How to Make a VTuber 3D Avatar: A Beginner’s Guide to Blender + GPT-6 Astra

Creating your own 3D VTuber model involves much more than simply turning a 2D character into a 3D one. From modeling the head and body to rigging clothes, setting up materials, creating facial expressions with Shape Keys, adding hair movement, and applying toon shading, every step affects how natural the final character looks.

Besides knowing tools such as Blender and Unity, you also need some art and 3D knowledge to judge whether facial proportions, clothing shapes, mesh deformation, and movement look right.

This is also why many people choose to hire professional 3D modelers. If a model is built from scratch and includes modeling, texturing, rigging, expressions, and commercial-use requirements, public commission prices can easily reach thousands of dollars. The more complex the character design, expressions, and clothing requirements are, the higher the production cost becomes.

However, if you simply want to try making your first 3D VTuber yourself, there are now many easier ways to get started.

Many beginners start with VRoid Studio or Ready Player Me, using preset parts, adjustable parameters, and custom textures to create a character before exporting it as VRM. This approach has a lower learning curve, but unusual body types, complex outfits, or highly customized designs usually still require further editing in software such as Blender.

Another option is to use generative 3D tools such as Tripo or Meshy to turn character images into 3D models, then continue with rigging and motion setup. Tools such as Mixamo can help automatically rig some humanoid models and apply animations, but facial expressions, clothing deformation, and character details still require additional work.

Now, with GPT-6 Astra’s Computer Use capabilities, there is another way to create 3D characters. Astra can directly help operate Blender instead of only providing written instructions.

This guide is based on the Tripo GPT-6 Astra 3D Character Workflow, together with practical tips shared by creator Nano (@Dstudio_ai), covering the full beginner process from character images and 3D model generation to Blender assembly, rigging, and facial expressions.

TripoとGPT-6 Astraを使ってキャラクターモデルを作ってみたのでメモ！【PR：Tripo】

超重要なとこ！！！！！
・Blender MCPをなるべく使わせない。9割9分Computer useでやらせる。←マジのガチで一番大事… pic.twitter.com/kHLd6GIop2

— Nano(ナノ) (@Dstudio_ai) September 12, 2026

When using AI tools, make sure that any materials you provide comply with copyright rules and the relevant terms of use. Do not use copyrighted materials for commercial purposes, modify them, or create derivative works without permission from the rights holder.

What You Need Before Making a 3D Model

The roles in this workflow can be understood simply as follows:

• Character reference images: Provide references for the design, proportions, clothing, and colors.
• Tripo Studio: Generates 3D assets such as the body, head, and hair.
• Blender: Handles model assembly, rigging, expressions, materials, and animation.
• GPT-6 Astra: Helps inspect the screen and operate Blender.
• Unity + UniVRM: One possible workflow for preparing and exporting the model as VRM.

Tripo’s official tutorial uses Blender MCP to connect the AI assistant with Blender, while Nano uses a different approach. Nano strongly recommends minimizing the use of Blender MCP and mainly using Computer Use to operate Blender, because during testing, MCP repeatedly struggled with tasks such as selecting specific faces and identifying local parts of the model.

In other words, this workflow is not about handing the entire job to AI and walking away. Astra performs the operations while you check the results and continue giving directions for corrections.

Step 1: Separate the Character Image into the Head, Hair, and Body

Do not immediately upload the complete character image into a 3D generation tool. In Tripo’s workflow, the character is separated into three main parts:

1. The clothed body
2. The head without hair
3. The hair and head accessories

Keep the complete character image as a shared reference for proportions, clothing, and colors. A front-facing full-body image is best, with the hands and feet fully visible and the arms kept away from the torso. An A-pose or T-pose is easier to process than a complex pose, and Tripo’s official example uses a T-pose.

Separating the image is not simply a matter of cutting it into three pieces. For example, after removing the hair, you need to restore the scalp that was originally hidden. When preparing the body image, you should also leave enough of the neck visible, otherwise connecting the head and body later may become more difficult.

Based on Nano’s testing, the head and body may still produce usable results with only a front view. However, it is better to prepare multiple views for the hair, especially side and back views, because the depth and overall hairstyle shape can easily become distorted.

After preparing the images, manually check whether the hair length, accessory positions, and proportions are consistent across different views. Avoid giving the generation model conflicting references.

Step 2: Generate the 3D Parts Separately with Tripo

Next, generate the body, head, and hair separately in Tripo. Nano used Smart Mesh P2.0 and recommends starting with around 2,000 to 5,000 polygons for the head, making it easier for Astra to handle and edit later.

This is an experience-based range from this particular case, not a fixed polygon count that every VTuber model must follow. Rather than simply aiming for more polygons, it is more important to check whether the topology and facial structure are suitable for later editing.

For the head, pay particular attention to:

• Whether the eye sockets have a reasonable amount of depth instead of being completely flat.
• Whether the eyes, pupils, or related areas can be handled separately.
• Whether the mouth has enough 3D structure.
• Whether enough of the neck is preserved for connecting it to the body.

In particular, if the eyes are simply painted onto the face texture, you cannot assume that the model already supports eye rotation, blinking, or gaze tracking. If the generated structure is clearly unsuitable for later production, regenerating it may be better than forcing a repair.

After selecting the models, export the Body, Head, and Hair separately in a Blender-compatible format such as GLB, and name them clearly.

Step 3: Assemble the Character in Blender

After importing the Body, Head, and Hair into Blender, use the body as the scale reference and adjust the position, size, and rotation of the head and hair. At this stage, focus only on whether the character looks right. Do not rush to complete rigging, expressions, materials, and animation at the same time.

Check the model from at least the front, side, and back:

• Are the head proportions correct?
• Is the hair floating above the scalp?
• Are there broken or missing faces on the back of the head?
• Does the head connect naturally to the neck?
• Does the hairstyle silhouette match the original design?

For this kind of quick prototype, the head and neck do not necessarily need to be immediately joined into a single mesh. If the connection is hidden well, it may not look obviously wrong in a still image.

However, if the character later needs to turn the head significantly, look up, or look down, simply placing the two models together may reveal gaps. For a finished model, the mesh connection, weights, or other deformation methods should therefore be adjusted as needed.

Step 4: Set Up Expressions, Starting with “Model Switching”

In general 3D character production, Shape Keys are often used to create facial expressions. A Shape Key can store different shapes of the same mesh and interpolate between the Basis and target shapes, making it suitable for continuous changes such as blinking, smiling, and mouth shapes.

However, Nano did not ask Astra to create a complete Shape Key system in this project. Instead, different facial-expression versions of the head were first generated in Tripo, then the visibility of different Head Meshes was switched in Blender, such as neutral, smiling, and eyes closed.

This is a simpler form of “discrete expression switching.” Each head version should keep the face shape, proportions, and neck position as consistent as possible. Otherwise, switching expressions may make it look as if the entire face has suddenly changed.

This method works well for displaying a prototype, but switching between three head models is not the same as having a complete VTuber facial-tracking system.

If you want the eyes to gradually move from open to half-closed, the mouth to change continuously with speech, or the model to support facial tracking, you need suitable Morph Targets / Blend Shapes, bones, or other facial structures that can deform continuously.

If the final goal is to create a VRM model, blinking, mouth shapes, and facial expressions usually also need to be organized into the VRM Expression system. VRM itself defines controls for blinking, mouth shapes, and some standard expressions.

Step 5: Complete Rigging and Weights, Then Test Character Movement

Rigging creates the character’s skeleton and control structure so the model can be posed and animated. Weights determine which bones affect each vertex and how strongly they affect it. Blender’s Armature Modifier deforms the model based on the skeleton and the weights stored in Vertex Groups.

There is one important detail in Tripo’s official workflow: if the body has already been rigged in Tripo, it is recommended to reuse the Armature and Weights from the imported GLB instead of rebuilding the entire skeleton.

Then align the head and hair with the same skeleton. Only create a new Humanoid Rig in Blender if the original model does not have a usable rig.

The clothing also needs to deform properly with the body. If the clothing and body are already part of the same Mesh, their weights can be handled together. If the clothing is a separate Mesh, it also needs to be correctly bound to the same Armature or use another suitable method to follow the body.

Afterward, do not only check the standing pose. At minimum, test:

• Raising the arms
• Bending the elbows
• Turning the head left and right
• Bending the knees
• Leaning the body forward

Pay particular attention to whether the shoulders collapse, the elbows twist, the neck splits apart, or sleeves, skirts, and accessories clip through the body.

When you find a problem, it is better to specify the exact body part and pose. For example: “When the right arm is raised to shoulder height, the right shoulder collapses. Please adjust only the Weight around the Shoulder and Upper Arm.”

This makes it easier to get a controlled result than simply asking AI to “fix the Rigging.”

Nano believes Astra performs particularly well in Rigging and secondary-motion setup. However, Nano also noted that this workflow used a self-developed Rigging helper tool, so the result cannot be considered something Astra completed entirely on its own.

Step 6: Adjust Hair Movement, Materials, and Toon Shading

Once the character’s basic movement works correctly, work on the hair, skirt, earrings, and other Secondary Motion. This makes it easier to tell whether a problem comes from the Rigging or from the physics setup.

If the goal is to use the model for actual livestreaming, you will also need a secondary-motion system suitable for real-time use.

Materials and textures should also be checked:

• Do the face and neck have matching skin tones?
• Are the eyebrows, eyes, and mouth clear?
• Are the clothing patterns distorted?
• Do the hair and body use a consistent material style?
• Are Normal, Roughness, and other maps connected to the correct Shader?

To achieve the toon-style look commonly seen in Japanese VTubers, you also need to adjust shadow boundaries, lighting direction, and outlines.

Simply increasing the polygon count does not necessarily make the character look more anime-like. If you plan to export the model as VRM, remember that the Shader and Physics you see in Blender may not behave exactly the same after export.

For example, VRM 1.0 supports material systems such as PBR, Unlit, and MToon, and also defines Spring Bone as the standard mechanism for moving parts such as hair and skirts. MToon and Spring Bone each have their own VRM 1.0 Extensions.

Step 7: Check the Model with a Short Animation

After completing the basic setup, do not rush to apply a full dance animation.

A short animation of around three seconds, such as waving, turning the head, or making a simple weight shift, is useful for checking Rigging and mesh-clipping problems.

For example, you can test: “Create a three-second waving animation. Keep both feet in place, raise the right hand and make a small waving motion, slightly turn the head with the movement, then return to a natural standing pose.”

When you find a problem, clearly describe the timing, body part, and issue.

For example: “At Frame 20, the left hand passes through the jacket. Please reduce the arm movement while keeping the rest of the animation unchanged.”

This makes it easier for AI to identify what actually needs to be fixed than simply saying, “The hand looks strange.”

Tripo also recommends starting with short movements, checking the hands, feet, clothing, and character silhouette first, and then gradually moving on to more complex animation.

Finally, keep the Blender project file and a full-body motion preview.

At this point, you have completed a 3D character prototype that can be properly inspected. You can then decide whether you want to continue working on Shape Keys, VRM Expressions, Spring Bone, Look At, and compatibility with VTuber streaming software.

3D VTuber Model FAQ

Can I Start Even If I Know Nothing About Blender?

This method does reduce the amount of manual Blender work required, but you still need to judge whether the character proportions, broken faces, seams, clipping, and weights are correct.

AI can help operate the software, but human judgment is still important for deciding whether the result is actually correct.

Can I Use Only GPT-6 Astra and Skip Tripo?

It is strongly recommended to generate the model with Tripo first and then use Astra for the setup, rather than asking Astra to model everything from scratch, as Astra’s modeling capabilities in this area are not yet mature.

Should I Use Blender MCP or Computer Use?

Tripo’s official tutorial uses Blender MCP, while Nano’s testing recommends using Computer Use for most tasks, especially when you need to directly inspect the screen, select specific parts of a Mesh, and confirm the results.

The two tools are not mutually exclusive, but based on this case, it is not recommended to hand every operation over to MCP.

Also keep in mind that regenerating models in Tripo or using Astra’s Computer Use for long periods may consume a significant amount of plan usage and generation Credits.

Can AI Replace Professional 3D Modelers?

AI is beginning to lower the technical barrier to 3D character prototyping, allowing people without complete modeling experience to create character prototypes that can move and be tested.

However, accurately recreating a character design, maintaining clean topology, creating natural facial expressions, achieving stable weights and clothing deformation, supporting livestreaming software, and continuing to add outfits and features all involve many details.

Working with an expert can save you a great deal of trial and error.