Recently I've been messing around with a number of different tools to help improve reference material that I have for some of my fun personal projects. I've always enjoyed finding good reference for things and luckily with Star Wars stuff I've collected a lot of material over the years including photos people have taken at exhibits of props, costumes, and miniatures over the years along with all kinds of different official imagery and anything that can be useful for trying to recreate stuff from the movies.
When playing around with AI there isn't really anything useful that can generate 3D content, I've tested out some different AI tools and it can only do some of the most basic stuff and the results are just not that great or usable. However, I've found some ways to generate some useful images for reference when doing normal 3D asset modeling work. This I think is a way to use AI that's not replacing any part of the process, it's providing you with something that isn't otherwise possible. I'm not super experienced with character modeling so I've been trying to gather as much useful reference of Star Wars characters that's available out there and with some of the characters if they are popular enough (which would mean that AI has more training material from them) then I've found it's possible to generate good character reference to use for 3D modeling. In this example I'm trying to get reference for Obi-Wan as he is seen in Attack of the Clones. I was able to provide one promo photo and then with the prompt it was able to generate an orthographic character sheet with him in the A-pose that will help to model him.
Character sheet of Obi-Wan Kenobi from Star Wars Attack of the Clones, based on the costume in the reference image. Front view and side view and back view, all views orthographic, all views are in an A-pose, neutral lighting, flat white background, hyper-realistic, 8k, photorealistic render style. He wears the textured beige outer tunic, light-yellow inner tunic, dark brown leather utility belt with pouches, and tall reddish-brown leather boots. The images are tightly packed to fill the entire canvas width and height as much as possible while maintaining a tiny bit of spacing and consistent scale between each other. --ar 3:2 Let's call this image "Image_001" for later reference
My source image I first upscaled with Topaz Gigapixel and then I used Gemini to help develop the prompt because without some of the specific direction for his AOTC look it would default to his look in ROTS due to more training on that version. I was also able to get it to render the gray clay version for any of the generated images using this prompt:
Do the same image, but instead of photorealistic style render it in a uniform medium matte gray clay material with very strong ambient occlusion (AO)
And by assigning a name to the full character sheet, I'm able to easily reference back to it to generate different images like a closeup of his face.
While this generates something that looks like a complete 3D model, this is just an image and like I mentioned before there aren't any AI tools that can generate any 3D models even remotely close to this. When closely analyzing the results you'll find small variations between each view along with image artifacts so while these images are very useful for 3D modeling reference they can't be used for much else.
For another example I wanted some reference of Zam Wesell but she's not well known enough for the AI to already know enough about her to be able to produce her from just a prompt and since I was able to get good orthographic A-Pose reference of Obi-Wan, my goal was to try to achieve that with Zam as well.
Some of the best reference available is in books like Star Wars Chronicles which has multiple angles of pretty much every prop, character, and vehicle that you can think of. I've gotten cheap used copies of these reference books so that I can take them apart to more easily scan the photos without having to deal with trying to flatten the binding of the book against the scanner. When trying to enhance the scans and clean up the printing roughness Gigapixel wasn't able to deal with the printing artifacts very well and it would sometimes try to enhance those as details instead of ignoring them as flaws, but with Gemini it's able to remove the noise and other printing artifacts and also remove any leader lines and captions.
Analyze the provided scanned image. First, remove all the halftone printing dot patterns and other noise artifacts from the image to create a clean, smooth surface. Remove all text paragraphs, captions, titles, and also remove leader lines that point from the captions to areas of the photographic parts of the scanned image. Then, upscale and enhance the image, sharpening all the details, improving clarity, increasing the brightness and making it look like a high-resolution, professional-quality photograph.
For Zam, the next step was to try and get it to make my orthographic character sheet in the A-pose. While the source photos are useful, having a sheet where everything in each view is lined up and there's minimal perspective distortion can help with modeling. After several tries generating the image it would just reproduce the source image but with different lighting, but after around 10 tries it was able to generate what I was looking for.
Reference the costume details, colors, and materials in the image above precisely. Create a technical 3D character modeling reference sheet for this character (Zam Wesell).
CRITICAL POSE OVERRIDE: Ignore the relaxed poses in the reference image. The character must be rendered in a rigid, symmetrical A-Pose (arms out at 45 degrees, legs straight) for all views.
VIEWS & STYLE: Three separate views arranged horizontally: Front A-Pose, Side Profile A-Pose, Back A-Pose. The style must be a flat Orthographic Projection. No perspective distortion, no dramatic shadows. Photorealistic, 8K style with neutral, flat studio lighting against a solid medium grey background.
COSTUME DETAILS: Ensure all details from the reference are translated to the A-pose: the silver helmet with the red scarf draped on the right side, the purple ribbed bodysuit, the chest armor with tubing, the utility belt with holsters, and the distinctive square-tiled leather skirt.
Let's call this generated image "Image_001" for later reference.
Negative Prompt: Perspective, foreshortening, cinematic lighting, dramatic shadows, relaxed pose, asymmetry, photorealistic photography, dynamic pose, shapeshifting.
Again, I used Gemini to help develop the prompt, the more descriptive and strict the better. Another thing to keep in mind is that the overall size of the image that it generates has a max of 2048x2048 pixels, so you'll want to try to have minimal amount of empty space as you can so layout is very important. Getting the full character sheet is the first step and having a name for that image allows you to create isolated renders of each view so that you can get higher quality results. Here's an example of my next prompt to get just an image of the front view:
I need to generate a high-resolution isolate of ONLY the Front View from "Image_001".
CRITICAL INSTRUCTIONS:
ASPECT RATIO CHANGE: Do not use the wide aspect ratio of the source image. You must re-render the canvas into a TALL, VERTICAL PORTRAIT orientation (approx 2:3 or 9:16 ratio) suited for a single standing figure.
FRAME MAXIMIZATION: In this new vertical frame, zoom in closely on the Front View character. The subject should fill the frame almost entirely from top to bottom to maximize resolution. Eliminate excessive empty space above the head and below the feet.
STRICT CONSISTENCY: The generated image must be a perfect match to the design, pose, armor details, and coloring of the Front View in "Image_001". Do not alter the design; just re-frame it into the new vertical canvas.
ISOLATION: Do not include any parts of the other views.
The details ended up getting lost from all of this generation like the writing on her armor, but those details can easily be corrected when modeling and I have other photo references for those. What's important here is the proportions that can help a lot before getting to any of the details.
I was also able to use this to get some clearer reference of her pistol.
If you have any questions feel free to ask. Right now I've been doing this all in Gemini, Midjourney has worked well for some things and the benefit there is being able to generate multiple variations at once but Gemini has gotten better over time.