Figuring out more about figures

I’ve been happy enough with the cannery worker figures that I’ve continued to experiment with making figures over the last few weeks. Here’s a couple interesting observations.

Let the AI propose figures

The current workflow - I think of a description of a figure, let gemini.google.com draw a picture of a person based on that description, and I let tripo3d.ai turn the image into a 3d printer model - is pretty easy. Choosing which figures to make requires a little bit of care. What are suitable figures for my era and place? How do I describe someone in 1930’s clothing? How do I mix up my choices do I don’t have a bunch of similar figures?

As I was doing lots with the AI / large language models to create images, I realized that the generative AI systems are also good at analyzing photos too. So… we could also get the AI chatbot to propose a bunch of possible figures, and I could pick out the ones I want.

The AI chatbots work completely from text, so all I need is the right prompt. I ended up taking one of the photos of the workers at a cannery (such as the panoramic photo from California Canneries in Campbell. I dragged the photo into the chatbot at gemini.google.com, and gave it the following command:

Find the human figures in the attached photo from 1930’s California. For each person, write a four sentence physical description of the person that can serve as a prompt to generate a 3d figure of the person. Each description should start with "Create a front, side, and rear image of a human figure suitable for creating a 3d model of the person where the figure represents a person in 1930’s California. The figure is…" For each figure, the description should indicate details of body and face type: build, height, facial features, etc. inspired by the original photo The four-sentence description should also include a description of movement that would be appropriate for that person: either that the figure is walking (most figures should show some movement), perhaps carrying an item (personal item, fruit lug, tool, bag, appearing in a quarter of the figures), or indicate a reasonable pose for a normal person of the time. The description can also contain facial features or other details that might make the figure unique or animated. The output should be a sequence of description prompts that can be used later to create an image of the figure. Each description should contain a summary or label for the figure (“the lean plant mechanic”), followed by the four-sentence description in plaintext. Do not generate the images at this time.

Note that I started with a smaller prompt, but added detail and restrictions when the AI didn’t quite do what I wanted. For example, the gory detail about “show the output this way” happened because the chatbot would switch between giving textual output and just giving me a picture in a form I didn’t want. The comment about adding movement and animation was needed because otherwise the figures were all just standing with arms at their side.

Running that command on gemini.google.com quickly throws out three or four potential figure descriptions. Pasting those descriptions/prompts into Gemini generates nice images, sometimes with details or roles I hadn’t intended. I was particularly surprised when the chatbot started identifying potential roles.

Here’s a on example Gemini created:

The industrious cannery worker: Create a front, side, and rear image of a human figure suitable for creating a 3d model of the person where the figure represents a person in 1930’s California. The figure is a sturdy, broad-shouldered woman with a kind, experienced face and graying hair pulled back under a simple white head covering. She wears a long-sleeved dress covered by a practical, thick canvas apron, and she is holding a small metal fruit-sorting tool in one hand. The figure is posed standing upright with a slight forward lean, as if preparing to return to an assembly line.

and here's the image that Gemini created for that prompt. Don't ask me what the tool she's holding is.

Now, I don’t want to have figures done automatically: I still want to be choosing what kinds of figures match my era and setting, and what figures help tell the story. But automatically generating these is a neat way to get inspirations about what figures to do next, and give me options I might not have considered.

The most annoying part of AI is how the work doesn’t fit what usually keeps me focused. Each request to generate an image or generate a 3d model can take from 30 seconds to a few minutes. That’s annoying because it involves a bunch of waiting and often bumps me out of a flow state where I’m focused on work. The pauses encourage me to go off and do another task, and then by the time I come back, I’ve forgotten what stage I’m at with each image. Worse, the AI tools often do something screwy - producing images without a background, or disregarding part of the prompt, or generating an incorrect figure. Using the generating AI systems often feels like trying to get a teenager to help - sometimes it does really well, sometimes it misunderstands completely, and some times it does something stupid because it just wants to go play outside. A figure might be shown walking, but in different poses in the front and side images. Another figure may have odd details added, or might appear period-inappropriate.

Dealing with the 3d printing process ends up being in some ways easier. Taking each 3d model, adding support structures, and arranging multiple figures on the build platform is tedious work, but at least it’s constant work that keeps me engaged.

Painting

I’ve usually found painting figures taxing because it takes so long, because I’m not happy with my color choices, or because of delays waiting for paint to dry when painting a few figures.

My first big improvement was painting bigger groups of figures. I’d attach 10-20 figures to a strip of styrene, prime all the figures, then go down the line painting each color - flesh color first, then clothing, then hair (once faces had dried), then shoes, then hat.

For color choices, I started out this time writing down potential color choices for clothing, then made sure I had those colors on my pallette. Having the color ranges figured out meant that I had all my colors ready, and had ideas about variations to try ahead of time. I did one table of colors for the regular worker figures, and another for some of the unusual figures (such as management).

white, yellow

Clothing	Men workers	Women workers
shoes	brown, black	brown, black, tan
pants/dresses	tan, brown, black, blue, olive	yellow, white, light blue, dark blue, tan
shirts/aprons	white, blue, red check, tan
hat	brown, black, tan	white
hair	brown, black, yellow, white	brown, black, yellow, white
shoes	brown, black	brown, black, tan

All this sped up figure painting so I could do a large set in an evening.

Next steps

Once we’re doing figures of random folks, the next question is obviously “can we do figures of people we know?” I’ve also been doing a few figures based on friends. I’ll share some of those results another time.