This Free Claude Skill Turns Images into 3D Models

Today I am going to show you how to turn any 2D image into a 3D asset using Claude. I gave Claude Code a single photo of a bike plus one public GitHub repo, and it rebuilt the whole thing as a working procedural Three.js model, created one hundred percent in code and editable on the spot. No 3D software, no paid generators, just your Claude subscription.

What you need

• Claude Code (this workflow runs there)
• The link to the img2threejs GitHub repo (it is public, so you do not even need a GitHub account)
• One or more photos of the object you want to rebuild

1. Any image into a 3D asset (0:00)

The repo we are using is called img2threejs. What it does is simple to describe: you take a picture, or multiple pictures, of a single item, you give it to AI, and the AI builds that item from zero in 3D. It will not give you a one hundred percent photorealistic copy of the 2D image, but we are getting very, very close to reusable 3D assets. And the more images you give it, the more precise it gets.

2. The showcase: what this skill can build (0:28)

The repo has a demo gallery you can browse before committing to anything. Every 3D model in there is built entirely using AI, with no pre-made assets whatsoever. What makes this special is that it does not use anything Claude cannot do by default. The repo simply contains a single skill that the AI uses to build those assets. It is the first iteration of this type of skill that I have seen, and I am confident that in the next three to six months we will see more and more capacity from these models to generate 3D assets.

3. Ask Claude what is inside the repo (1:40)

First things first: copy the repo link, go into Claude Code, and ask it to read the repository before doing anything.

Hey Claude, what's inside of this GitHub repo?

A repository is pretty much just a Google Drive in fancy terms, and this one contains everything you need. Give Claude a couple of seconds to read through it. It comes back describing the demo gallery and confirms that every model is built entirely with AI. While you are there, I highly recommend you give the repo a star, because this skill is incredibly useful.

4. Send the image and launch the skill (2:39)

Now the simple part. I copy my reference image (right click, copy image), paste it into Claude Code, and ask:

Please use this skill and also anything that you can find inside of that repo to recreate this image into a 3D asset.

What is fantastic is that you can combine image generation models with this skill. Say you are building a game and you want it heavily themed with a specific artistic direction: generate the 2D assets first, create multiple views of the same object, then hand it all to Claude and it will build those 3D sprites for you.

5. First check-in: decomposing the image into parts (3:45)

Generation is slow. Count between ten and twenty minutes per asset, so I cut the video and came back. Even before finishing, the skill had already produced a lot of elements from my single image. What it actually does is decompose the image into multiple elements and then extract each section, putting the 3D object together one piece at a time. It works with pure geometry: tubes, circles, ellipses, nothing else.

6. Detail passes: handlebars and wheels (4:51)

This is the part that makes the skill so insane. It analyzes each section of the bike it is building in 3D to find the imperfections. On the left of my screen, Claude is treating the handlebar. On the right, it is rendering the wheel to place all the necessary lines. A bicycle is a very complex 3D model, which is why my run went well past my estimate: I was nearly at the fifty-minute mark, using Opus 5 on high settings, which is likely part of why it took so long.

7. The final 3D model, side by side (5:35)

When it finishes, I have a working 3D model of the bicycle. The output includes a side-by-side comparison between the reference image and the model, plus an overlay of the two images showing how they compare, which is what Claude used to build it. Remember: this came from a single reference image. The skill accepts up to twelve, and we did not even use two. The level of similarity is already incredible.

8. Fully editable 3D, built entirely with code (6:17)

The model is created one hundred percent with code. That means I can change it on the spot using Claude: colors, materials, even the diameter of any part of the bike. And the quality goes up exponentially with the number of images you provide. Do not hesitate to give five, ten, or fifteen images from different viewpoints, because that is exactly what Claude is doing in the back: taking as many viewpoints as possible and building the model from them.

With the models we have now, Opus 5, Fable 5, GPT 5.6 Sol, hardware creation possibilities are getting close. Within the next year or so we will be able to create 3D models that pop into 3D printers very easily.

Try it yourself (7:31)

Grab any object you have laying around, take a few images, give your AI the repo link, tell it to read what is in there, and then use the skill on your images. You do not need anything beyond your Claude subscription, or your favorite AI subscription, to build these models from here on out. And leave the author a star, because what he built is absolutely unbelievable.

Pitfalls and tips

• This skill is slow by nature: plan for ten to twenty minutes per asset, and complex organic shapes can take far longer. My bike on Opus 5 high was still running at nearly the fifty-minute mark.
• Geometric objects like boxes, chests, and barrels finish much faster than curved, detailed ones.
• Do not expect photorealism from one photo. More viewpoints means exponentially better output.

Where to go next

• Curious about the model I ran this on? See my real-world tests in Claude Opus 5: Fable 5 Killer or Benchmark Hype?.
• Not sure where this kind of workflow should live? Read Claude Cowork vs Claude Code: What's the Difference?.