Test log
Decohere AI test data
Raw results for the Decohere AI 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 | Character & Subject Consistency | – | n/a |
| T02 | Editing, Inpainting & Image-to-Image | – | n/a |
| T03 | Prompt Adherence | Elements correct: 123 of 179 | 3/5 |
| T04 | Photorealism & Anatomy | Images that pass: 15 of 24 | 3/5 |
| T05 | Text in Images | Words correct: 5 of 61 | 1/5 |
| T06 | Style Range | Styles hit: 8 of 10 | 4/5 |
| T07 | Resolution & Upscaling | Native size: 768 px (0.59 MP) | 2/5 |
| T08 | Speed per Image | Median per image: 1.6 s | 5/5 |
| T09 | Usage Rights, Watermarks & Privacy | Points met: 3 of 7 checked | 2/5 |
| T10 | Pricing & Value | Free images I got in one day: 197 | 5/5 |
| Overall (in the review) | 3.7/5 | ||
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 completed.
Raw notes
A described woman and a described thermos in six scenes each.
Not run: the daily limit ("Daily free image limit reached") ended the session after the style prompts, so none of the twelve images was made. The AI characters feature that holds a person steady is a paid plan feature.
- Run on a paid plan, or on the next day of the free 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 realtime page offers an Upload button and five buttons under each image (Save, Share, Station, Upscale and Video), but the limit was reached before any edit could be made, and no masked fill or frame expansion was offered.
- Run on a paid plan, or on the next day of the free plan.
Prompt Adherence
3/5Tested- Method
- 34 of the 40 fixed prompts (6 portraits, 6 product shots, 6 set-wording images, 6 busy scenes, 10 style prompts), typed in by script on the Realtime AI page, first image kept, every written element counted at full size.
- Elements correct
- 123 of 179
- Busy scenes
- 14 of 42
- Remark
- Scale 3: 65 to 79 percent of elements correct. The sample was 34 prompts and 179 elements, not 40 prompts and about 240; the six character scenes were not run.
Raw notes
34 of the 40 fixed prompts (6 portraits, 6 product shots, 6 set-wording images, 6 busy scenes, 10 style prompts), typed in by script on the Realtime AI page, first image kept, every written element counted at full size.
123 of 179 elements were right (68.7 percent). Portraits 31 of 35, products 27 of 32, style prompts 38 of 40, set-wording images 13 of 30, busy scenes 14 of 42.
Kitchen scene 3 of 8: two apples, a red jug instead of a blue one, and the dog and the black animal on the counter, not under the table and on the chair. Market scene 2 of 7: more than two people and none passing a melon. Beach scene 0 of 7: four large umbrellas out of order, no child at a sandcastle and no boat. Street scene 2 of 8: two pigeons, and no man on a bench.
Added content that changed the scene: a second balloon on the girl in the field, and invented lines of text on the poster, the sign and the cover.
- 34 of 40 prompts. The six character scenes of the consistency test were not run because the daily limit was reached.
Photorealism & Anatomy
3/5Tested- Method
- 24 images: the first and the second image Decohere drew for each of the six portrait and six product prompts, inspected at 100 percent for faults that give a generated image away.
- Images that pass
- 15 of 24
- Hand faults
- 4
- Remark
- Scale 3: 13 to 17 of 24 pass.
Raw notes
24 images: the first and the second image Decohere drew for each of the six portrait and six product prompts, inspected at 100 percent for faults that give a generated image away.
15 of 24 passed. Passed: the man on the bench, the laughing woman, the girl in the field, the grey-haired woman, the pancakes, the pizza and the perfume bottle, with no fault at full size.
Failed (9): both chef images (the hand and knife merge into the tomatoes, and the jacket badge is made-up lettering), both friends-with-headphones images (hands merge into the cable and gloves), both watch dials (malformed numerals), both olive oil labels (letters that read nothing) and one image of three trainers instead of two.
Hand faults 4. Face faults 0. Faults in text-like marks on objects 5.
- The second image for each prompt is the second of the five that Decohere drew in the same run, not a separate run.
Text in Images
1/5Tested- Method
- The six set-wording prompts: the coffee shop sign and board (11 words), the honey jar label (8), the night market poster (13), the book cover (8), the fridge note (10) and the street signs (11), 61 words in all.
- Words correct
- 5 of 61
- Images fully correct
- 0 of 6
- Remark
- Scale 1: under 40 percent of words correct, or no readable text.
Raw notes
The six set-wording prompts: the coffee shop sign and board (11 words), the honey jar label (8), the night market poster (13), the book cover (8), the fridge note (10) and the street signs (11), 61 words in all.
5 of 61 words were spelled right (8 percent): "Bean", "Wild", "Summer", "The" and "the". None of the six images was fully correct.
The shop sign reads "HIOORUR BEAN", the honey label "WILD BIOSSOM", the poster "SUMMER NLIGHT MAKET" and the book cover "THE OIR OUTA CHIARD". The fridge note starts "Pelense" and the street scene has no readable sign, and no bus stop sign.
Every image added extra text that was not asked for, as lines of letters that read nothing.
- The sample is all six set-wording prompts.
Style Range
4/5Tested- Method
- Ten style prompts for one lighthouse on a rocky coast at dusk, first image kept, each judged on whether the style shows in the medium and the lighthouse, rocks and dusk light are all there.
- Styles hit
- 8 of 10
- Remark
- Scale 4: 8 or 9 styles hit, subject kept.
Raw notes
Ten style prompts for one lighthouse on a rocky coast at dusk, first image kept, each judged on whether the style shows in the medium and the lighthouse, rocks and dusk light are all there.
8 of 10 styles hit: oil painting, watercolour, pencil sketch, flat vector, Japanese woodblock print, anime cel, 1970s film photograph and black ink line art. The lighthouse, rocky coast and dusk light were in all ten.
Missed (2): the pixel art prompt gave a smooth illustration with no pixel grid, and the clay 3D render prompt gave a smooth computer render with no clay.
- The anime cel image is a borderline hit: flat colour blocks, but soft shading on the rocks. Y3 and Y4 were run again after two of my runs overlapped.
Resolution & Upscaling
2/5Tested- Method
- Every saved image was measured, the aspect ratios were read in the app, and the upscaler was looked for on the free plan.
- Native size
- 768 px (0.59 MP)
- Upscaler on free
- None
- Remark
- Scale 2: under 1 megapixel native. No upscale was possible on the free plan, so the scale's upper marks were not reached.
Raw notes
Every saved image was measured, the aspect ratios were read in the app, and the upscaler was looked for on the free plan.
All 171 images I kept came out at 768 by 768 pixels (0.59 megapixels), drawn in the page as WebP data. The app offers three frames: wide 16:9, square 1:1 and tall 9:16.
The upscaler (4x for 2 credits) is listed on the paid plans only, so no upscale was run. A download was not tested after the daily limit was reached.
- Native size was measured on the square frame only; the wide and tall sizes were not read. No upscale was run.
Speed per Image
5/5Tested- Method
- Every generation in the prompt adherence test was timed from filling the prompt box to the first full-size image on the page.
- Median per image
- 1.6 s
- Slowest image
- 2.1 s
- Remark
- Scale 5: a median under 10 seconds.
Raw notes
Every generation in the prompt adherence test was timed from filling the prompt box to the first full-size image on the page.
Median 1.6 seconds, fastest 0.8, slowest 2.1 across 34 prompt runs. My timer checks about every 0.8 seconds, so each figure is within that.
Each run drew five images; the first one appeared in under two and a half seconds every time. No generation failed, and there was no queue until the daily limit.
- 34 runs, not 40. The timer reads in steps of about 0.8 seconds.
Usage Rights, Watermarks & Privacy
2/5Tested- Method
- The terms of service (read 6 November 2026), the plans page and its questions were checked against the ten-point list.
- Points met
- 3 of 7 checked
- Remark
- Scale 2: 4 or 5 of 10. 3 of 7 checked is 43 percent, so about 4 of 10.
Raw notes
The terms of service (read 6 November 2026), the plans page and its questions were checked against the ten-point list. Seven points could be checked without a download.
Met (3): the terms say who owns the output ("you own all Assets you create"); the Creator plan has a commercial license, which its pricing question says gives you all the rights; the terms say content goes in public galleries only when you opt in.
Not met (4): free output is licensed under Creative Commons Noncommercial, so no commercial use on free (the Explorer plan has no rights either); Decohere may use your content without asking to improve the service, with no opt-out offered; the only training source stated is "public datasets"; none of the 171 saved files holds a content credentials label.
Not tested (3): a watermark on paid downloads, a watermark on free downloads (the plans page says free images carry one), and whether real people and brands are refused.
- Terms read by me, not a lawyer. No download was made and no paid plan was bought.
Pricing & Value
5/5Tested- Method
- The free tier was used until the app stopped me, and the plans page was read in both billing views.
- Free images I got in one day
- 197
- Cost per image on Creator
- about $0.019
- Remark
- Scale 5: under $0.02 an image with at least 10 free images a day. The paid per-image price comes from the plans page, because the quality generator was not open on the free plan.
Raw notes
The free tier was used until the app stopped me, and the plans page was read in both billing views.
Free: the app announced 200 realtime images for the day, and I got 197 from 38 prompt runs before it said "Daily free image limit reached" and "Claim more free images tomorrow".
The cheapest plan with a commercial license is Creator at $29.99 a month ($19.99 a month paid yearly) for 400 credits a month. A quality image costs 1 credit for 4 images, so about $0.019 an image on the monthly price and about $0.012 paid yearly.
Costs on top: 2 credits for a 4x upscale, 1 credit for a 4-second video, 50 credits to train an AI character. Unused credits expire at the end of the billing period; refill credits do not.
- No plan bought; the credit costs are those the pricing questions state.
Coverage
Tested
- 34 of the 40 fixed prompts, first image kept: 6 portraits, 6 product shots, 6 set-wording images, 6 busy scenes, 10 styles
- Time from filling the prompt box to the first full-size image, for every prompt
- The size of every saved image, and the aspect ratios the app offers
- Terms of service, privacy page, plans page and the pricing questions read in full
- The quality image generator, which showed the message "Upgrade to generate more images"
Not tested or limited
- The free plan's daily limit ended the session after 197 images: the six character scenes and six product scenes of the consistency test were not run
- Editing was not run: the limit was reached before the edit prompts, and inpainting, image restyles and frame expansion were not offered on the realtime page
- Photorealism used 24 images (the first and second image of each of 12 prompts), not two separate runs of 12 prompts
- No download was tested after the limit, so the watermark the plans page says free images carry, the file format of a download and any content credentials label were not checked
- No plan was bought: the quality generator, the upscaler and paid ownership were read from the plans page and the terms only
- P1 was typed once while I learned the page, and Y3, Y4 and C1 were run again after two of my own runs overlapped; the images graded are from the clean runs
Still to test
- Consistency and editing, on a paid plan or the next free day
- The quality generator and the 4x upscaler
- A download, its watermark and its file
- Ownership and cancellation after buying Creator