Nano Banana: Massive Hypes, Nano Satisfaction

So with all the insane hype surrounding "nano banana" before its release, I was actually interested to try it out and see what it could do. What I found was a classic case of a potentially cool piece of tech being completely crippled by its own company.

Here are my grievances after using it on both AI Studio and the Gemini browser:

First off, the model is extremely censored to the point of being stupid. For example, my screenshots. I tried to get it to edit my friend's face into the "Giga Chad" meme, and it flat-out refused multiple times because it thought I was trying to maliciously edit a "real person." It has zero understanding of internet culture, memes, or stylistic intent. This aggressive filtering makes it useless for a huge range of creative and fun edits.

The worse part is something I've never really seen before: an image model that "hallucinates." It often fails to even understand instructions. You'll ask for an edit and it will do nothing at all, ignore parts of your request, get stuck in a loop, or add stuff you never asked for. Sometimes it produces a completely different image or changes the style randomly for no reason.

The user interface on Google's own platforms is absolute garbage. It wasn't designed for continuous editing. On top of that, the underlying model is clearly not like GPT or Sora; its natural language understanding is very limited. It lacks the nuance for proper editing context, has almost no memory of previous steps, and consistently fails to follow multiple instructions in a single prompt.

Multi-image editing was awful. I tried something as simple as "add the hat from image 1 to the person in image 2 and make it look natural," and most of the time it just did a bare-minimum crop and paste job, with no blending or adjustment. It also really struggles to change the style of illustrations, like turning a 2D object into 3D or making it look realistic.

Frustrated, I decided to try it on LMArena, where all the hype started. The UI there is the bare minimum and obviously not for real editing, but the actual function of "nano banana" finally started to work. The censorship is reduced, and when the edits actually go through, it's pretty fun to play with.

The process is still horrendous though, you can only do one edit per chat, and then you have to copy the new image and paste it back in to edit it again. Jesus, that's awful.

But here, I finally saw what people were excited about:

• Style Consistency: It sticks to the current style of an image really well. When it adds new details, it tries hard to match the existing aesthetic.
• Lighting and Reflections: It's surprisingly good at how light and surface reflections work. I was shocked when it managed a realistic take on an object half-submerged in water, the rippling refractions were visible and looked really cool.
• Different Angles: It's also pretty good at generating different angles or views of the same subject, distances too.

None of these features are brand new, but it's good at them. In longer editing chains, it will eventually start to hallucinate and change details, but it continues to stick to the image's overall style surprisingly well, even as the consistency of other details degrades.

The Manufactured Hype and The Tribalism

So, I want to quickly talk about this. Before its release, the hype for this thing was unreal, especially on social media, Reddit, and even among tech 'journalists' who were happy to run stories based on impressive LMArena scores and curated demos. From the very beginning, it felt very manufactured. Even the name they chose, "nano banana," was clearly a deliberate choice for something memey that could go viral easily.

The primary weapon in this hype campaign was the constant boasting about LMArena results. But let's be real, LMArena is not proof against astroturfing or paid operations. The hype claimed this model was the new "GPT killer" and that its image generation capabilities were the best around, like it was some massive new frontier model.

The funniest part, which was later revealed by Google, was that the "incredible" new generations were just secretly using Imagen 4. "Nano banana" was only being used when you specifically asked it to edit an image. This shows how much these so-called "testers" on LMArena actually knew. They were praising an editor for its generation skills, skills it didn't even have. The entire foundation of the hype was built on a misunderstanding, or a straight-up lie of omission.

Then, we have the tribalism of GenAI, it got so bad, especially on places like Reddit. You'd see people literally foam at the mouth whenever "Patrick Logan" or other Google devs would tweet something vague. These are the exact same people who would shit on Sam Altman when he hypes his own stuff. The hypocrisy was insane. They'd post screenshots of a developer's tweet and worship it, it was a mentality of teenagers and their console wars.

Even now that the model has been released and its issues are obvious, the defenders are out in full force. You know the pattern, it's the same old same old:

• "You're using it wrong."
• "It works fine for me."
• "Oh, but it's great at XYZ feature!"