Test log
Artbreeder test data
Raw results for the Artbreeder review. The verdict is in the review.
- Tested
- 6 November 2026
- Tester
- Miles Foster
- Account
- Free demo account, signed up by RankDude, nothing paid to the brand
- Tests
- 10
- Screenshots
- 8
- Method
- Published methodology
- Environment
- Google Chrome on desktop
Results
| ID | Test | Key measurement | Score |
|---|---|---|---|
| T01 | Style Range | – | n/a |
| T02 | Character & Subject Consistency | – | n/a |
| T03 | Editing, Inpainting & Image-to-Image | – | n/a |
| T04 | Prompt Adherence | Elements correct: 51 of 59 | 4/5 |
| T05 | Photorealism & Anatomy | Images that pass: 4 of 5 | 4/5 |
| T06 | Text in Images | Words right: 14 of 19 | 3/5 |
| T07 | Resolution & Upscaling | Native size: 1,024 px square | 3/5 |
| T08 | Speed per Image | Median per image: 8.0 s | 5/5 |
| T09 | Usage Rights, Watermarks & Privacy | Points met: 4 of 8 checked | 2/5 |
| T10 | Pricing & Value | Free Flux images a month: 12 | 4/5 |
| Overall (in the review) | 4/5 | ||
Style Range
n/aNot Tested- Method
- Ten style prompts for one lighthouse.
- Remark
- Not graded: no style prompt was run.
Raw notes
Ten style prompts for one lighthouse.
Not run: ten images on Flux cost 2.5 credits, and the free plan had 0.5 left. Running them on SDXL Lightning would have changed the model halfway through the test.
- Run on a plan with more credits.
Character & Subject Consistency
n/aNot Tested- Method
- A described woman and a described thermos in six scenes each.
- Remark
- Not graded: no consistency prompt was run.
Raw notes
A described woman and a described thermos in six scenes each.
Not run: twelve images on Flux cost 3 credits, more than the whole monthly balance.
- Run on a paid plan.
Editing, Inpainting & Image-to-Image
n/aNot Tested- Method
- Twelve fixed edits.
- Remark
- Not graded: no edit was run.
Raw notes
Twelve fixed edits.
Not run: the free credits went on the prompt tests and 0.5 were left, so no edit was tried in the Outpainter or Splicer.
- Run on a paid plan, with the Outpainter and Splicer.
Prompt Adherence
4/5Tested- Method
- 10 of the 40 fixed prompts (3 portraits, 2 product shots, 2 set-wording images, 3 busy scenes), typed in by script on the Flux model, first image kept, every written element counted at full size.
- Elements correct
- 51 of 59
- Busy scenes
- 16 of 21
- Remark
- Scale 4: 80 to 89 percent of elements correct. The sample was 10 prompts and 59 elements, not 40 prompts and about 240.
Raw notes
10 of the 40 fixed prompts (3 portraits, 2 product shots, 2 set-wording images, 3 busy scenes), typed in by script on the Flux model, first image kept, every written element counted at full size.
51 of 59 elements were right (86.4 percent). Portraits and products: 28 of 29. The two set-wording images: 7 of 9. The three busy scenes: 16 of 21.
Kitchen scene 7 of 8: the cat sits on the table, not the chair. Market scene 5 of 7: a pile of oranges, not five in a row. Living room 4 of 6: three green cushions and one white, not two and two. The man on the bench wears square glasses, not round.
Added content that changed the scene: none beyond the invented lettering counted in the text test.
- 10 of 40 prompts. The groups not run (styles and character scenes) and three of the six busy scenes are the ones the free credits could not pay for.
Photorealism & Anatomy
4/5Tested- Method
- Five portrait and product images (P1, P2, P3, F1, F2), one run each, inspected at 100 percent for faults that give a generated image away.
- Images that pass
- 4 of 5
- Hand and face faults
- 0
- Remark
- Scale 4: 18 to 21 of 24 pass. The sample is 5 images, so 4 of 5 (80 percent) stands for about 19 of 24.
Raw notes
Five portrait and product images (P1, P2, P3, F1, F2), one run each, inspected at 100 percent for faults that give a generated image away.
Passed (4): the man on the bench, the laughing woman, the girl with the balloon and the olive oil bottle, with no hand or face faults.
Failed (1): the trainers carry a printed mark on the tongue that looks like letters but spells nothing.
Hand faults 0. Face faults 0. Faults in text-like marks on objects 1.
- Five images in one run, not 24 images in two runs.
Text in Images
3/5Tested- Method
- Two of the six set-wording prompts: the coffee shop sign and board (11 words) and the honey jar label (8 words), every word checked at full size.
- Words right
- 14 of 19
- Images fully correct
- 0 of 2
- Remark
- Scale 3: 70 to 84 percent of words correct. Large headlines came out right; small lines did not.
Raw notes
Two of the six set-wording prompts: the coffee shop sign and board (11 words) and the honey jar label (8 words), every word checked at full size.
14 of 19 words were spelled right, in the place asked for (74 percent). The shop sign "Harbour Bean Coffee Roasters" was perfect, and so was "Wild Blossom Honey" on the label.
The board beside the door reads "OPEN Daily 7 am 6 rp to 5pk" instead of "Open Daily 7 am to 5 pm". The honey label reads "Raw and Unfiltem" and "250 g" with a broken character, not "Raw and Unfiltered 250 grams".
The board's "to" is spelled right and sits between the opening and closing times, so it counts; "5pk", "Unfiltem", the raised small zero in "250" and the broken g do not. Neither image was fully correct. Both carry extra lettering that was not asked for, such as the tiny lettering on the label.
- Two of six prompts; the poster, book cover, note and street sign were not run.
Resolution & Upscaling
3/5Tested- Method
- Every saved image was measured, the five frame shapes were read in the app, and the Download button was used once and compared with the stored file.
- Native size
- 1,024 px square
- Upscale price
- 0.15 credits
- Remark
- Scale 3: 1 megapixel or more native, with an upscaler I could not check. A 4 needs a clean upscale, and none was run.
Raw notes
Every saved image was measured, the five frame shapes were read in the app, and the Download button was used once and compared with the stored file. The upscale was priced but not run.
All 10 images came out at 1,024 by 1,024 pixels (1.05 megapixels) as JPEG. The other shapes are 1,280 by 768, 1,152 by 832, 832 by 1,152 and 768 by 1,280.
The Download button gave the same file, byte for byte, with no watermark. The Upscale button shows a price of 0.15 credits an image.
I did not run the upscale: it would have taken the account to 88 percent of its credits, past the 85 percent line. The pricing page shows a 2,048 pixel download as a paid feature.
- No upscale was run, so no image was compared before and after.
Speed per Image
5/5Tested- Method
- Each of the 10 generations was timed from pressing generate to the full-size image on screen.
- Median per image
- 8.0 s
- Slowest image
- 8.3 s
- Remark
- Scale 5: a median under 10 seconds.
Raw notes
Each of the 10 generations was timed from pressing generate to the full-size image on screen.
Median 8.0 seconds, average 7.8, fastest 7.2, slowest 8.3.
Every image took between 7 and 9 seconds, whatever the prompt. Times are read to the nearest tenth with a quarter-second check.
No generation failed. There was no queue on the free plan at the time.
- 10 generations, not 40.
Usage Rights, Watermarks & Privacy
2/5Tested- Method
- The terms of use (read 6 November 2026, last updated 20 November 2019) and the downloaded file were checked against the ten-point list.
- Points met
- 4 of 8 checked
- Remark
- Scale 2: 4 or 5 of 10. 4 of 8 checked is 50 percent, so about 5 of 10.
Raw notes
The terms of use (read 6 November 2026, last updated 20 November 2019) and the downloaded file were checked against the ten-point list. Eight points could be checked on the free plan.
Met (4): the terms say each image is owned by the user who created it; the same terms release every image under the CC0 licence, so anyone may use it for any purpose, which covers business use on free and paid plans; and the free download had no visible watermark.
Not met (4): all images are public domain by the terms and private mode is a paid feature; the terms give no way to opt out of training; they name no source for training data; the downloaded file carries no content credentials label.
Not tested (2): a watermark on paid downloads, and whether real people are refused when asked for.
- Terms read by me, not a lawyer. No paid plan was bought.
Pricing & Value
4/5Tested- Method
- The free balance was used until it reached the 85 percent line, and the plans page was read in both billing views.
- Free Flux images a month
- 12
- Cost per Flux image on Starter
- about $0.022
- Remark
- Scale 4: $0.02 to $0.049 an image on the cheapest plan. The free plan allows fewer than 10 Flux images a day, so it cannot earn a 5.
Raw notes
The free balance was used until it reached the 85 percent line, and the plans page was read in both billing views.
Free: 3 credits a month. Flux costs 0.25 credits an image, so one month gives 12 Flux images; I made 10 and stopped at 0.5 credits. Cheaper models cost far less: SDXL Lightning is 0.017 credits an image and Flux Schnell 0.03.
The cheapest plan is Starter at $8.99 a month ($7.49 a month paid yearly). Credits cost about $0.09 each on the monthly plan, so a Flux image costs about $0.022, and about $0.019 paid yearly.
Costs on top: 0.15 credits for an upscale. Private mode, Animator and the paid file size need a paid plan. Commercial use is allowed on every plan, because the terms release all images as CC0.
- No plan bought; the credit cost of an image on a paid plan was assumed to match the free plan.
Coverage
Tested
- 10 of the 40 fixed prompts, first image kept: 3 portraits, 2 product shots, 2 set-wording images, 3 busy scenes
- Time and credit cost of every image
- The frame shapes and their sizes, read in the app
- The product's own Download button, compared byte for byte with the stored image
- The terms of use (PDF linked from the site's footer) and the plans page in both billing views
Not tested or limited
- The free plan's 3 monthly credits paid for 10 Flux images: 10 of the 40 prompts were run
- Style range, consistency and editing were not graded: no credits were left under the 85 percent line, and changing to a cheaper model halfway would have changed the test
- Flux Krea, at 0.6 credits an image, was not tried, so I cannot say whether it would have done better
- Photorealism used one run of 5 images, not two runs of 12 prompts
- The upscale (0.15 credits) was not run: it would have taken the balance past the 85 percent line, so resolution is graded on native size only
- No plan was bought: private mode, the paid file and the cancel path were not tested; prompts naming real people were not run
Still to test
- Style range, consistency and editing on a paid plan
- Photorealism with 24 images, and the other 30 prompts
- The upscale at its largest size
- Private mode, paid downloads and cancellation after buying Starter