Auditable scoring · not vibes

The Sunflowers Fidelity Test

One text description. No reference image supplied to either model. The question: how close can a text-to-image model get to a specific, known composition — and can we score it in a way readers can check themselves?
Van Gogh's Sunflowers (1888, National Gallery NG3863) is public domain — the artist died in 1890. These generated images are original photographic interpretations of the composition, not reproductions of the painting.
▸ View the exact prompt sent to both models
A photorealistic photograph of a still life, vertical portrait orientation, 3:4 aspect ratio (noticeably taller than wide). Shot on a medium-format camera with soft, even, diffuse daylight.

Subject: exactly fifteen real sunflowers arranged in a simple round earthenware vase.

Composition, follow precisely:
- The vase sits slightly below center, resting on a plain table surface.
- The table surface meets the back wall about one third up from the bottom of the frame.
- The background wall is a flat, PALE yellow — soft and light, almost cream-yellow, NOT a deep orange or saturated ochre. No pattern, no window, no other objects.
- The vase is glazed earthenware, divided by a single horizontal line across its middle: the upper half is pale cream, the lower half is a warmer tan-brown.
- The bouquet fills the upper two thirds of the frame in a broad, slightly ragged fan shape, wider than the vase, reaching nearly to the top edge.

The fifteen flowers must show a full range of life stages, spread throughout the arrangement rather than grouped:
- Four in full bloom, facing the camera, bright golden-yellow petals fully open.
- Three in full bloom seen from the side or three-quarter angle.
- Four past their peak, petals drooping, curling and browning at the tips.
- Four spent seed heads with almost no petals left, showing large dark brown textured centers and green spiky bracts.

Style: real botanical photography. Sharp focus across the flower heads, natural petal texture, fine hairs visible on the thick green stems, subtle imperfections and blemishes on the older petals. Warm but light color palette of yellows, creams, ochres, greens and browns.

Absolutely no text, no signature, no picture frame, no canvas weave, no visible brushstrokes, no painterly or impasto effect. This must read as a photograph of real flowers, not as a painting or a painting reproduction.
The Original
Van Gogh, 1888 · National Gallery NG3863 · public domain
Drop the public-domain file here to compare side by side.

Source: Wikimedia Commons — 4719 × 6000, public domain.
Ground truth: 15 sunflowers, pale-yellow ground, two-tone earthenware vase, every flower at a different life stage. 92.1 × 73 cm (3:4 portrait).
Gemini 3 Pro Image
generated from text only
Native output: 896 × 1200 (ratio 0.747)
GPT-5 Image
generated from text only
Native output: 1024 × 1024 (ratio 1.000)

The scoring rubric

Weighted and tiered deliberately. Tier A is machine-verifiable — already scored, no judgment involved. Tier B is countable — you count, and anyone can re-count to check you. Tier C is subjective and capped at 20% of the total on purpose, so taste can't swing the result. Set the dropdowns; totals update live.
CriterionWeightGemini 3 ProGPT-5 Image
Tier A — Machine-verifiable (30 pts) · pre-scored from the files
Aspect ratio 3:4 portrait
Read directly from image dimensions. Tolerance ±0.03.
15 PASS · 0.747 FAIL · 1.000 (square)
No text, signature, frame or canvas texture
Visual check for prohibited elements the prompt explicitly banned.
15 PASS PASS
Tier B — Countable (50 pts) · you count, readers can verify
Total flower count = 15
Count every distinct flower head. Full = exactly 15. Partial = within 2. Fail = off by 3+.
15
Life-stage mix present (bloom / fading / spent seed head)
All three stages clearly represented = full. Two stages = partial. One = fail.
10
Two-tone vase with horizontal division
Pale upper half, darker lower half, single clean dividing line.
10
Background is flat PALE yellow (not deep ochre)
Eyedrop the wall. Light cream-yellow = full. Mid ochre = partial. Orange/brown = fail.
8
Horizon line ≈ one third up from bottom
Where table meets wall. Within 10% of the one-third mark = full.
7
Tier C — Subjective (20 pts) · capped so taste can't decide the winner
Photorealism (reads as a photo, not a painting)
Petal texture, stem hairs, believable depth of field, no impasto or brushstroke artifacts.
10
Compositional echo of the original
Does the silhouette and fan-shaped spread evoke the painting's arrangement?
10
100
Maximum possible score
Gemini 3 Pro Image
GPT-5 Image
Two honest notes before you publish any number.

1. I corrected the target mid-test. My first spec asked for 12 flowers against a deep ochre wall. The National Gallery describes the actual painting as 15 sunflowers against a pale yellow ground. The prompt above is the corrected version — worth mentioning in the edition, because "we were scoring against the wrong answer at first" is a genuinely useful lesson about benchmark design.

2. I have not filled in the Tier B counts for you. Counting flower heads in a dense bouquet with overlap and occlusion is exactly the kind of judgment I shouldn't quietly make on your behalf and present as fact — two careful people can legitimately disagree by one or two. Tier A is pre-scored because it comes straight off the file dimensions. Tier B is yours to count, and the whole point of the rubric is that any reader can re-count and check you.
Images generated live from text only, no reference image supplied. Original painting: Vincent van Gogh, Sunflowers, 1888, National Gallery London (NG3863), public domain.
National Gallery catalogue entry · AI Super Simplified