Google’s latest model, Nano Banana, is promising for use cases like marketing, branding, and storytelling due to its generation speed and consistency across many images.
TLDR
- Nano Banana produces consistent results for exact image replication or positional/angular variations of input images. It also reliably writes & edits text.
- I prefer Midjourney for style transfer and “raw” art creativity.
- Sora editor kept crashing 💀
**
I started watching the anime Dandadan and was blown away by the art style. I wanted to see if I could emulate a Dandadan-like manga scene using Nano Banana.
Scenes from Dandadan:
Explorations
Before trying to one-shot a manga scene, I wanted to take some time to explore the capabilities of Nano Banana (particularly in comparison to Midjourney).
First, I messed around with “panel completion” — I took an existing manga page, removed the last panel, and prompt the model to make the last panel with a text description of the scene. The purpose of this was to test how well Nano Banana could do art style transfer.
Not bad.
I then started messing around with multi-panel panels. Prompt (plus a picture of Killua from HxH as a character reference):
Create a 6-panel manga sequence. Make sure the character and art style is consistent in black and white.
Panel 1: "Create a young manga protagonist discovering a mysterious artifact"
Panel 2: "Same character examining the artifact closely - keep character design identical"
Panel 3: "Wide shot of character being surrounded by ominous magical energy and a glimpse of a massive demon-like alien"
Panel 4: "Close up of the character's eyes very wide in fear, filled with dread"
Panel 5: "A front, straight-on view of character running frantically from the alien - same art style"
Midjourney had a more authentic anime art style in my opinion and the panel framing was also very clean, but all the text was gibberish. I didn’t like the art style of Nano Banana as much (it kind of reminds me of those printable black and white coloring pages) but the panels made “more sense” as a sequential scene.
I gave Midjourney the same prompt but this time without the Killua reference image. Again the words were gibberish but the actual art and character design were pretty stunning and creative.
**
One thing that I really liked about the Midjourney prompt terminal was its separation between Image Prompts, Style Prompts, and Omni Reference. On the other hand, Google’s AI lab wasn’t very reliable at style transfer and would sometimes just return one of the images 1:1 without any edits.
Generally, I think there’s a lot of room for novel AI editor interfaces. Especially in Google AI Studio, it was extremely annoying to (a) try to reference 1 character when there were multiple characters in an image, (b) differentiate which images were for style and which were for character consistency, and (c) rapidly iterate – aka parallelize image generation.
**
I then started experimenting a bit with character design. I first gave Nano Banana some images of characters from HxH and Dandadan.
I am writing a manga and need your help to visualize the story. The art style should be consistent, only black-and-white art and no words.
The story has two main characters, a boy and a girl, and is set in Japan. Please design these two characters for me. I've attached some examples of characters that I like in other mangas.
Make a full body character profile of the boy and the girl for me to use throughout the story.
Using these profiles as a starting point, I wanted to make reference sheets in an attempt to keep characters consistent across scenes. Nano Banana is very good at character consistency – especially variations given an input character. On the other hand, Midjourney is quite bad at this.
**
I was pretty surprised at Nano Banana’s ability to edit text. While keeping everything else the same, Nano Banana successfully changed “candiat to candid”:
**
After all this fiddling around, I made “The Coup” using a mix of Nano Banana and Midjourney:
It’s way shorter than I intended but it was pretty hard to keep character consistency across many scenes in the way I wanted to (notice the style isn’t super consistent either…). I ended up using a combination of Nano Banana for character/style consistency and Midjourney to “seed” the initial style and character design.
**
Interesting directions to explore in the future:
- Train a model specifically for style as a Nano Banana post-processing step
- Better AI editors for storytelling (e.g. a storyboard phase that generates reusable context prompts for consistency across scenes)
- Character consistency (e.g. upload characters that are passed into every scene generation’s context window)
- Post-generation editing (e.g. layers, text, etc.)