Generating 3D Models from Images with Blender MCP

After completing the Blender MCP setup in the previous article, I tried actually generating 3D models.

The tool this time is Hyper3D Rodin — a service that generates 3D models from text prompts or images, integrated with Blender MCP.

First: Verify the Connection

Before using Hyper3D, confirmed that Blender and Claude are properly connected.

Instructed it to “create a white silhouette.”

Works fine. Confirmed that Claude can operate Blender.

Side note: if you don’t delete the default cube that appears when Blender starts, it gets buried inside your generated model. Didn’t know that as a Blender beginner.

Text Prompt Generation

Next, used Hyper3D Rodin to generate a 3D model.

Told Claude “turn ‘anime girl with bikini’ into a 3D model,” and Hyper3D Rodin kicked off and generated one.

Generation result:

Item	Value
Vertices	22,890
Polygons	23,332
Material	model
Textures	diffuse, metallic-roughness, normal (512×512 each)

An anime-style character in a green bikini was generated in tens of seconds, textures included.

Image-to-3D: Battling Errors

Next, tried generating a 3D model from an existing image. That’s where the trouble started.

Wanted to turn this into a 3D model.

The reason for the swimsuit: figure detail in 3D is hard, so something close to the base body seemed like a better starting point. Gemini was finicky about it, so used SuperGrok instead.

The error

Specifying the image path and starting generation returned this:

Error: Input buffer contains unsupported image format
at Sharp.toBuffer (/app/node_modules/sharp/lib/output.js:163:17)
at TaskService.rodinPrompting (/app/dist/src/task/task.service.js:968:14)

The Sharp library on the Hyper3D server couldn’t recognize the image format.

Everything I tried (all failed)

Attempt	What	Result
1	Resize to 1024×1024	Aspect ratio broken (rejected)
2	Convert to JPG preserving aspect ratio	Same error
3	Convert to PNG	Same error
4	Convert to PNG with sips	Same error
5	Convert to PNG via Pillow RGB	Same error
6	Convert to JPEG via Pillow RGB	Same error

Not a format or size issue — something more fundamental.

Root cause analysis

The problem was happening when the MCP tool sent the file to the API. Traced through the source code.

MCP server (server.py) processing:

# server.py lines 800-802
images.append(
    (Path(path).suffix, base64.b64encode(f.read()).decode("ascii"))
)

Reads the image file and converts it to a Base64 string. Correct, since binary can’t go through JSON directly.

Blender addon (addon.py) processing:

# addon.py lines 1174-1175
files = [
    *[("images", (f"{i:04d}{img_suffix}", img)) for i, (img_suffix, img) in enumerate(images)],
]
# passed to requests.post's files parameter

This was the bug. The files parameter in requests.post expects raw binary data. But img is still a Base64 string.

What was happening:

1. MCP server converts the image to a Base64 string
2. Blender addon passes that string directly to requests.post’s files
3. requests sends the Base64 string as-is, expecting binary
4. Hyper3D’s Sharp tries to parse a Base64 string as an image and fails

With text-prompt-only requests, images is an empty list, so this bug never triggered.

Fix

Modified the create_rodin_job_main_site function in addon.py:

# Before (buggy)
files = [
    *[("images", (f"{i:04d}{img_suffix}", img)) for ...],
]

# After
files = [
    *[("images", (f"{i:04d}{img_suffix}", base64.b64decode(img))) for ...],
]

Added Base64 decoding so binary data gets sent instead.

Result after fix

Sent the source image (1536×2752px) without resizing, and 3D model generation succeeded.

Generated model:

Item	Value
Vertices	23,273
Polygons	23,332
Material	model
Textures	diffuse, metallic-roughness, normal (512×512 each)

Tip: Improve Accuracy with Multiple Images

Hyper3D Rodin accepts up to 5 images at once. With a single image, the back and sides are missing — hair and the body’s reverse side don’t reconstruct well.

Multiple angles improve accuracy (probably):

# Multi-image generation example
input_image_paths = [
    "/path/to/front.png",   # front
    "/path/to/back.png",    # back
    "/path/to/side.png"     # side
]

Tips:

• Use the same character and outfit across all images
• T-pose is ideal if possible
• Simple backgrounds work better
• The first image becomes the base for texture generation

---

Blender MCP + Hyper3D Rodin can generate 3D models from text or images. Hit an error with image-based generation, but tracking it down to a missing Base64 decode in the MCP tool was the kind of thing that’s only possible because it’s open source.

Next up: trying multiple images to get higher-quality model generation.

Related Files

Files that needed modification (macOS):

• MCP server: ~/.cache/uv/archive-v0/.../lib/python3.10/site-packages/blender_mcp/server.py
• Blender addon: ~/Library/Application Support/Blender/5.0/scripts/addons/blender_mcp.py

References

• Previous: Blender MCP Setup
• BlenderMCP GitHub
• Hyper3D Rodin