TLDR for those who want to cut right to the chase:
- Use Nano Banana
- Use it through a different tool like Dzine or Higgsfield
- Do LOTS of generations (don't give up, it'll come out like you want in the end)
- Come up with a storyboard to decide all the frames / angles you need before getting started
Making consistent characters is the first step to any AI-generated video that requires any kind of characters.
For my latest series (aka "Meet the Praetorians" or MTP) I'm doing videos in the style of LADBible Stories or the Soft White Underbelly interviews, which feature a specific character who gets interviewed for the entire duration of the video. Since there are multiple different camera angles, it's important to create a consistent character for the narrative to stay believable throughout.
In this article I'll share how you can use Google's Nano Banana image generation model to create consistent characters. Nano Banana is available from within Google Gemini and the allowed number of image-generation credits vary based on your Google One plan. You can also get access to Nano Banana in most of the AI tool aggregators like Dzine, Higgsfield, OpenArt, and others.
I have found, in fact, that using Nano Banana directly in Google Gemini is sometimes worse than using it through other tools. For example, it'll often reject a prompt to make alternates of an image because it says it's a "recognizable person" (even though in my cases they're all AI-generated!), while when using Dzine I've never had this problem.
My reference for this article is Hodlor. The reference image below is from a video generated using Google Veo 3.
Before getting started, it's important to have a rough storyboard so you can determine what angles / frames you need to create.
Once I have my storyboard, and using the reference image, I asked Nano Banana to create a profile image of the character with the people in the background removed:
I wanted his skin to be a bit more realistic and textured:
Next, I zoom out a few levels:
While I'm at it, I get a view of the empty bench to use as my starting frame for the "walk in" video:
Angles are next. Now another issue I found is with background consistency - if you want to have a character stay consistent AND the background stay consistent when the camera MOVES, that's something Nano Banana may struggle with.
Take this example, where I've asked Nano Banana to make a three quarters angle view:
You can see that while it succeeds at moving the character, it's clear the street in the background hasn't moved, so it won't be believable if used in the same video as my profile shot. I struggled quite a bit getting Nano Banana via Gemini to correctly shift the background, so I tried on Dzine and was able to get it working (though it took many attempts). Another (big) advantage of using these other tools is you can generally submit multiple concurrent jobs simultaneously, which is hard to do in the Gemini interface. In fact, I'd say that's really the best advice I can give on these - try a lot of times to get it right. If you hope
Here are a few examples (done in Dzine) where we can see that the houses on the side of the street are at a more believable angle, so while it's not perfect it's more believable with the shift in the background.
With my various angles ready to go, I'm now working on getting the audio files through ElevenLabs, then lip sync via Dzine, and finally stitch things together in Davinci Resolve. Stay tuned for the Hodlor episode.
The art on this page was create using Nano Banana AI