Dataset Viewer
Auto-converted to Parquet Duplicate
Model
stringlengths
8
22
Best fit / published position
stringlengths
16
50
Where tested
stringclasses
5 values
Tested cost or access
stringlengths
13
41
Quality used
stringlengths
7
94
Face test
stringlengths
43
227
Hospital-screen test
stringlengths
58
207
Product-ad test
stringlengths
44
162
Main limitation
stringlengths
165
575
Reroll / exception
stringclasses
5 values
Tested period
stringclasses
5 values
Published rank
int64
1
18
Face place
int64
1
18
Hospital-screen place
int64
1
18
Product-ad place
int64
1
18
Sum of places
int64
8
45
Snapshot version
int64
6
6
GPT Image 2.5 Flare
Faces and Product Adverts
WaveSpeed
$1.00 per run
4K, max quality; 2576 × 3216 PNG output
visible pores, individual hairs and convincing eyes; warmer and more flattering than Sunburst, with smoother under-eye skin.
crisp, accurate text and a natural monitor angle; the dashboard still looks like a clean render, with extra chart-axis numbers.
the most believable campaign photograph, with both hands, late side light, smoked glass and the most reasonable bottle scale among images following the brief.
visible pores, individual hairs and convincing eyes; warmer and more flattering than Sunburst, with smoother under-eye skin. crisp, accurate text and a natural monitor angle; the dashboard still looks like a clean render, with extra chart-axis numbers. the most believable campaign photograph, with both hands, late side...
First output; no editing or upscaling.
October 2026
1
2
5
1
8
6
GPT Image 2.5 Sunburst
Close-Up Facial Detail
WaveSpeed
$1.00 per run
4:5, 4K, max quality; 2576 × 3216 PNG output
the least retouched face, with open pores, fine under-eye creases, irregular freckles, separate eyebrow hairs and faint veins in the eye whites; the colour grade is faded.
crisp and exact angled text, a convincing person and a patient bed behind the station; the dashboard still looks like a clean render.
crisp lettering and plausible glass, but one hand is in a pocket and the bottle is larger than Flare's.
the least retouched face, with open pores, fine under-eye creases, irregular freckles, separate eyebrow hairs and faint veins in the eye whites; the colour grade is faded. crisp and exact angled text, a convincing person and a patient bed behind the station; the dashboard still looks like a clean render. crisp letterin...
First output; no editing or upscaling.
October 2026
2
1
6
3
10
6
Midjourney 8.2
Visual Style and Art Direction
Midjourney website
$10 Basic monthly plan
V8.2, Raw, HD; first image of four
striking editorial beauty shot with a wet sheen and larger lips; it diverged from the neutral studio brief.
the best-looking photograph, with coherent teal lighting, but OXYGEN and NEXT CHECK read “989b” and the requested 14:30 is missing.
polished tailoring and visual direction, with one hand in a pocket and a bottle that is too large.
striking editorial beauty shot with a wet sheen and larger lips; it diverged from the neutral studio brief. the best-looking photograph, with coherent teal lighting, but OXYGEN and NEXT CHECK read “989b” and the requested 14:30 is missing. polished tailoring and visual direction, with one hand in a pocket and a bottle ...
First image of each four-image batch.
August 2026
3
5
3
10
18
6
GPT Image 2
Portraits and ChatGPT Workflows
WaveSpeed
$0.72 per run
4K, High, PNG
strong eyebrows, lashes and irises; cheeks are smoother than either new 2.5 result.
accurate, neat requested text in an ordinary, plain scene.
strong campaign quality with an oversized bottle.
strong eyebrows, lashes and irises; cheeks are smoother than either new 2.5 result. accurate, neat requested text in an ordinary, plain scene. strong campaign quality with an oversized bottle.
First output; no editing or upscaling.
August 2026
4
3
10
5
18
6
MiniMax H3-Image
Large, Detailed Portraits and People-Led Adverts
WaveSpeed
$0.09 per run
4:5, 4K, PNG; no prompt-enhancer switch on the form; 3648 × 4576 PNG output
pores, moles, fine facial hair and a natural asymmetry hold up at full size, with a convincing neck and collarbones; clumped, mascara-like lashes and far more freckles than requested keep it behind GPT Image 2.
exact text and a believable doctor, but a flat dashboard with oversized labels, a heart trace with no heartbeat, a zigzag oxygen chart, two active tabs and an office by a window instead of a nurses' station.
a believably sized bottle, an expensive-looking suit and a man with a real face; the bottle is empty and clear where liquid and smoked glass were requested.
pores, moles, fine facial hair and a natural asymmetry hold up at full size, with a convincing neck and collarbones; clumped, mascara-like lashes and far more freckles than requested keep it behind GPT Image 2. exact text and a believable doctor, but a flat dashboard with oversized labels, a heart trace with no heartbe...
First output; no editing, upscaling or rerolls.
17 November 2026
5
4
11
6
21
6
Ideogram 4.5
Facial Detail and Product Adverts at $0.22 per Run
WaveSpeed
$0.22 per run
3:4 (no 4:5 offered), 2K, High quality, prompt expansion off; 1728 × 2304 PNG output
Real pores, well-drawn ears, good hair texture, no artefacts; mascara added despite ban on dramatic makeup.
All nine required lines exact; doctor looks past the screen; monitor well under 45% of frame; dashboard looks pasted on.
VAREN 92 correct; both hands with visible veins; convincing man, tailoring and concrete; bottle far oversized and clear glass with amber liquid, not smoked glass.
Real pores, well-drawn ears, good hair texture, no artefacts; mascara added despite ban on dramatic makeup. All nine required lines exact; doctor looks past the screen; monitor well under 45% of frame; dashboard looks pasted on. VAREN 92 correct; both hands with visible veins; convincing man, tailoring and concrete; bo...
First output; no editing, upscaling or rerolls.
9 November 2026
6
6
12
4
22
6
Ideogram 4.0
Screen Text and Facial Detail
WaveSpeed
$0.20 per run
2K, High, JPEG
strong eye detail and real skin texture, with heavy mascara despite the ban on dramatic makeup.
2nd on the documented rerun: a convincing patient-camera panel and working dashboard idea, with requested chart elements missing.
strong brutalist architecture, but the most oversized bottle in the test, wider than the man's torso.
strong eye detail and real skin texture, with heavy mascara despite the ban on dramatic makeup. 2nd on the documented rerun: a convincing patient-camera panel and working dashboard idea, with requested chart elements missing. strong brutalist architecture, but the most oversized bottle in the test, wider than the man's...
Hospital wording rerun; ranked output is attempt 2.
August 2026
7
7
2
14
23
6
Seedream 5.5
Low-Cost Volume Work and People Shots
WaveSpeed
$0.033 per run ($0.035 list, 5% discount)
4:5, 2K (highest tier), PNG, background auto, Prompt Enhancer not used; 1831 × 2289 PNG output
real pores, fine hairs, properly drawn ears and good hair without 5.0 Pro's beauty-filter glow; the bun is looser than requested, the frame crops the top of the head and the 2K file holds less fine detail than the 4K portraits.
every required line spelled correctly and the doctor looks at the screen, but the dashboard is too neat, every chart axis carries gibberish numbers and the doctor's second hand is missing.
smoked glass, a matte cap and the right light, but an absurdly large bottle, one hand in a pocket and no liquid showing through the glass.
real pores, fine hairs, properly drawn ears and good hair without 5.0 Pro's beauty-filter glow; the bun is looser than requested, the frame crops the top of the head and the 2K file holds less fine detail than the 4K portraits. every required line spelled correctly and the doctor looks at the screen, but the dashboard ...
First output; no editing, upscaling or rerolls.
17 November 2026
8
9
8
9
26
6
Seedream 5.0 Pro
Hospital Photography and Low Test Cost
WaveSpeed
$0.09 per run
2K, JPEG
pores and uneven redness are visible, but a beauty-filter glow gives the face a CGI edge.
the monitor exposure matches the room and every requested element is present; extra half-readable labels in the tabs and sidebar violate the brief.
complete setting, but a big bottle and readable text on a book spine that the brief did not permit.
pores and uneven redness are visible, but a beauty-filter glow gives the face a CGI edge. the monitor exposure matches the room and every requested element is present; extra half-readable labels in the tabs and sidebar violate the brief. complete setting, but a big bottle and readable text on a book spine that the brie...
First output; no editing or upscaling.
August 2026
9
13
1
13
27
6
Reve 2.1
Editing Workflows
WaveSpeed
$0.28 per run
4:5, JPEG; no resolution control shown
good eyes and glowing skin, but black hair instead of the requested dark blonde and less fine detail than the leaders.
natural monitor perspective, correct core text and a coherent photograph; a few tiny chart labels break up at full size.
a usable, controlled advert with one hand in a pocket and a bottle on the large side.
good eyes and glowing skin, but black hair instead of the requested dark blonde and less fine detail than the leaders. natural monitor perspective, correct core text and a coherent photograph; a few tiny chart labels break up at full size. a usable, controlled advert with one hand in a pocket and a bottle on the large ...
First output; no editing or upscaling.
August 2026
10
14
4
12
30
6
GPT Image 2.6 Sunburst
Lower-Cost OpenAI Runs and Extra Output Formats
WaveSpeed
$0.85 per run
4:5, 4K, max quality, high complexity; 2576 × 3216 PNG output
a crispy, grainy skin finish like Grok's, a long, narrow face with mirrored-looking symmetry and ears that don't sit right; re-placed to 15th in November.
All nine required lines exact; adds oxygen-chart axis labels 100%-92%; hand grips a missing mouse; screen almost front-on.
VAREN 92 correct; both hands; believable man and scene; bottle oversized.
a crispy, grainy skin finish like Grok's, a long, narrow face with mirrored-looking symmetry and ears that don't sit right; re-placed to 15th in November. All nine required lines exact; adds oxygen-chart axis labels 100%-92%; hand grips a missing mouse; screen almost front-on. VAREN 92 correct; both hands; believable m...
First output; no editing, upscaling or rerolls.
8 November 2026
11
15
9
8
32
6
GPT Image 2.6 Flare
Quick OpenAI Images Where the Text Must Be Right
WaveSpeed
$0.85 per run
4:5, 4K, max quality, high complexity; 2576 × 3216 PNG output
a crispy, grainy texture, a narrow, oddly mirrored face and grey skin around the eyes; the moles and freckles don't save it, and it finished 16th, one place behind 2.6 Sunburst.
All nine required lines exact; no extra readable text; hand rests flat on desk; screen almost front-on, reads as a crop.
VAREN 92 correct; both hands; bottle far oversized; smooth glossy generated finish.
a crispy, grainy texture, a narrow, oddly mirrored face and grey skin around the eyes; the moles and freckles don't save it, and it finished 16th, one place behind 2.6 Sunburst. All nine required lines exact; no extra readable text; hand rests flat on desk; screen almost front-on, reads as a crop. VAREN 92 correct; bot...
First output; no editing, upscaling or rerolls.
8 November 2026
12
16
7
11
34
6
MAI-Image 2.6 Preview
Bottle Material, with a Failed Hospital Test
MAI Playground
Free during the limited preview
Highest option available
generic portrait compared with the leaders.
no valid hospital scene returned after roughly ten attempts; the published asset is a failure capture.
the best bottle material and believable foreground scale, with one hand in a pocket and inconsistent lighting on the man.
generic portrait compared with the leaders. no valid hospital scene returned after roughly ten attempts; the published asset is a failure capture. the best bottle material and believable foreground scale, with one hand in a pocket and inconsistent lighting on the man.
No hospital image after roughly ten attempts; failure capture retained.
August 2026
13
17
18
2
37
6
Meta Muse Image
Facial Detail in a Free Test Route
Meta AI
Free during the test
No manual quality control
convincing pores, peach fuzz and upper-lip hair at 1408 × 1760, ahead of all three Nano Bananas.
exact text but a rigid, lifeless doctor, a pasted dashboard with an oversized header and distracting desk props.
good image quality but an oversized bottle and a man placed as though he is standing in the reflecting pool.
convincing pores, peach fuzz and upper-lip hair at 1408 × 1760, ahead of all three Nano Bananas. exact text but a rigid, lifeless doctor, a pasted dashboard with an oversized header and distracting desk props. good image quality but an oversized bottle and a man placed as though he is standing in the reflecting pool.
First output; no editing or upscaling.
August 2026
14
8
17
15
40
6
Nano Banana 2.1
The Highest-Placed Nano Banana in This Test
WaveSpeed
$0.135 per run
4:5, 4K; Web Search and Image Search off; 3712 × 4608 PNG output
cleaner and more convincing than the two older Nano Bananas, though skin remains smooth; the recorded PNG output is 3712 × 4608.
strong monitor perspective and exact text, but the dashboard looks like a separate pasted layer and the doctor is lifeless.
good hands and a convincing setting, but the man's head looks too large for his body.
cleaner and more convincing than the two older Nano Bananas, though skin remains smooth; the recorded PNG output is 3712 × 4608. strong monitor perspective and exact text, but the dashboard looks like a separate pasted layer and the doctor is lifeless. good hands and a convincing setting, but the man's head looks too l...
First output; no editing or upscaling.
October 2026
15
10
14
16
40
6
Grok Imagine Image 2.0
Editing Controls
Grok
Included in the subscription used
Quality mode
decent eyes and lips beneath an obvious applied grain.
a rigid, generic doctor, a nearly flat screen and an added star logo that the brief banned.
its strongest image, with a realistic man, polished setting and a large bottle that is less extreme than the Nano Bananas.
decent eyes and lips beneath an obvious applied grain. a rigid, generic doctor, a nearly flat screen and an added star logo that the brief banned. its strongest image, with a realistic man, polished setting and a large bottle that is less extreme than the Nano Bananas.
First output; no editing or upscaling.
August 2026
16
18
16
7
41
6
Nano Banana Pro
Recurring Characters in Separate Reference Work
WaveSpeed
$0.24 per run
4K, PNG
attractive, detailed face behind Nano Banana 2.1 and the leading models.
the least lifeless doctor in the lower group, with some station depth and a faint reflected hand painted into the interface.
the bottle approaches half the man's height.
attractive, detailed face behind Nano Banana 2.1 and the leading models. the least lifeless doctor in the lower group, with some station depth and a faint reflected hand painted into the interface. the bottle approaches half the man's height.
First output; no editing or upscaling.
August 2026
17
11
13
17
41
6
Nano Banana 2
Fast Editing Workflows
WaveSpeed
$0.14 per run
4K, PNG; Web Search and Image Search off
a good face behind Nano Banana 2.1 and Pro.
invented a second oxygen reading, “OXYGEN 16%”, making the screen incorrect.
the bottle approaches half the man's height.
a good face behind Nano Banana 2.1 and Pro. invented a second oxygen reading, “OXYGEN 16%”, making the screen incorrect. the bottle approaches half the man's height.
First output; no editing or upscaling.
August 2026
18
12
15
18
45
6

Best AI Image Generators in 2026: Tested and Ranked

These Guys Know tested 18 models on three shared briefs: a close-up face, exact medical text inside a photographed hospital monitor, and a VAREN 92 fragrance advert. Ten models were tested in August 2026 and three were added in October. GPT Image 2.6 Flare, GPT Image 2.6 Sunburst and Ideogram 4.5 joined on 8 and 9 November, and MiniMax H3-Image and Seedream 5.5 on 17 November. Earlier outputs were re-ranked against the additions, not retested.

Snapshot v6 · August to November 2026 · 54 model/test combinations · 53 usable outputs · one failed combination.

This dataset follows the TGK article updated on 17 November 2026. The dataset release was synchronised on 17 November 2026.

The original ten models' portrait outputs from August 2026

Historical comparison of the original ten models. The current dataset includes eight later additions.

What the Ranking Means

Each model receives a distinct editorial place from 1 to 18 in each test. The three places add up to the total, so lower totals rank higher. Equal totals are separated by the better placement in each model's strongest test, which is how Meta Muse Image and Nano Banana 2.1 are separated on 40, and Grok Imagine Image 2.0 and Nano Banana Pro on 41. Midjourney 8.2 and GPT Image 2 finished level on 18 with the same three places, so we compared their images again: Midjourney's hospital photograph is far stronger, while GPT Image 2's advert is only slightly sharper, so Midjourney keeps third. These are human editorial judgments, not measurements of image quality or a universal benchmark score.

Rank Model Face Hospital Advert Total
1 GPT Image 2.5 Flare 2 5 1 8
2 GPT Image 2.5 Sunburst 1 6 3 10
3 Midjourney 8.2 5 3 10 18
4 GPT Image 2 3 10 5 18
5 MiniMax H3-Image 4 11 6 21
6 Ideogram 4.5 6 12 4 22
7 Ideogram 4.0 7 2 14 23
8 Seedream 5.5 9 8 9 26
9 Seedream 5.0 Pro 13 1 13 27
10 Reve 2.1 14 4 12 30
11 GPT Image 2.6 Sunburst 15 9 8 32
12 GPT Image 2.6 Flare 16 7 11 34
13 MAI-Image 2.6 Preview 17 18 2 37
14 Meta Muse Image 8 17 15 40
15 Nano Banana 2.1 10 14 16 40
16 Grok Imagine Image 2.0 18 16 7 41
17 Nano Banana Pro 11 13 17 41
18 Nano Banana 2 12 15 18 45

Round Winners and Recorded Costs

GPT Image 2.5 Sunburst wins the close-up face test, Seedream 5.0 Pro wins the hospital-screen test, and GPT Image 2.5 Flare wins the VAREN 92 advert test. Flare ranks first overall on 8, followed by Sunburst on 10. Of the November additions, MiniMax H3-Image placed highest, fifth overall.

The two GPT Image 2.5 variants each cost $1.00 per recorded October run. The GPT Image 2.6 variants cost $0.85 per run on 8 November, and Ideogram 4.5 cost $0.22 per run on 9 November. On 17 November, MiniMax H3-Image showed $0.09 per 4K run (3648 × 4576 output), and Seedream 5.5 was charged $0.033 per 2K run (1831 × 2289 output) after a 5% discount on a $0.035 list price. Seedream 5.0 Pro's $0.09 run belongs to the August test. Midjourney's $10 Basic price is a monthly plan, not a per-image charge. These are recorded test costs, not current price quotes.

Files

  • model_results.csv: 18 model records with test dates or periods, settings, access routes, recorded costs, findings, places, totals and exceptions.
  • evidence_records.csv: 54 model/test combinations plus Ideogram 4.0's earlier hospital attempt, for 55 evidence rows. URLs identify published outputs or the recorded failure capture.
  • test_prompts.csv: the three published prompts and their settings notes.
  • snapshot_v6.json: the complete maintained research snapshot, including model/test associations, recorded generation times where we have them and their source dates. Missing timings are null.
  • CHANGELOG.md: what changed in each snapshot.
  • LICENSE: CC BY 4.0 for TGK-authored metadata, prompts, annotations, methodology and tables. Model-generated images and third-party material are excluded.

Method and Exceptions

We used the highest practical native quality, 4:5 where offered and the closest portrait ratio otherwise. We did not edit or upscale results to improve them.

  • Ideogram 4.0 hospital: the ranked image is the second attempt, after the word "vertical" made the first monitor upright. The first attempt remains in the evidence records; its exact earlier prompt was not retained.
  • MAI-Image 2.6 Preview hospital: no valid hospital scene arrived in roughly ten attempts. Its evidence URL is a failure capture, not a successful generated scene.
  • Midjourney 8.2: the first image in each four-image batch was judged.
  • November additions: the first output was kept, with no rerolls. Both GPT Image 2.6 variants used 4:5, 4K, max quality and high complexity through WaveSpeed. Ideogram 4.5 used 3:4 because 4:5 was unavailable, 2K, High quality and prompt expansion off. MiniMax H3-Image used 4:5, 4K and PNG; its form has no prompt-rewrite switch and the enhancer button was not used. Seedream 5.5 used 4:5, 2K (its highest tier), PNG and background auto, with the Prompt Enhancer not used.
  • November re-placement: when the 17 November additions went in, we compared every portrait again at full size and moved GPT Image 2.6 Sunburst and Flare down to 15th and 16th in the face round. In the advert round, Grok Imagine Image 2.0 now sits seventh, one place above GPT Image 2.6 Sunburst, whose bottle is about as big as the man's torso. Every other change in places comes from the two new models taking positions.
  • Mixed dates and formats: every model retains its test period. Published WebP or JPEG images may be converted or resized from the recorded PNG output.

This comparison covers three first-output jobs. It does not establish performance across every editing workflow, maximum quality after rerolls, licensing or local deployment. Provider wrappers can change defaults, prompt rewriting, compression and controls. Any separate recommendation for an editing or character-reference workflow comes from the article's discussion, not a score invented from these three tests.

Version History and Reuse

Snapshot v6 adds MiniMax H3-Image and Seedream 5.5, recounts every place from 1 to 18, re-places the GPT Image 2.6 pair in the face round and GPT Image 2.6 Sunburst in the advert round, and refreshes the two GPT Image 2.6 face findings. The August ten-model dataset, the October thirteen-model Snapshot v4 and the sixteen-model Snapshot v5 remain in the repository's commit history and as earlier Zenodo versions. GPT Image 2 evidence remains labelled GPT Image 2.

Attribute reused TGK-authored data to These Guys Know, link the canonical article, and retain the model version, test date or month, access route and relevant limitation. The dataset licence does not grant rights to individual generated outputs, provider interfaces, logos or trademarks; applicable provider terms may govern those materials.

Dataset mirrors and citation

Hugging Face: https://huggingface.co/datasets/These-Guys-Know/tgk-ai-image-generators Kaggle: https://www.kaggle.com/datasets/mikehazard/tgk-ai-image-generators Zenodo (this version): https://doi.org/10.5281/zenodo.24322759 Zenodo (v5): https://doi.org/10.5281/zenodo.24316617

Downloads last month
138

Collection including These-Guys-Know/tgk-ai-image-generators