AI for landscape visualisation: 5 image generators tested
As a 3D visualiser for landscape design, I find myself asking this more and more:
How can AI image generators actually support visualisation in landscape design?
Not just as inspiration or quick conceptual sketches, but as visuals based on the actual design can that you can use to present your landscape design to clients.
From what I noticed while testing, the results can look impressive at first glance, but don’t often hold up when you compare them to the actual design. Layouts shift, materials change, details get lost and AI-slop get added.
I put it to the test and tried 5 different image generators.
Test setup: same input, different models
To put different AI image generators to the test, I setup a quick experiment to see if the images would be usable (without further altercations) in a client presentation.
I used a street view image of a typical Dutch polder landscape and tried to create a before/after scenario. The ‘before’ scenario contained a road, a narrow canal and some trees in the foreground and a grassland and a few buildings in the background. And I wanted the ‘after’ scenario to have a flooded wetland with reed grass and birds instead of the grassland. To create a usable before/after scenario I prompted the AI models to keep the perspective the same Here is the image I used:
Test
I compared 10 image generator that use a reference image and a prompt to generate a new image. The goal? To see which is most accurate and useful for a before and after scenario, and which comes best out of the test when comparing result vs cost.
Results
Model: Krea 1
+ Suprisingly realistic in terms of the landscape, only a bit repetitive maybe.
+ lighting and colors look good.
+ useful for further upscaling and photoshopping.
+ 4 generated images.
+ only 14 credits!
- Buildings are different, but not bad per se
- It’s interpreting some weird pyramids from a mount of gravel in the original image.
- Doesn’t match original perspective.
- Missing described birds
- Cyclist looks funny.
- Added white lines and weird looking stuff on the left side.
Overall: It’s the first time I’ve tried this model and I’m honestly quite surprised by it. It looks realistic and has good colours and lighting. It does come up with some funny things like the gravel pyramids and a strange looking cyclist. It also misses some of the prompted suggestions like adding specific birds and keeping the same perspective. Seeing as it’s just 14 credits for 4 (slightly different) generated images it’s worth having a closer look at this model!
Model: Nano Banana Pro
+ Very realistic image.
+ Kept the buildings in the background (and added some extra).
+ Basically exactly as described in the prompt.
- Trees are different and have moved.
- Road and water transition in the foreground looks a bit funny.
- Green and yellow colours slightly oversaturated.
- Perspective is different from the original image.
Overall: As I’ve used a couple of times before, I knew it would probably give a good result. And it doesn’t disappoint. It generates exactly was prompted, so that’s great. But it also takes it a step further and changes things that weren’t described, like the trees and the buildings. Fixing that may require a better/stronger/clearer prompt though… It’s one of the more expensive models, but the results are worth it, I think!
Model: Flux 2 Max
+ correct birds species added.
+ Lighting seems correct.
+/- kind of realistic landscape
+/- interesting combination of the original and newly generated elements.
- Road disappeared
- Funny looking perspective
- Landscape repetitive
- Colours are oversaturated.
Overall: The trees from the original image are kept and the lighting seems correct. The birds are also as prompted, but other than that, it isn’t great. Funny thing is, it looks like it keeps the same perspective as the original image but just goes wild with adding and removing stuff. So for the number of credits this costs (=112), it’s a shame that it doesn’t really generate a useful image.
Model: Seedream 5 Lite
+ Keeps to the original perspective.
+ landscape plants and water look alright.
+ Generates exactly what is prompted.
+/- Lighting and colours look good in some parts, in others not so much.
+/- Creates more an overlay than a newly generated image
- places elements in illogical sports
Overall: It looks weird at first, with that bird standing in the puddle in the middle of the road, but it’s not as stupid as it looks. The model actually generates what was prompted (a birds standing in a puddle), it’s placement just isn’t really logical… So, other than that, if we’re really looking for the before after scenario, this image can be quite useful as it just replaces the grassland with wetland and adds the birds. And with 2 images for only 55 credits, it interesting to see if it can do this consistently, without the random birds..
Model: ChatGPT 1.5
+ Keeps to the original perspective.
+ Generates what is prompted.
+ lighting and colors look good.
+ correct birds species added and correctly placed
- doesn’t match the original perspective
- buildings and trees have disappeared
- most expensive of all models
Overall: I had used ChatGTP to add some photorealistic effects before, so I knew that the realism of the images was quite good. And this test proves it again. It gives you exactly what you describe, but that’s also where you’d have to be careful because it removes anything that wasn’t prompted. For the 316 (!) credits, it’s a very useful image in terms of before/after, the only comments would be that the perspective doesn’t quite match and that it can quickly consume your credits when prompting revisions of this.
Conclusion
I think the clear result is that generating images like this doesn’t give you exactly what you would want in terms of a before-after scenario. Most generated images don’t match the perspective and if they do, they lack in either realism, placement of objects, repetitiveness, colours or just general AI-slop.
Some of the results, however, like the Krea 1, Nano Banana Pro and ChatGPT 1.5 images can still be very useful for further editing. Adjusting the original prompt and generating multiple images will give you a lot of options for creating a decent before-after scenario.
When talking about generating multiple images the costs also come into consideration of course. It’s clear that the more expensive models give more realistic and accurate images but generating multiple images of the same scenario from image models like Nano Banana Pro and ChatGTP 1.5 would quickly cost you a lot of credits. The Krea 1 model is a surprising exemption to that rule by the way!
To decide which AI image model is most useful based on this test, I’ve made a short ranking of the 5 tested image generators:
| Model | Developer | Credits |
| 1. Nano Banana Pro | Google DeepMind | 238 |
| 2. Krea 1 | Krea AI | 14 |
| 3. ChatGPT 1.5 | OpenAI | 316 |
| 4. Seedream 5 lite | ByteDance | 55 |
| 5. Flux 2 Max | Black Forest Labs | 112 |
Nano Banana Pro is in my opinion the winner in this round of testing. It’s one of the most expensive ones, but the result is impressive. Maintaining the perspective while generating everything that is prompted and even that the buildings are important is an example of thinking and understanding that I haven’t seen in the others.
Krea 1 definitely deserves the second place as its price vs. result alone is already worth it. It doesn’t generate exactly as prompted but with a bit of editing you’ll surely have a decent before-after scenario.
Third comes ChatGPT 1.5. The realism is great but it’s disappointing that it doesn’t match in perspective and that it filles the area completely, but that might be a prompting issue…
Fourth place is for the Seedream 5 lite as it keeps to the original perspective and does as prompted, but misses out on detail, realism and colours.
Then finally in last place there’s the Flux 2 Max, it vaguely refers to the original image and interprets the prompt in the least logical way. Besides that its realism and colours also aren’t great while it’s quite expensive compared to the other image generators.
What’s next?
An important factor to consider in generating images is the consistency of the models. It’s something I haven’t really tested yet and is an interesting subject for another round of testing. The biggest question would be: Can an AI image model produce the same image with slight modifications in the prompt? This would be useful in revising an image without having to edit it a lot afterwards.
I’ll also be checking out some more image generators, existing ones as well as the new (or upgraded) ones that pop up!
Thanks for reading and I hope it’s useful to you and your projects!
Disclaimer:
I’m not affiliated with any company, developer, or AI related provider. I do not earn money or receive any benefits by writing this article. This test and the results are purely my own experience and opinion of the current generative AI image models.
Over de auteur
Timo Hoogstra
Ik ben een landschapsontwerper die zich heeft gespecialiseerd in 3D-visualisatie. Ik maak gedetailleerde modellen voor landschaps- en stedelijk ontwerp. Ik werk met SketchUp, Rhino en Revit, en breng modellen tot leven met Lumion of V-Ray. Je krijgt nauwkeurige, meeslepende 3D-beelden—met duidelijke communicatie en ruimte voor revisies.