🎮 Turn a single image into a textured 3D game asset using a local AI model, directly from Claude Code, Cursor or your terminal.

Meet MCPBytes Image-to-3D, a free, open-source CLI and MCP server built around Tencent's Hunyuan3D 2.1.

It's not another newly trained 3D model. Instead, it packages an existing generative model into a complete workflow that AI coding agents can actually use.

What does it do?

Give it a PNG, JPEG or WebP image of an object.

The tool automatically:

▪️ Removes the background: Isolates the subject using U²-Net, or preserves an existing transparency mask.

▪️ Reconstructs geometry: Uses Hunyuan3D 2.1 to generate a 3D mesh from the reference.

▪️ Generates PBR materials: Creates base-color and metallic-roughness textures.

▪️ Cleans the mesh: Welds vertices, removes disconnected debris and normalizes the object to unit size.

▪️ Positions the asset: Places the model on the ground plane with +Y as the up axis.

▪️ Optimizes for the web: Produces a separate GLB with Draco-compressed geometry and WebP textures.

▪️ Creates previews: Automatically renders the model from six directions, including front, back, sides, top and bottom.

The output is more than a single 3D file.

You get:

▪️ model.glb: Full geometry with 1024px PBR textures.
▪️ model.web.glb: Compressed version for browsers and engines.
▪️ model_preview.png: Six-view preview sheet.
▪️ Individual preview images and camera metadata.
▪️ Generation settings, timing, checksums and provenance records.
That last part is particularly useful for automated asset pipelines.
And it works through MCP.
Connect the server to Claude Code, Cursor or another compatible AI client, then ask it to generate a 3D object from a reference image.
The agent can call create_3d_model_from_image, inspect the generated previews and use the resulting files in a larger workflow.

No manual GUI interaction required.

Performance and hardware:

The developers measured two modes on an NVIDIA A40:

▪️ Reduced: 40K faces, 512px painting views, 28.2GB peak VRAM, 203 seconds.

▪️ Full: 120K faces, 768px painting views, 41.2GB peak VRAM, 386 seconds.

Both measurements include approximately two minutes of model loading.

Requirements:

▪️ Linux or Windows via WSL2.
▪️ NVIDIA GPU with at least 32GB VRAM.
▪️ CUDA 12.8-compatible driver.
▪️ Around 19GB for model weights, plus 16GB for Python dependencies.

Install with one command:

uvx --python 3.12 --torch-backend=auto mcpbytes-image-to-3d chair.jpg

No powerful GPU? There's a cloud option.

Add --cloud to send the job to MCPBytes servers.
▪️ 300 free credits at signup.
▪️ Economy generation: 30 credits.
▪️ High quality: 100 credits.
▪️ Max quality: 150 credits.
Cloud processing requires an API key. Local mode doesn't upload your images.
Another interesting detail: reproducibility.

The developers report that the same reference image, random seed, settings and GPU produce byte-for-byte identical outputs.

That's useful for repeatable asset generation, although reproducibility across different hardware hasn't been established.

Important limitations:

▪️ The back and underside are inferred, not reconstructed from actual observations.
▪️ Models aren't automatically rigged, watertight or ready for 3D printing.
▪️ The wrapper uses AGPL-3.0, while Tencent's model license has separate restrictions, including geographic exclusions for the EU, UK and South Korea.

Why this matters:
The interesting development isn't just generating 3D objects from images. We've had that capability for a while.
It's making the entire process accessible to autonomous development agents.
Imagine an AI game-development workflow that creates an object, generates its 3D asset, checks multiple viewpoints and imports it into the project, all through tools.

MCPBytes provides one practical building block for that workflow.

💻 https://github.com/MCPBytes/mcpbytes-image-to-3d

📖 https://mcpbytes.com/docs/image-to-3d-model

#GenerativeAI #GameDev #MCP #OpenSource