How a Blender beginner used Claude Code to connect 6 tools and create a VRChat avatar in one day

This is a two-day record of how I, a Blender beginner, decided to create a VRChat avatar from my own photo.

The trigger was a ChatGPT image prompt that recently went locally viral on X. It's a tool that lets you create your own character from a single photo and a prompt.

When I imported my photo, a super cute character was created. GPT-image-2 is truly amazing.

Wait, I wonder if I can turn this into an avatar. I realized it could probably create a four-view drawing easily too.

Create a four-view drawing of the avatar with GPT-image-2 and send it to TRIPO 3D

As of May 2026, tools for creating 3D models from images have become quite robust, with affordable and easy-to-use options like Meshy and TRIPO 3D. If I upload the four-view drawing to TRIPO 3D, could I use it for a VRChat avatar?

Import the generated four-view drawing into TRIPO 3D → Generate a 3D model. In a few minutes, the avatar model was finished.

TRIPO 3D → Blender was the first hurdle

The next hurdle was importing the model created in TRIPO 3D into Blender. I didn't know how to do it.

When I asked Claude Code by connecting it to the official Blender MCP, it told me there is a feature called "Tripo DCC Bridge for Blender" that transfers data directly to Blender. I connected the bridge MCP to Blender and succeeded in importing it with a single button press.

But it's rough. The mesh around the face is jagged, and the texture is coarse.

I decided to turn it into an avatar as is, thinking that getting it moving was the priority.

Adding a rig with Adobe Mixamo

I asked Claude Code for the procedure.

• Export as FBX

• Upload to Adobe Mixamo

• Manually place the spheres for joints on the corresponding parts of the avatar

• Bones are automatically inserted and bound to the mesh

• Download the rigged FBX

Mixamo is free. Since it's a web app, no software installation is required, and the operation is simple. However, because it's a free tool, the precision of the movement when brought into VRChat is low.

Just by manually placing the spheres for joints on the corresponding parts of the avatar and clicking the button, a rigged model is completed. Download that as 'FBX for Unity'.

To Unity via Blender

Bring the FBX downloaded from Mixamo into Blender.
I had Claude Code handle this import process.

Since the armature was already set up, Code told me I could take it to Unity as is. I checked the movement in Blender just in case before going to Unity. (I don't know how to operate Blender, so I left it to Code)

I didn't know how to apply materials either, so when I consulted Code, it created a tool that automatically sets up materials and shaders. I just dropped the tool and the model into Unity as instructed, and the setup was completed with a single button.

After that, as guided by Code, I added the VRC Avatar Descriptor and performed the initial settings. After repeating a few errors, the build to VRChat was completed.

(By the way, building from Unity to VRChat might be difficult without basic knowledge of VRChat avatar uploading and modification.)

I made my first original avatar ever.

Starting to worry about the roughness of the face

The build was successful. But the face mesh is way too rough.

This is because I'm cramming both the face and body into a limited polygon count. I came up with the idea of separating the mesh for just the face.

I created a three-view drawing of just the face with a bald head and sent it to TRIPO 3D. (Assuming I'll put on hair from an asset I bought on Booth later)

I created the head and body separately and combined them in Blender. As expected, generating just the head significantly improved the precision of both the mesh and the texture. I'm happy.

I cut the unnecessary mesh around the neck (I had Claude Code do this). I manually adjusted the size and position in detail to make the neck seam as seamless as possible.

After integrating the head and body, I added a rig in Adobe Mixamo and moved to Unity.

Something I noticed here. Since it was generated from a four-view drawing, the front and side views look well-formed at a glance, but the proportions look wrong when viewed from an angle. From an angle, the cheeks lack volume and the chin is too sharp, making the head look like a triangular pyramid.

Even so, I finished it for now by adding hair and glasses from assets I bought on Booth, and built it for VRChat.

Since TRIPO 3D is an American service, I concluded that it's probably difficult to create models with the anime-style aesthetic popular in Japan from images. I switched to a method of creating the head using only prompts.

Trying to generate a head model using only prompts

When I have TRIPO 3D generate just the head, for some reason it turns into a heavy, Western-looking realistic face.

I had ChatGPT analyze the style of the original illustration and kept repeating the prompt gacha. I managed to get an Asian-faced torso that I thought was cute, and imported it into Blender.

I adjusted the head size, lengthened the lower legs, and adjusted the proportions to my liking. I've gone through so much trial and error to get to this point that I've learned how to move mesh vertices to smooth out the seam between the head and body.

The mysterious phenomenon where materials broke was a real headache.

I don't know the reason, but when I took the body merged in Blender to Mixamo and brought the rigged version back into Blender, the materials applied to the head became a mess.

I got stuck here for 5 hours. As the context grew longer, Claude Code became increasingly stupid. Even the supposedly strongest Code is quite useless when it comes to Blender.

While getting angry repeatedly (I was able to vent without reservation since I separated my personality from Chat/Cowork Maro...), I started new sessions several times and somehow managed to get it looking decent.

I still don't really understand why this happened or why it was fixed (lol).

When I imported it into Unity, the material application was still slightly off, but I decided to proceed as is for today.

Completion, and then greed sets in.

I modified the rigged model in Unity and took it to VRChat.

Oh... this time, it looks beautiful from every angle.

A friend happened to be in VRChat, so I moved there and we took a commemorative photo.

Once it gets to this point, I start wanting more. Next, I want to touch the textures myself.

However, unfortunately, the UV map is too messy to do anything.

A modeler friend told me, 'Why don't you just paint directly on the 3D model in Adobe Substance 3D Painter?' I thought, 'That's a great idea,' but when I looked into getting it, the price made me hesitate.

When I asked Claude Code, it told me there are tools in Blender for painting colors directly onto 3D models. It set up the screen for me, so I started painting away. How convenient...

However, there's a limit to how well you can paint with a mouse, so I want to get a pen tablet before I continue. It would be great if I could do it with an iPad + Apple Pencil.

As of now, it's nearly impossible for someone without technical skills to create a body from scratch that allows for costume modification with clean rigging. Next time, I plan to buy a generic VRChat body on Booth and attach the head I made to it. More on that later!

What an amazing era we live in.

Up to this point, I have only spent a total of 3 days actually touching Blender, including the time I made a lily flower from TRIPO 3D previously. As for avatar generation, I was able to build it for VRChat on the very first day.

Although there are still mountains of challenges, considering the evolution of generative AI since I started creating images with Midjourney 3 years ago, I thought that in about 2 more years, we might be able to create decent avatars with beautiful rigging, UV maps, and facial shape keys using AI.

Summary of the entire process

1. Create face and body materials

• Generate characters from your own photos, illustrations, or prompts using ChatGPT's Image 2

• Generate a four-sided view (front, right, left, back) of that character

2. 3D modeling (TRIPO)

• Upload the four-sided view or prompt to TRIPO (※For anime-style avatar heads, generating directly from prompts rather than images results in a more beautiful, unbroken shape)

• A T-pose 3D model is generated in a few minutes

3. Import into Blender, merge, and shape

• Connect TRIPO 3D × Blender MCP bridge

• Import TRIPO model into Blender and adjust

• If the face and body were generated separately, join them in Blender (*at that time, perform shaping and fine-tuning such as smoothing the seams)

4. Add rig (Mixamo)

• Export from Blender to FBX

• Upload to Mixamo

• Manually place joint spheres → automatic binding with bones complete

• Download rigged FBX for Unity

5. Check in Blender

• Check if the rig is working properly

• Export as FBX with armature included

6. Setup in Unity

• Import into Unity

• Apply materials/shaders using the automatic material setting tool created by Claude Code

• Add Booth assets such as hair and glasses

• Add VRC Avatar Descriptor component, perform initial settings, and build to VRChat

7. VRChat

• Verify operation

Tools used

• Claude Code: Consultant for all processes, Blender operations, Blender MCP bridge, and generation of Unity setting tools

• ChatGPT (Image 2): Character image generation from photos, illustrations, prompts, etc.

• TRIPO 3D: Four-view images or prompts → 3D model conversion

• Blender: Importing, merging, mesh correction, texture painting

• Adobe Mixamo: Automatic rigging, skin binding

• Unity: Material/shader settings, VRC Avatar Descriptor

• VRChat: Implementation

Points where I got stuck and workarounds

1. When generating from four-view images, the drawing breaks down when viewed from an angle

• Cause: Since four-view images lack diagonal information, TRIPO 3D fills in the gaps using only front, side, and back information

• Workaround: Generating the head directly using only prompts makes it easier to get a well-proportioned 3D shape

2. When generating using only prompts, TRIPO 3D tends to produce realistic Western-style faces

• Cause: Presumed to be due to training data being centered on Western models

• Workaround: Have ChatGPT analyze the style of the reference illustration and repeat the prompt gacha

3. Returning to Blender via Mixamo can sometimes break the head material

• Cause: Unknown (I spent 5 hours stuck on this, but the reason for the resolution remains unknown)

• Workaround: Start a new session to reset Claude Code and try again through trial and error. It will eventually work if you re-assign it on the Unity side

4. Claude Code's Blender operations become sloppy as the context grows long

• Cause: Model followability drops as the conversation gets longer

• Workaround: Switch sessions for each work block. Creating skills (saving procedures) is also effective

5. UV maps derived from generative AI are too messy for texture editing

• Cause: Automatically generated UV unwrapping has not undergone manual optimization

• Workaround: Rather than editing UVs, it is more practical to paint directly onto the 3D model in Blender (an alternative to Substance 3D Painter)