Speed Up Character Image Creation with Flora

One of the AI tools I’ve been using more lately is Flora. Instead of a chat interface like many AI tools, Flora gives you a canvas with multiple nodes where you can see and organize multiple results. Flora includes most of the main image generation models, so you can switch between tools easily. For my post comparing how different image models handled the same prompt, I used Flora’s built-in feature to run a single prompt with multiple models at once. In this post, I’ll show how I used Flora’s batch feature to generate character images in multiple poses quickly. After experimenting with this, I think this process can really speed up character image creation for scenarios. You can generate multiple poses of a character with consistent aspect ratios and even transparent backgrounds much faster than if you prompt one at a time in other tools.

This isn’t a sponsored post; I’m sharing my own experiences so you can see what worked for me. I’ll share some of my mistakes and lessons learned as well.

Reference image

The first step for generating character image sets is always generating a reference image. If your reference image isn’t good,

In this case, I wanted to replicate the visual style from another image I’d previously generated for a presentation. I uploaded that image to Claude and asked it to describe the style so I could use it in an image prompt. Claude gave me much more detail than I wanted, but I picked several of those phrases to include in my prompt.

I used a structured prompt with labels. Nano Banana and ChatGPT work well with this kind of prompt. It’s simpler than JSON prompting, and you get some of the same benefits of repeatability without needing to be so precise in your syntax.

Style: Contemporary digital painting, visible brushstrokes with flat color blocking, realistic body proportions, acrylic painting blended with vector-style cel shadingSubject: A Latino man, mid 40s, black hair, well-trimmed short beardClothes: Navy blue button-down shirt, khakis, brown Oxford shoes, black glassesPose: Standing, looking at the camera, smiling, relaxed poseBackground: Solid white

Here’s my reference image using Nano Banana 2 after a few attempts. I tried the ChatGPT Image 2.0 model too, but that was much slower.

Pose list in CSV

Next, I wrote a list of poses and saved them in a CSV file. I used 8 simple poses for my examples for this post, but Flora can handle up to 100 at once in a batch. Practically speaking, I probably wouldn’t do more than 20 images at a time. If something goes wrong and you need to adjust your prompt or settings, it’s easier to adjust with smaller numbers of images.

Talking gesturing with handsJoyous laughingListening thoughtful arms crossedDisappointed slight frown looking downFrustrated upsetConfused head tilted slightlyTurned to the side talking and gesturing with handsTurned to the side listening with arms crossed

I uploaded the CSV with the list above to a Batch node in Flora. Flora separates out each row into a separate prompt.

Combine the reference image and batch prompt list

This is where Flora really speeds up the process of character image generation. Instead of prompting 8 different times, I connect the node with my character reference image to my batch of prompts. With the batch, I can run 8 generations with a single click.

Running the batch of 8 prompts gave me these variations on the original character image. Note that even though it’s a single node on the right, there are actually 8 variations all stacked in the same node. Click each image in the gallery below to view a larger version. (If you’re reading this in email, you may have to view it on my blog instead.)

Transparent backgrounds

While I often generate scenario images of characters directly in scenes with context, sometimes it’s easier to work with separate character images with transparent backgrounds and then manually place them against a background image later. Flora makes this easy; removing the background is one of the tool options on image nodes. Since all of my character variations are in a single node, I just have to click remove the background in all 8 images in one step.

Repeat the process

I generated another reference image for a character named Maia and used the same batch prompts for the variations. This time I used a 3:4 aspect ratio instead of a 1:1 ratio. I don’t know why Nano Banana matches the aspect ratio just fine in Flora but ignores it in prompts in the regular interface. But I get consistent ratios in Flora.

I also generated a set of images for a character named Ji-hun. These were a little less consistent, but still usable. I’m not sure sure she suddenly has wavy hair when she’s frustrated, and I don’t know why the laughing image wasn’t a full body picture. These are good examples of why you probably don’t want to run too large of a batch at once, and sometimes you’ll have to manually adjust and re-run specific images.

What didn’t work?

I tried a few things that didn’t quite work out and that I’d do differently next time. Hopefully you can learn from my mistakes!

The standing pose is really similar in all of these results because I didn’t include any specific body changes in my prompt list. I think with a better list of prompts I could get more variety in the results. The facial expressions are good, but the poses don’t vary enough.

I didn’t realize that Ji-hun’s reference image wasn’t on a solid white background; it had a paper texture instead. The flaw in the character reference image meant I got more variety in the backgrounds later. That’s an easy fix if I remove the backgrounds, but it’s a good reminder to be more careful with the reference image.

With Maia’s character image set, I tried putting the reference image on a transparent background. I’d hoped it would generate all of the subsequent images also on a transparent background, but it it just put them on a solid black background instead. Again, an easy fix with removing the backgrounds, but I created some extra work for myself in needing a step to put them all on a white background for this blog post.

Some of these characters have shadows around their feet, but not all of them. That’s easy enough to fix in an image editing program though. I could probably add “no shadows” to the prompts as well.

Reflections

Overall, I think this is a really promising workflow for generating images for scenarios. I often generate a “shot list” for branching scenarios as a table (sometimes using Claude to generate a list based on my Twine export), so I already have a table describing scenes. This would work for generating images in context for scenarios all in a batch like this.

If I wanted to generate multiple sets of characters in standard poses with transparent backgrounds, this would be significantly faster than generating them one-by-one. Even if I just start with a standard set of poses and then prompt individually for a few specific images, this would speed up my process.

While I still like Midjourney for generating blog cover images and my initial reference images, I’m finding myself using Flora more often. It’s a definite shift in process, and I feel like I still need to spend more time to get really efficient at it. But I’m getting better the more I practice and refine my processes and prompts.

More on AI image generation

Learn more about generating character images using other AI tools.

Related