HiDream-I1 Full100%80%35%20%20%20%20%0%20%0%20%10%0%0%0%0%0%
FLUX 3 Image Klein100%100%30%30%30%30%20%20%10%20%20%10%0%5.0%0%0%0%

Hall of Fame

The most valid generated images, one per model (real photos left out).

Hall of Shame

The least valid images, one per model. Only images read as invalid count: an inconclusive one may be our reader's failure rather than the model's.

Samples

Each model's best image by default; show more per model, filter, or page through every attempted image below. Each card shows its validity and rubric points, and which stickers to recolor for the nearest valid cube. The squares are the 17 checks in order; hover for names. Where a failed check has a clear culprit (a duplicate center, an impossible piece, a face that is not 3x3, ...) it is outlined on the image and listed below it. Twist, flip and parity failures belong to the whole cube, so they are not outlined.

PassInconclusiveFailBlockedproblem outline

Failure flags

Images carrying each flag (one image can carry several).

Checks

#CheckPasses whenDecided when read
1Cube detectedA cube-shaped object is found in the image.–
2Cube shapeThree faces form a consistent corner view of a cube.a cube outline
33x3 gridEvery visible face has 3 columns and 3 rows of cubies.three faces
4Cubies in placeEvery cubie sits where a real cube puts it: none sunk in, sticking out or deformed on its own. Found by posing an ideal cube to the image; a lone cell that is mostly not sticker color holds such a cubie.the image read as a 3x3 cube
5Plain stickersStickers carry no letters or other markings (a logo on a center is allowed).the image read as a 3x3 cube
6Single-colored stickersNo sticker shows two clearly different colors (a logo on a center is allowed; a split between neighbouring hues such as red and orange is not detected).the image read as a 3x3 cube
7At most six colorsNo more than six distinct sticker colors. Fewer is allowed: every sticker of a color can sit on the three hidden faces, and the solver checks that the missing colors really fit there.the image read as a 3x3 cube
8One shade per colorOn each face, all stickers of one color share a shade (no lime next to green). Real cubes stay within 3 degrees of hue; a sticker more than 15 degrees off is a second shade.a 3x3 cube with confidently named colors
9Standard paletteEvery visible color is white, yellow, red, orange, green or blue.a 3x3 cube with confidently named colors
10Standard color pairsNo visible corner or edge combines colors that sit on opposite sides of a standard cube (white-yellow, red-orange, blue-green).a 3x3 cube with confidently named colors
11ScrambledLooks like a random scramble: at least 15 of the 24 visible stickers differ from their face's center, and at most one face shows a symmetric pattern. Fewer than 1% of random scrambles fail this; a cube one move from solved, or a designed pattern, does.a 3x3 cube with confidently named colors
12Distinct centersNo two visible centers share a color.a 3x3 cube with confidently named colors
13Standard center orderCenters fit the standard scheme: standard colors, no opposite colors side by side, correct handedness.a 3x3 cube with confidently named colors
14Possible piecesEvery visible corner and edge is a real cubie under some color scheme. Without a scheme (e.g. two centers share a color) only pieces impossible under every scheme, such as one color twice, are known.a 3x3 cube with confidently named colors
15Unique piecesNo cubie is needed twice. Decided once every piece is possible, or when two pieces show the same colors.a 3x3 cube with confidently named colors
16SolvableA legal, reachable cube state matches everything visible.a 3x3 cube with confidently named colors
17Solvable (standard)A legal state matches using the standard color scheme.a 3x3 cube with confidently named colors