Gemini 2.5 Flash (more commonly known as the “Nano Banana” model) was released in August, but I haven’t had much chance to experiment with it until now. Like ChatGPT’s image generation tool, Nano Banana can generate and edit images in a chat format. In my testing, I found that the images look more realistic than ChatGPT, and the characters stay more consistent. However, I struggled to get Nano Banana to change the angle of images and had mixed results generating scenes with multiple characters.
Nano Banana is a tool I will use in my work, but only in combination with other tools. That seems to be what most of the experts in AI image and video generation are doing: they switch between multiple tools depending on their strengths and weaknesses or when a particular tool fails.
Character generation
With Nano Banana, you can prompt with sentences and paragraphs in natural language, similar to how you would prompt an LLM. The image prompt structure I typically use still works here, but you don’t have to be as strict about removing filler words as you do with tools like Midjourney.
To experiment with character generation, I generated an initial image on a white background. Then, I prompted it to change the expression. “Keep the character details the same (hair, glasses, blouse, etc.) but change her expression to [a slight frown; a confused expression; a frustrated expression.” I found that if I didn’t remind it to keep the details the same that it shifted too much, but it was pretty consistent as long as I asked for that. I was able to generate multiple images fairly quickly.
Her makeup is different in the initial image, but the rest of the series kept the details more stable. Her skin tone changes somewhat, but I think it wouldn’t be that noticeable in a scenario where you only saw one image at a time instead of looking at the side-by-side like this.
At one point, Nano Banana reverted back to an earlier image from the same chat. I had to upload one of the images I wanted as a reference image to reset it from drifting away. There’s much less drift than ChatGPT, but it still needs reminders sometimes.
I zoomed out the image to show her from the waist up talking, and then I generated a few side view images. Again, it kept the character consistency pretty good for these images.
Scene with multiple characters
To test a scene with multiple characters, I tried to generate something similar to my previous tests with ChatGPT image generation. The scenario is a graduate student and a professor having a conversation in a conference room.
I was pretty happy with the initial generation of the graduate student taking notes and of the same character talking. Her notebook disappeared in the second image, but the character consistency was pretty good.
Then, I zoomed out the image to show more of the conference room before adding another character. I’ve seen recommendations from others to make one change at a time for these kinds of edits, which make sense to me to maintain consistency.
My first attempt at adding a second character on the other side of the table was hilarious, but not useful. The professor is in the table instead of sitting at a chair behind the table.
I prompted to “Change the professor so it looks like he is sitting in a chair, not floating on top of the table.” She doesn’t look like she’s looking right at him for a conversation, but at least it doesn’t ignore the laws of physics.
When I tried to change the camera angle to generate an over the shoulder shot, it understood how to add the shoulder. However, it didn’t rotate the camera. In fact, none of my attempts to get different camera angles in this scene (other than zooming out) were successful. Somehow, the professor is simultaneously in two places. We see the camera behind his shoulder, but the rest of the image hasn’t rotated to a new perspective.
I’m sure at least some of this failure is due to my prompts; there may be some other tricks to getting it to change the perspective. However, some quick searching online reveals that others are also having trouble with changing perspective in Nano Banana. Sometimes it works, and sometimes it doesn’t, and it’s not clear why the results are mixed. I obviously was able to get different angles with one isolated character but not with multiple characters in a scene.
It may be that this is just a limitation of the tool right now. Changing to a different tool like ChatGPT or Runway to generate an image in a different perspective might get better results, especially for multi-character scenes.
Of course, the tools are always improving. In six months, I may get more consistent results with multi-character scenes.
More on consistent character image generation
To see how Nano Banana compares to other tools, check out my examples and tips for generating consistent character images.
Upcoming events
Writing Scenarios: Compelling Characters and Distinctive Dialogue
Set the Stage: Make Elearning Relevant and Authentic with Scenarios
Rapid Prototyping Branching Scenarios
Stock Photos to Stunning: Generate Custom Images with AI
Friday, November 14 at 10AM PacificDevLearn in Las VegasSave an extra $100 off your DevLearn registration with code CHRISTY.
