This is amazing. Not that long ago, even getting a model to reliably output the same character multiple times was a real challenge. Now we’re seeing this level of composition and consistency. The pace of progress in generative models is wild.
Huge thanks to the author (and the many contributors) as well for gathering so many examples; it’s incredibly useful to see them to better understand the possibilities of the tool.
Through that testing, there is one prompt engineering trend that was consistent but controversial: both a) LLM-style prompt engineering with with Markdown-formated lists and b) old-school AI image style quality syntatic sugar such as award-winning and DSLR camera are both extremely effective with Gemini 2.5 Flash Image, due to its text encoder and larger training dataset which can now more accurately discriminate which specific image traits are present in an award-winning image and what traits aren't. I've tried generations both with and without those tricks and the tricks definitely have an impact. Google's developer documentation encourages the latter.
Nano-Banana can produce some astonishing results. I maintain a comparison website for state-of-the-art image models with a very high focus on adherence across a wide variety of text-to-image prompts.
I recently finished putting together an Editing Comparison Showdown counterpart where the focus is still adherence but testing the ability to make localized edits of existing images using pure text prompts. It's currently comparing 6 multimodal models including Nano-Banana, Kontext Max, Qwen 20b, etc.
Gemini Flash 2.5 leads with a score of 7 out of 12, but Kontext comes in at 5 out of 12 which is especially surprising considering you can run the Dev model of it locally.
I've come to realize that I liked believing that there was something special about the human mental ability to use our mind's eye and visual imagination to picture something, such as how we would look with a different hairstyle. It's uncomfortable seeing that skill reproduced by machinery at the same level as my own imagination, or even better. It makes me feel like my ability to use my imagination is no more remarkable than my ability to hold a coat off the ground like a coat hook would.
While these are incredibly good, it's sad to think about the unfathomable amount of abuse, spam, disinformation, manipulation and who know what other negatives these advancement are gonna cause. It was one thing when you could spot an AI image, but now and moving forward it's be basically increasingly futile to even try.
Almost all "human" interaction online will be subject to doubt soon enough.
Hard to be cheerful when technology will be a net negative overall even if it benefits some.
The #1 most frustrating part of image models to me has always been their inability to keep the relevant details. Ask to change a hairstyle and you'd get a subtly different person
While I think most of the examples are incredible...
...the technical graphics (especially text) is generally wrong. Case 16 is an annotated heart and the anatomy is nonsensical. Case 28 with the tallest buildings has the decent images, but has the wrong names, locations, and years.
so many little details off when the instructions are clear and/or the details are there. Brad Pitt jeans? The result are not the same style and missing clear details which should be expected to just translate over.
Another one where the prompt ended with output in a 16:9 ratio. The image isn't in that ratio.
The results are visually something but then still need so much review. Can't trust the model. Can't trust people lazily using it. Someone mentioned something about 'net negative'.
Personally, I'm underwhelmed by this model. I feel like these examples are cherry-picked. Here are some fails I've had:
- Given a face shot in direct sunlight with severe shadows, it would not remove the shadows
- Given an old black and white photo, it would not render the image in vibrant color as if taken with a modern DSLR camera. It will colorize the photo, but only with washed out, tinted colors
- When trying to reproduce the 3 x 3 grid of hair styles, it repeatedly created a 2x3 grid. Finally, it made a 3x3 grid, but one of the nine models was black instead of caucasian.
- It is unable to integrate real images into fabricated imagery. For example, when given an image of a tutu and asked to create an image of a dolphin flying over clouds wearing the tutu, the result looks like a crude photoshop snip and copy/paste job.
This is the first time I really don't understand how people are getting good results. On https://aistudio.google.com with Nano Banana selected (gemini-2.5-flash-image-preview) I get - garbage - results. I'll upload a character reference photo and a scene and ask Gemini to place the character in the scene. What it then does is to simply cut and paste the character into the scene, even if they are completely different in style, colours, etc.
I get far better results using ChatGPT for example. Of course, the character seldom looks anything like the reference, but it looks better than what I could do in paint in two minutes.
It's not just you and there's a ton of gaslighting and astroturfing happening with Nano Banana. Thanks to this article we can even attempt to reproduce their exact inputs and lo and behold the results are much worse. I tried a bunch of them and got far worse results than the author. I assume they are trying the same prompts again and again until they get something slightly useful.
I'm pretty sure these are cherry-picked out of many generation attempts, I tried a few basic things and it flat out refused to do many of them like turning a cartoon illustration into a real-world photographic portrait, it kept wanting to create a pixar style image, then when I used an ai generated portrait as an example, it refused with an error saying it wouldn't modify real world people...
I then tried to generate some multi-angle product shots from a single photo of an object, and it just refused to do the whole left, right, front, back thing, and kept doing things like a left, a front, another left, and weird half back/half side view combination.
After looking at Cases 4, 9, 23, 33, and 61, I think it might be suited to take in several wide-angle pictures or photospheres or such from inside a residence, and output a corresponding floor plan schematic.
If anyone has examples, guides, or anything to save me from pouring unnecessary funds into those API credits just to figure out how to feed it for this kind of task, I'd really appreciate sharing.
Man, I hate this. It all looks so good, and it's all so incorrect. Take the heart diagram, for example. Lots of words that sort of sound cardiac but aren't ("ventricar," "mittic"), and some labels that ARE cardiac, but are in the wrong place. The scenes generated from topo maps look convincing, but they don't actually follow the topography correctly. I'm not looking forward to when search and rescue people start using this and plan routes that go off cliffs. Most people I know are too gullible to understand that this is a bullshit generator. This stuff is lethal and I'm very worried it will accelerate the rate at which the populace is getting stupider.
Computer graphics playing in my head and I like it! I don't support Technicolor parfaits and those snobby little petit fours that sit there uneaten, and my position on that is common knowledge to everyone in Oceania.
Unfortunately NSFW in parts. It might be insensitive to circulate the top URL in most US tech workplaces. For those venues, maybe you want to pick out isolated examples instead.
(Example: Half of Case 1 is an anime/manga maid-uniform woman lifting up front of skirt, and leaning back, to expose the crotch of underwear. That's the most questionable one I noticed. It's one of the first things a visitor to the top URL sees.)
59 comments
[ 1.9 ms ] story [ 75.6 ms ] threadHuge thanks to the author (and the many contributors) as well for gathering so many examples; it’s incredibly useful to see them to better understand the possibilities of the tool.
Since the page doesn't mention it, this is the Google Gemini Image Generation model: https://ai.google.dev/gemini-api/docs/image-generation
Good collection of examples. Really weird to choose an inappropriate for work one as the second example.
Is this model open? Open weights at least? Can you use it commercially?
Through that testing, there is one prompt engineering trend that was consistent but controversial: both a) LLM-style prompt engineering with with Markdown-formated lists and b) old-school AI image style quality syntatic sugar such as award-winning and DSLR camera are both extremely effective with Gemini 2.5 Flash Image, due to its text encoder and larger training dataset which can now more accurately discriminate which specific image traits are present in an award-winning image and what traits aren't. I've tried generations both with and without those tricks and the tricks definitely have an impact. Google's developer documentation encourages the latter.
However, taking advantage of the 32k context window (compared to 512 for most other models) can make things interesting. It’s possible to render HTML as an image (https://github.com/minimaxir/gemimg/blob/main/docs/notebooks...) and providing highly nuanced JSON can allow for consistent generations. (https://github.com/minimaxir/gemimg/blob/main/docs/notebooks...)
I recently finished putting together an Editing Comparison Showdown counterpart where the focus is still adherence but testing the ability to make localized edits of existing images using pure text prompts. It's currently comparing 6 multimodal models including Nano-Banana, Kontext Max, Qwen 20b, etc.
https://genai-showdown.specr.net/image-editing
Gemini Flash 2.5 leads with a score of 7 out of 12, but Kontext comes in at 5 out of 12 which is especially surprising considering you can run the Dev model of it locally.
Almost all "human" interaction online will be subject to doubt soon enough.
Hard to be cheerful when technology will be a net negative overall even if it benefits some.
..guess that's solved now.. overnight. Mindblowing
...the technical graphics (especially text) is generally wrong. Case 16 is an annotated heart and the anatomy is nonsensical. Case 28 with the tallest buildings has the decent images, but has the wrong names, locations, and years.
so many little details off when the instructions are clear and/or the details are there. Brad Pitt jeans? The result are not the same style and missing clear details which should be expected to just translate over.
Another one where the prompt ended with output in a 16:9 ratio. The image isn't in that ratio.
The results are visually something but then still need so much review. Can't trust the model. Can't trust people lazily using it. Someone mentioned something about 'net negative'.
- Given a face shot in direct sunlight with severe shadows, it would not remove the shadows
- Given an old black and white photo, it would not render the image in vibrant color as if taken with a modern DSLR camera. It will colorize the photo, but only with washed out, tinted colors
- When trying to reproduce the 3 x 3 grid of hair styles, it repeatedly created a 2x3 grid. Finally, it made a 3x3 grid, but one of the nine models was black instead of caucasian.
- It is unable to integrate real images into fabricated imagery. For example, when given an image of a tutu and asked to create an image of a dolphin flying over clouds wearing the tutu, the result looks like a crude photoshop snip and copy/paste job.
I get far better results using ChatGPT for example. Of course, the character seldom looks anything like the reference, but it looks better than what I could do in paint in two minutes.
Am I using the wrong model, somehow??
which goes to show that some of these amazing results might need 18 attempts and such.
[0] https://imgur.com/a/aSbOVz5
I then tried to generate some multi-angle product shots from a single photo of an object, and it just refused to do the whole left, right, front, back thing, and kept doing things like a left, a front, another left, and weird half back/half side view combination.
Very frustrating.
[1] https://github.com/JimmyLv/awesome-nano-banana
If anyone has examples, guides, or anything to save me from pouring unnecessary funds into those API credits just to figure out how to feed it for this kind of task, I'd really appreciate sharing.
(Example: Half of Case 1 is an anime/manga maid-uniform woman lifting up front of skirt, and leaning back, to expose the crotch of underwear. That's the most questionable one I noticed. It's one of the first things a visitor to the top URL sees.)