Posing Characters in Scenes with Google's Gemini Nano Banana

Many of the AI video generation tools in ComfyUI require a starting image for the shot: you provide a start image (and in same cases and end image) and describe the actions to generate a video clip. This means forming the starting image correctly is very important. So then how to pose a character?

There are several models that can do this. Recent versions for ComfyUI support “API nodes” which allow you to set up ComfyUI billing and then use API calls to cloud services, such as Google Gemini “Nano Banana” (an internal project name that got traction externally as well). At the time of writing, two levels of Gemini are available, 2.5 and 3.0 (aka “Pro”). Pro is generally better, can generate higher resolution images, and of course costs more to use. Most of the images below cost around 4 cents with 2.5 and 14 cents with 3.0 (Pro). While not that expensive, it does add up as you frequently need to retry multiple times.

In the following example, I provided 3 input images (left) - a really rough sketch (from a storyboard I crudely drew with a mouse), a character reference sheet, and a background - and it generated the final result (right).

Here is the prompt I used: “Make the male robot in image 2 pose according to the simple sketch in image 1. Only the upper torso of the robot can be seen. Place the robot in front of the background image 3. The robot is shorter than the door in the background scene.” Pretty cool huh? It did a pretty good job even with a terrible sketch.

But there were some gotchas and problems that I have not yet solved. I don’t have the compositional control I want yet.

First little gotcha. I had to resize (pad) the character reference image so it had the same aspect ratio. Without that the image chopped the head and feet off the robot, ending in a less than desirable result (it made up the head to add).

The bigger problem however is positioning the character. First render the reference pose was too high, making things feel out of proportion. The robot would not fit through the door. So while it got the reference pose, you still have to worry about scale and placement.

Also, the horizontal positioning was still not correct. In cinematography, there is a “rule of thirds” where important things often look better on a third reference line (you put a tic-tac-toe board over the top. You can see it in the sketch I made.

But the final image above the character was centered (allowing for the arm). This is a common problem with AI image models - they have a tendency to center things.

So I explored some other avenues. First, I asked Gemini 2.5 to output the character with a transparent background.

Can you see the problem? Look up in the top left corner. The checkerboard pattern is not regular! Actually, the checkboard is not transparency at all, it is a checkerboard pattern baked into the image!!

Another approach is to remove the background then manually place the character onto the background. The following is an example of the time you can waste trying to find a path that works.

First, I tried to remove the background from the above image using a node in ComfyUI. It missed under the arm.

Next, I tried on a greenscreen. Of course it decided to not obey the pose.

Trying again finally worked.

This is common. I find a lot of the work is not just to get something to work once, but find sequences that work more reliably.

With the solid color behind, the background removal works better, but not perfect. The antenna on his head got lost, and sometimes you can get a green glow around the character as the image was not generated taking the background into account. There are of course other background removal algorithms, but it is a common problem, especially if there is blending between the edge and the background, which is common to resolve anti-aliasing.

Bottom line is precise control with Nano Banana is not something I have achieved yet with a satisfactory, repeatable, workflow. That, and it costs money for every experiment. Some things it does great (extracting pose from low quality sketches), but without a complete working workflow I am not planning on using it for shots where I want detailed control over shot composition.

This does not mean models like Nana Banana are never useful, just that for all the hype you hear about how fantastic these models are, if they do not do what you need, then you need to keep looking.

Here is an example from other test with Nano Banana Pro. Giving it a room and two characters, it generated pretty consistent views of the characters from different angles, at 4K, allowing me to resize/crop different starting images for shots. This is useful when I want a sequence but don’t need complete control.

Want thirds? Snip at the right position!

A cheat? Maybe, but its practical. And you can take the image and improve its quality by feeding it back through another ComfyUI workflow. It may be a good Nano Banana path is to put up with the character being centered, then crop the image afterwards.

Next experiments? I want to try the same with another free open source model - QWEN Image Edit. And I also want to try a scene with multiple characters. But that is for another post.