Gemini is instructed to reply in images first, and if it thinks, to think using the image thinking tags. It cannot seemingly be prompted to show verbatim the result 4+5 without showing the answer 4+5=9. Of course it can show whatever exact text that you want, the question is, does it prompt rewrite (no) or do something else (yes)?
We can do the same exercises with Flux Kontext for editing versus Flash-2.5, if you think that editing is somehow unique in this regard.
Is prompt rewriting "thinking"? My point is, this article can't answer that question without dElViNg into the nuances of what multi-modal models really are.
Use Google AI Studio to submit requests, and to remove watermark, open browser development tools and right click on request to “watermark_4” image and select to block it. And from next generation there will be no watermark!
Theres lots these models can do but I despise when people suggest they can do edits with "with only the necessary aspects changed".
No, that simply is not true. If you actually compare the before and after you can see it still regenerates all the details on the "unchanged" aspects. Texture, lighting, sharpness, even scale its all different even if varyingly similar to the original.
Sure they're cute for casual edits but it really pains me people suggesting these things are suitable replacements for actual photo editing. Especially when it comes to people, or details outside their training data theres a lot of nuance that can be lost as it regenerates them no matter how you prompt things.
The kicker for nano banana is not prompt adherence which is a really nice to have but the fact that it's either working on pixel space or with a really low spatial scaling. It's the only model that doesn't kill your details because of vae encode/decode.
>Nano Banana is terrible at style transfer even with prompt engineering shenanigans
My context: I'm kind of fixated on visualizing my neighborhood as it would have appeared in the 18th century. I've been doing it in Sketchup, and then in Twinmotion, but neither of those produce "photorealistic" images... Twinmotion can get pretty close with a lot of work, but that's easier with modern architecture than it is with the more hand-made, brick-by-brick structures I'm modeling out.
As different AI image generators have emerged, I've tried them all in an effort to add the proverbial rough edges to snapshots of the models I've created, and it was not until Nano Banana that I ever saw anything even remotely workable.
Nano Banana manages to maintain the geometry of the scene, while applying new styles to it. Sometimes I do this with my Twinmotion renders, but what's really been cool to see is how well it takes a drawing, or engraving, or watercolor - and with as simple a prompt as "make this into a photo" it generates phenomenal results.
Similarly to the Paladin/Starbucks/Pirate example in the link though, I find that sometimes I need to misdirect a little bit, because if I'm peppering the prompt with details about the 18th century, I sometimes get a painterly image back. Instead, I'll tell it I want it to look like a photograph of a well preserved historic neighborhood, or a scene from a period film set in the 18th century.
As fantastic as the results can be, I'm not abandoning my manual modeling of these buildings and scenes. However, Nano Banana's interpretation of contemporary illustrations has helped me reshape how I think about some of the assumptions I made in my own models.
I added a CLI to it (using Gemini CLI) and submitted a PR, you can run that like so:
GEMINI_API_KEY="..." \
uv run --with https://github.com/minimaxir/gemimg/archive/d6b9d5bbefa1e2ffc3b09086bc0a3ad70ca4ef22.zip \
python -m gemimg "a racoon holding a hand written sign that says I love trash"
I just merged the PR and pushed 0.3.1 to PyPI. I also added README documentation and allowed for a `gemimg` entrypoint to the CLI via project.scripts as noted elsewhere in the thread.
In my own experience, nano banana still has the tendency to:
- make massive, seemingly random edits to images
- adjust image scale
- make very fine grained but pervasive detail changes obvious in an image diff
For instance, I have found that nano-banana will sporadically add a (convincing) fireplace to a room or new garage behind a house. This happens even with explicit "ALL CAPS" instructions not to do so. This happens sporadically, even when the temperature is set to zero, and makes it impossible to build a reliable app.
The author overlooked an interesting error in the second skull pancake image: the strawberry is on the right eye socket (to the left of the image), and the blackberry is on the left eye socket (to the right of the image)!
This looks like it's caused by 99% of the relative directions in image descriptions describing them from the looker's point of view, and that 99% of the ones that aren't it they refer to a human and not to a skull-shaped pancake.
Extroverts tend to expect directions from the perspective of the skull. Introverts tend to expect their own perspective for directions. It's a psychology thing, not an error.
> It’s one of the best results I’ve seen for this particular test, and it’s one that doesn’t have obvious signs of “AI slop” aside from the ridiculous premise.
It’s pretty good, but one conspicuous thing is that most of the blueberries are pointing upwards.
For images of people generated from scratch, Nano Banana always adds a background blur, it can't seem to create more realistic or candid images such as those taken via a point and shoot or smartphone, has anyone solved this sort of issue? It seems to work alright if you give it an existing image to edit however. I saw some other threads online about it but I didn't see anyone come up with solutions.
I tried asking for a shot from a live-action remake of My Neighbor Totoro. This is a task I’ve been curious about for a while. Like Sonic, Totoro is the kind of stylized cartoon character that can’t be rendered photorealistically without a great deal of subjective interpretation, which (like in Sonic’s case) is famously easy to get wrong even for humans. Unlike Sonic, Totoro hasn’t had an actual live-action remake, so the model would have to come up with a design itself. I was wondering what it might produce – something good? something horrifying? Unfortunately, neither; it just produced a digital-art style image, despite being asked for a photorealistic one, and kept doing so even when I copied some of the keyword-stuffing from the post. At least it tried. I can’t test this with ChatGPT because it trips the copyright filter.
I have been generating a few dozen images per day for storyboarding purposes. The more I try to perfect it, the easier it becomes to control these outputs and even keep the entire visual story as well as their characters consistent over a few dozen different scenes; while even controlling the time of day throughout the story. I am currently working with 7 layers prompts to control for environment, camera, subject, composition, light, colors and overall quality (it might be overkill, but it’s also experimenting).
I also created a small editing suite for myself where I can draw bounding boxes on images when they aren’t perfect, and have them fixed. Either just with a prompt or feeding them to Claude as image and then having it write the prompt to fix the issue for me (as a workflow on the api). It’s been quite a lot of fun to figure out what works. I am incredibly impressed by where this is all going.
Once you do have good storyboards. You can easily do start-to-end GenAI video generation (hopping from scene to scene) and bring them to life and build your own small visual animated universes.
We use nano banana extensively to build video storyboards, which we then turn into full motion video with a combination of img2vid models. It sounds like we're doing similar things, trying to keep images/characters/setting/style consistent across ~dozens of images (~minutes of video). You might like the product depending on what you're doing with the outputs! https://hypernatural.ai
> The more I try to perfect it, the easier it becomes
I have the opposite experience, once it goes off track, its nearly impossible to bring it back on message
That sounds intriguing. 7 layers - do you mean its one prompt composed of 7 parts, like different paragraphs for each aspect?
How do you send bounding box info to banana? Does it understand something like that? What does claude add to that process? Makes your prompt more refined?
Thanks
I dont get how these tools are considered good when they cant even do a simple thing decribing this scene.
> i was to bring awareness to the dangers of dressing up like a seal while surfboarding (ie. wearing black wetsuites, arms hanging over the board). Create a scene from the perspective of a shark looking up from the bottom of the ocean into a clear blue sky with silhouettes of a seal and a surfer and fishing boat with line dangling in the water and show how the shark contemplates attacking all these objects because they look so similiar.
I havnt found a model yet that can process that description, or any varition, into a scene that usable and makes sense visually to anyone older the a 1st grader. They will never place the seal, surfer, shark or boat in the correct location to make sense visually. Typically everyone is under water, sizing of everything is wrong. You tell them to the image is wrong, to place the person ontop of the water, and they cant. Please can someone link to a model that is capable or tell me what i am doing wrong? How can you claim to process words into images in a repeatable way when these systems cant deal with multiple contraints at once?
> I also created a small editing suite for myself where I can draw bounding boxes on images when they aren’t perfect, and have them fixed. Either just with a prompt or feeding them to Claude as image and then having it write the prompt to fix the issue for me (as a workflow on the api)
Are you talking about Automatic1111 / ComfyUI inpainting masks? Because Nano doesn't accept bounding boxes as part of its API unless you just stuffed the literal X/Y coordinates into the raw prompt.
You could do something where you draw a bounding box and when you get the response back from Nano, you could mask that section back back over the original image - using a decent upscaler as necessary in the event that Nano had to reduce the size of the original image down to ~1MP.
> Once you do have good storyboards. You can easily do start-to-end GenAI video generation (hopping from scene to scene) and bring them to life and build your own small visual animated universes.
I keep hearing advocates of AI video generation talking at length about how easy the tools are to use and how great the results are, but I've yet to see anyone produce something meaningful that's coherent, consistent, and doesn't look like total slop.
65 comments
[ 2.7 ms ] story [ 68.2 ms ] threadokay, look at imagen 4 ultra:
https://aistudio.google.com/app/prompts?state=%7B%22ids%22:%...
In this link, Imagen is instructed to render the verbatim prompt “the result of 4+5”, which shows that text, and not instructed, which renders “4+5=9”
Is Imagen thinking?
Let's compare to gemini 2.5 flash image (nano banana):
look carefully at the system prompt here: https://aistudio.google.com/app/prompts?state=%7B%22ids%22:%...
Gemini is instructed to reply in images first, and if it thinks, to think using the image thinking tags. It cannot seemingly be prompted to show verbatim the result 4+5 without showing the answer 4+5=9. Of course it can show whatever exact text that you want, the question is, does it prompt rewrite (no) or do something else (yes)?
compare to ideogram, with prompt rewriting: https://ideogram.ai/g/GRuZRTY7TmilGUHnks-Mjg/0
without prompt rewriting: https://ideogram.ai/g/yKV3EwULRKOu6LDCsSvZUg/2
We can do the same exercises with Flux Kontext for editing versus Flash-2.5, if you think that editing is somehow unique in this regard.
Is prompt rewriting "thinking"? My point is, this article can't answer that question without dElViNg into the nuances of what multi-modal models really are.
(Do we say we software engineered something?)
No, that simply is not true. If you actually compare the before and after you can see it still regenerates all the details on the "unchanged" aspects. Texture, lighting, sharpness, even scale its all different even if varyingly similar to the original.
Sure they're cute for casual edits but it really pains me people suggesting these things are suitable replacements for actual photo editing. Especially when it comes to people, or details outside their training data theres a lot of nuance that can be lost as it regenerates them no matter how you prompt things.
Even if you
I figured that if you write the text in Google docs and share the screenshot with banana it will not make any spelling mistake.
So, use something like "can you write my name on this Wimbledon trophy, both images are attached. Use them" will work.
https://imgur.com/a/llN7V0W
>Nano Banana is terrible at style transfer even with prompt engineering shenanigans
My context: I'm kind of fixated on visualizing my neighborhood as it would have appeared in the 18th century. I've been doing it in Sketchup, and then in Twinmotion, but neither of those produce "photorealistic" images... Twinmotion can get pretty close with a lot of work, but that's easier with modern architecture than it is with the more hand-made, brick-by-brick structures I'm modeling out.
As different AI image generators have emerged, I've tried them all in an effort to add the proverbial rough edges to snapshots of the models I've created, and it was not until Nano Banana that I ever saw anything even remotely workable.
Nano Banana manages to maintain the geometry of the scene, while applying new styles to it. Sometimes I do this with my Twinmotion renders, but what's really been cool to see is how well it takes a drawing, or engraving, or watercolor - and with as simple a prompt as "make this into a photo" it generates phenomenal results.
Similarly to the Paladin/Starbucks/Pirate example in the link though, I find that sometimes I need to misdirect a little bit, because if I'm peppering the prompt with details about the 18th century, I sometimes get a painterly image back. Instead, I'll tell it I want it to look like a photograph of a well preserved historic neighborhood, or a scene from a period film set in the 18th century.
As fantastic as the results can be, I'm not abandoning my manual modeling of these buildings and scenes. However, Nano Banana's interpretation of contemporary illustrations has helped me reshape how I think about some of the assumptions I made in my own models.
I added a CLI to it (using Gemini CLI) and submitted a PR, you can run that like so:
Result in this comment: https://github.com/minimaxir/gemimg/pull/7#issuecomment-3529...is this just a manual copy/paste into a gist with some html css styling; or do you have a custom tool à la amp-code that does this more easily?
- make massive, seemingly random edits to images - adjust image scale - make very fine grained but pervasive detail changes obvious in an image diff
For instance, I have found that nano-banana will sporadically add a (convincing) fireplace to a room or new garage behind a house. This happens even with explicit "ALL CAPS" instructions not to do so. This happens sporadically, even when the temperature is set to zero, and makes it impossible to build a reliable app.
Has anyone had a better experience?
This looks like it's caused by 99% of the relative directions in image descriptions describing them from the looker's point of view, and that 99% of the ones that aren't it they refer to a human and not to a skull-shaped pancake.
It’s pretty good, but one conspicuous thing is that most of the blueberries are pointing upwards.
I didn't expect that. I would have definitely counted that as a "probably real" tally mark if grading an image.
I also created a small editing suite for myself where I can draw bounding boxes on images when they aren’t perfect, and have them fixed. Either just with a prompt or feeding them to Claude as image and then having it write the prompt to fix the issue for me (as a workflow on the api). It’s been quite a lot of fun to figure out what works. I am incredibly impressed by where this is all going.
Once you do have good storyboards. You can easily do start-to-end GenAI video generation (hopping from scene to scene) and bring them to life and build your own small visual animated universes.
> i was to bring awareness to the dangers of dressing up like a seal while surfboarding (ie. wearing black wetsuites, arms hanging over the board). Create a scene from the perspective of a shark looking up from the bottom of the ocean into a clear blue sky with silhouettes of a seal and a surfer and fishing boat with line dangling in the water and show how the shark contemplates attacking all these objects because they look so similiar.
I havnt found a model yet that can process that description, or any varition, into a scene that usable and makes sense visually to anyone older the a 1st grader. They will never place the seal, surfer, shark or boat in the correct location to make sense visually. Typically everyone is under water, sizing of everything is wrong. You tell them to the image is wrong, to place the person ontop of the water, and they cant. Please can someone link to a model that is capable or tell me what i am doing wrong? How can you claim to process words into images in a repeatable way when these systems cant deal with multiple contraints at once?
https://lmarena.ai/c/019a84ec-db09-7f53-89b1-3b901d4dc6be
https://gemini.google.com/share/da93030f131b
Obviously neither are good but it is better.
I think image models could be producing a lot more editable outputs if eg they output multi-layer PSDs.
Are you talking about Automatic1111 / ComfyUI inpainting masks? Because Nano doesn't accept bounding boxes as part of its API unless you just stuffed the literal X/Y coordinates into the raw prompt.
You could do something where you draw a bounding box and when you get the response back from Nano, you could mask that section back back over the original image - using a decent upscaler as necessary in the event that Nano had to reduce the size of the original image down to ~1MP.
I keep hearing advocates of AI video generation talking at length about how easy the tools are to use and how great the results are, but I've yet to see anyone produce something meaningful that's coherent, consistent, and doesn't look like total slop.