Vivenso Labs

Research — AI image models

Four image models, put to two tests.

First, make a picture from a written brief. Second, take a picture that already exists and change it: recolour something, remove something, add something, relight the scene. Both tests were run on an architectural picture, an action picture and a picture from a fictional setting. Every result is shown below, side by side, as each model returned it.

Run on 2 October 2026 One run per model Taste not judged

The short version

Which model did which job best.

JobBest in this runWhat we saw
Make a picture from a detailed briefGPT Image 2Met every listed item in all three kinds of picture
Recolour one thingSeedream 5 Pro, then Nano Banana ProAll four made the change; Seedream altered the least of the rest
Remove somethingSeedream 5 Pro and GPT Image 2Both removed all three cleanly. One model left the object in place; one removed too much
Add somethingSeedream 5 Pro and Nano Banana ProAll four added it; these two placed it best and disturbed the least
Relight the whole sceneGPT Image 2 and Nano Banana ProBoth changed the light and kept the subject in all three
Shadows that agree with the lightNo clear winnerNo model passed all four shadow checks; Seedream 5 Pro and GPT Image 2 passed the most
SpeedNano Banana ProAbout 25 seconds a picture; the others took 60 to 110
PriceKling Image O3About 3 US cents a picture, and the loosest with every instruction

Across the nine changes that can be measured, Seedream 5 Pro altered the least of the picture in eight. Which pictures look best is not judged here. That is what the pictures are for.

How far to trust it

One brief, one run.

Each model made each picture once. That shows how a model behaved on one occasion, not how it behaves on average, and a second run could come out differently. The tables record whether the thing asked for appears in the picture, which is something a still image can show.

The models

Four models from four makers.

ModelMakerPrice per pictureTypical time
GPT Image 2OpenAINot recorded89 seconds
Nano Banana ProGoogle$0.13425 seconds
Seedream 5 ProByteDance$0.067572 seconds
Kling Image O3Kuaishou$0.02861 seconds

Every model received exactly the same words, with nothing tuned for any one of them, and was asked for the same shape of picture at about the same size. None of them declined any brief or any change.

Test 1

Making a picture from a brief.

Each brief lists things that must appear. The tables under each set of pictures record which of them did.

Test 1 · Architectural

A house on a cliff, with straight lines to get right.

The brief: a two-storey house of board-formed concrete and vertical cedar slats, cantilevered over a coastal cliff at dusk, with straight verticals, a staircase seen through the glass, a flat roof and a narrow reflecting pool.

Four pictures of a concrete and cedar house on a coastal cliff at dusk, one from each model, each labelled with the model's name.
The same brief, as each model answered it.
Asked forGPT Image 2Nano Banana ProSeedream 5 ProKling Image O3
Two storeysYesNo, oneYesYes
Concrete and cedar slatsYesYesYesYes
Cantilevered over the cliffYesYesYes, most clearlyYes
Staircase seen through the glassYesYesYesYes
Reflecting poolYesYesYesUnclear

Test 1 · Action

A runner in mid-stride, seen from the side.

The brief: a woman running on a coastal trail at sunrise with both feet off the ground, seen from the side with the camera panning, so that the background is streaked and she is sharp; a watch on her left wrist; no text or brand marks.

Four pictures of a woman running on a coastal trail at sunrise, one from each model, each labelled with the model's name.
The same brief, as each model answered it.
Asked forGPT Image 2Nano Banana ProSeedream 5 ProKling Image O3
Seen from the sideYesNo, from the frontYesYes
Both feet off the groundYesUnclearYesYes
Background streaked by the panYes, stronglyNoYesNo
Watch on the left wristYesYesYesYes
No marks or logosYesA small mark on the shortsYesYes

Nano Banana Pro's picture reads most like an ordinary photograph, and it is the one that departs furthest from the brief.

Test 1 · A fictional setting

Does a model know a world it was not built for?

The brief named two kinds of warrior from the Warhammer 40,000 setting and said what each carries. It tests whether a model knows a fictional world in detail, and whether it will draw it at all. None of the four refused.

Four pictures of a gold-armoured guardian with a spear facing a red-armoured warrior with an axe in a ruined hall, one from each model, each labelled with the model's name.
The same brief, as each model answered it.
Asked forGPT Image 2Nano Banana ProSeedream 5 ProKling Image O3
Gold guardian on the left, cloaked, bare-headedYesYesYesYes
Spear held in both handsYesYes, with the gun built into itYesOne hand
Red warrior in a horned helmetYesYesYesNo, a bare horned head
Axe with a toothed, motorised bladeAn axe, raisedYesNo, reads as a hammerA plain axe
Ruined hall, light through a breachYesYesYesHall, under open sky

On this set

The pictures of this setting, here and in the second test, are unofficial test pictures made by AI models. Warhammer 40,000 is a trademark of Games Workshop Limited. Vivenso Labs is not affiliated with Games Workshop, and nothing here is endorsed by it.

Test 2

Modifying a picture that already exists.

Every model was handed the same three source pictures and asked for the same four kinds of change to each: recolour one thing, remove one thing, add one thing, and relight the whole scene. A good result makes the change and leaves everything else alone.

For the first three kinds, the figure beside each model's name is the share of the picture that differs from the source. A small figure is good only if the change was made. Relighting changes the whole picture on purpose, so no figure is given for it.

Test 2 · Recolour

Change the colour or material of one thing.

The house: replace the cedar slats with white render. The runner: make the navy top a strong bright blue. The guardian: make the crimson cloak white.

A source picture of a cliff house with cedar slats, followed by four versions with white render, each labelled with the model's name.
The house: cedar slats to white render.
A source picture of a runner in a navy top, followed by four versions with a blue top, each labelled with the model's name.
The runner: navy top to bright blue.
A source picture of a gold-armoured warrior in a crimson cloak, followed by four versions with a white cloak, each labelled with the model's name.
The guardian: crimson cloak to white.
PictureGPT Image 2Nano Banana ProSeedream 5 ProKling Image O3
The houseYes · 13.5%. Upper windows rearrangedYes · 13.4%Yes · 9.5%Yes · 15.1%. Upper floor rebuilt
The runnerYes, the strongest blue · 7.0%Yes · 2.9%Yes, lighter · 2.8%Turquoise · 6.1%
The guardianYes · 6.5%. Red strips at the waist went white tooYes · 7.2%Yes · 6.2%Yes · 11.2%

All four made every change. Seedream 5 Pro altered the least of the rest in all three pictures.

Test 2 · Remove

Take one thing out and fill what was behind it.

The house: remove the staircase, and its reflection in the pool. The runner: remove the watch from her wrist. The guardian: remove the cloak completely.

A source picture of a cliff house with a staircase behind the glass, followed by four versions with the staircase removed or kept, each labelled with the model's name.
The house: no staircase, and no reflection of it in the pool.
A source picture of a runner wearing a watch, followed by four versions without the watch, each labelled with the model's name.
The runner: no watch.
A source picture of a gold-armoured warrior in a crimson cloak, followed by four versions without the cloak, each labelled with the model's name.
The guardian: no cloak.
PictureGPT Image 2Nano Banana ProSeedream 5 ProKling Image O3
The houseYes, clean · 4.8%Staircase gone, but that part of the room was walled up · 9.7%Yes, clean · 1.0%No. The reflection went, the staircase stayed · 5.7%
The runnerYes · 0.3%Yes · 0.8%Yes · 0.1%Yes · 7.4%
The guardianYes · 5.5%Yes, and a fire was added on the left · 8.4%Yes · 6.6%Yes, with the figure redrawn · 16.8%

The staircase was the hard one, because its reflection had to go with it. GPT Image 2 and Seedream 5 Pro did both. Nano Banana Pro removed more than it was asked to, and Kling Image O3 removed only the reflection.

Test 2 · Add

Put one new thing into the scene.

The house: a woman in a long coat on the terrace, at the true scale of the building. The runner: a border collie running behind her. The guardian: a tattered banner with a two-headed eagle, hung from the pillar on the right.

A source picture of a cliff house with an empty terrace, followed by four versions with a woman standing on the terrace, each labelled with the model's name.
The house: a person on the terrace.
A source picture of a runner alone on a trail, followed by four versions with a border collie running behind her, each labelled with the model's name.
The runner: a dog behind her.
A source picture of a gold-armoured warrior before a dark pillar, followed by four versions with a crimson banner bearing a golden two-headed eagle, each labelled with the model's name.
The guardian: a banner on the pillar.
PictureGPT Image 2Nano Banana ProSeedream 5 ProKling Image O3
The houseYes · 6.5%Yes · 7.1%Yes · 0.3%Yes, but too large for the building · 1.9%
The runnerYes, though the dog touches her shoe · 5.3%Yes · 3.6%Yes · 3.2%Yes, but the dog overlaps her foot · 13.2%
The guardianYes, a large banner · 6.5%Yes · 4.9%Yes · 4.6%A flag on a pole, not a banner on the pillar · 9.9%

All four added what was asked for, each with a two-headed eagle. The differences are in placement: scale on the terrace, the distance between the dog and the runner, and whether the banner hangs where it was told to.

Test 2 · Relight

Change the light or the weather, and keep the subject.

The house: midday sun under a clear sky, interior lights off. The runner: a grey morning in steady rain, with a wet trail. The guardian: night, the fires burnt out, cold moonlight through the breach.

A source picture of a cliff house at dusk, followed by four versions in midday sun, each labelled with the model's name.
The house: dusk to midday.
A source picture of a runner at sunrise, followed by four versions in rain under a grey sky, each labelled with the model's name.
The runner: sunrise to rain.
A source picture of a gold-armoured warrior lit by fires, followed by four versions lit by blue moonlight, each labelled with the model's name.
The guardian: firelight to moonlight.
PictureGPT Image 2Nano Banana ProSeedream 5 ProKling Image O3
The houseYesYesYesSky changed, but the interior lights stayed on
The runnerYesYesYes, but her hair was restyledYes
The guardianYesYesYesYes, though he is still lit warm

This is the change that asks the most of a model, because everything is repainted and the subject still has to be the same one. The building, the runner and the guardian are recognisably the same in all twelve results.

Both tests · Shadows

Does the shadow agree with the light?

A picture can contain everything it was asked for and still look wrong, because something in it casts no shadow, or the wrong one. These close-ups are cut from the pictures above, in the three places where a shadow can be checked: the runner each model made from the brief, the dog each model added, and the house each model moved from dusk to midday.

Close-ups of the ground under the runner in each model's picture, each labelled with the model's name.
Test 1, the runner: low sun, so she should throw a long shadow on the trail.
Close-ups of the dog each model added behind the runner, each labelled with the model's name.
Test 2, the added dog: the long diagonal shadow is the runner’s and was already in the source picture. The instruction asked for the dog to cast its own.
Close-ups of the house facade in each model's midday version, each labelled with the model's name.
Test 2, the house at midday: the instruction asked for direct sun and crisp shadows.
CheckGPT Image 2Nano Banana ProSeedream 5 ProKling Image O3
The runner throws a shadow on the trailNoYes, and it matches the sunCannot be seen, the ground is darkNo
The added dog casts its own shadowNoNoYesFaint
Midday sun casts crisp shadows on the houseYesSoftYesNo, the light is flat
In the rain, the sunrise shadow is goneYesYesYesYes

No model got all four. Seedream 5 Pro was the only one whose added dog casts a shadow of its own. GPT Image 2 met every item of the runner brief and still left her with no shadow. The added person on the terrace is left out, because the dusk light there is too soft to throw one.

Method

How the changes were measured.

  • One source for everyone. For each change, all four models worked on the same source picture. For the house and the runner the source came from a fifth maker's model, so no model changed its own work. The fifth model declined the third brief, so that source was a picture of our own, made earlier with Nano Banana Pro. In that set one model was therefore changing its own work.
  • Counted, not judged. Each result was compared with its source, point by point, at the same size. Saving the source again without changing it reads 0%, so the figures are the models' own doing.
  • One correction. The first run of the guardian's recolour used a source of a different shape from the results, which inflated the figures. The source was cropped to match and the four results were made again. The figures on this page are from the second run.