What is nano-banana?!

I’ve been testing a new AI image editing model with some impressive capabilities.

In my post about my character-creation workflow, I demonstrated how one of the most important (and sometimes challenging) steps is generating images that faithfully represent the characters in new situations. These images can then be used as the starting frame for generating video.

I’ve used Midjourney, Runway, and ChatGPT for this. Each of them does a more-or-less decent job of recreating a character’s likeness in a new scene. But there is a new AI model that blows them all away, at least in my testing so far.

There is a website called LMArena where you can go to rank the outputs of different AI models in relation to each other. Each time you enter a prompt in the website's battle mode, you are presented with two AI-generated outputs and are asked to choose the one you prefer. A few days ago, visitors to LMArena noticed a new model had been added that edited images while preserving details better than any known model. This mysterious new model was labeled only as "nano-banana." People started to speculate that it was a new model from Google, a rumor which seems to have been confirmed.

Google will likely reveal it publicly soon (hopefully with a better name), but in the mean time, the only way to use it is in LMArena's battle mode. I tried it out to see how it fit into my workflow.

When I got the unlabeled results from two different models, it was always obvious which of the competing models was nano-banana—it was the one that looked like my character. Other AI image editing models will frequently change the features of people in their transformed images. But nano-banana did a great job preserving Maya’s likeness.

See for yourself; I provided as input Maya’s character sheet and a Midjourney image of her dancing through a meadow:

And this is what I got out:

I was very impressed by the model’s ability to represent her features faithfully!

A New Character

I decided to try nano-banana with another character I’ve been working on. Dax is a robot friend and companion for Carter. He is a flying robot, so I wanted him to look like a futuristic drone with a face. Surprisingly, I had a really hard time getting Midjourney or ChatGPT to generate realistic images of a drone. They always came out with their rotors distorted, or otherwise asymmetrical.

Thankfully, nano-banana did a much better job. Not only did it follow my instructions without messing up the original image, it also cleaned up the imperfections and distortions in the original!

This was my input, from Midjourney; notice how the rotors look kind of twisted and misaligned:

I also provided an image of a toroidal rotor (a clever design that promises to make drones quieter):

This was the output from nano-banana:

It didn't manage to integrate the toroidal rotors, but it really cleaned up the artifacts and distortion around the traditional rotors, and added the face while keeping the rest of the design precisely intact. I was pleased!

Midjourney Video

I’ve been using Midjourney’s built-in video features more. One reason is that Veo 3 is constantly giving me problems when I try to upload images of my characters. It erroneously flags them as images of prominent people almost every time. I’ve tried quite a few fictional characters, but they always get caught up in the filter.

At the same time, Midjourney’s video generation is really impressive. I like how it can extend a video and make it feel like one continuous take. You can ask it to automatically extend a video, add text to describe the action, or even set an end frame and the model will interpolate between the starting and ending images as naturally as it can.

I took the image of Maya dancing above and extended it twice without changing the prompt for a total of 9 seconds. Then I extended it again, using a nano-banana-generated image as the final frame. It turned out pretty well. The camera motion change feels abrupt, and in the last few frames, the flowers in the meadow change to match the last frame, but considering the difficult constraints of what I asked it to do, it was a pretty great!

It's an exciting time to be an independent filmmaker, with powerful new tools like these being released so frequently!