Intermediate20 min read

Complete TripoSplat ComfyUI Workflow Guide: Single 2D Image to 3D Gaussian Splat & Mesh Export

Step-by-step tutorial on running TripoSplat natively in ComfyUI — covering 2D-to-3D Gaussian splat generation, BiRefNet background removal, DINOv3 feature encoding, and GLB/SPZ export.

By CGFlowStudio

•

2026-08-20

Complete TripoSplat ComfyUI Workflow Guide: Single 2D Image to 3D Gaussian Splat & Mesh Export

Turning flat 2D concept art and character illustrations into interactive 3D assets used to require multi-view camera capture, complex photogrammetry software, or hours of manual modeling in Blender. TripoSplat — developed by VAST-AI and Tripo — completely transforms this pipeline.

By employing a fast feed-forward 3D Gaussian Splatting architecture, TripoSplat generates high-fidelity, controllable-density 3D Gaussian Splats directly from a single 2D reference image in seconds.

With official native support integrated into ComfyUI, converting 2D images to 3D models is seamless. In this workflow guide, we cover everything you need: installing model weights, loading the official TripoSplat: Image to Gaussian Splat template, tuning density settings, rendering camera orbits, and exporting ready-to-use .glb, .spz, and polygon mesh files.

1. Introduction & Why TripoSplat Matters

Unlike traditional NeRF or multi-view diffusion models that require generating 4 to 6 surrounding camera angles first, TripoSplat uses a direct single-pass vision encoder and triplane decoder to predict millions of 3D Gaussians instantly.

💡

What Makes 3D Gaussian Splatting Special? Traditional polygon meshes rely on flat vertices and faces. 3D Gaussian Splats represent objects as millions of volumetric ellipsoids (with position, scale, rotation, opacity, and color). This enables ultra-fast real-time rendering, soft translucency, and realistic specular highlights.

Key Advantages

Single 2D Image Input: No multi-view camera capture or multi-angle image prompts required.
Feed-Forward Speed: Generates 3D Gaussian representations in under 10 seconds.
Controllable Density (num_gaussians): Balance rendering fidelity against VRAM and export file size.
Multi-Format Exporting: Native export to .glb (for Blender, Unity, Unreal Engine), .spz (compressed 3D Gaussian format), 360° video previews, and polygon meshes.

2. Getting Started – Loading the Official Workflow

ComfyUI includes a built-in native workflow template for TripoSplat.

1 Update ComfyUI

Update ComfyUI to the latest release to ensure core 3D nodes (TripoSplat, RenderSplat, SaveGLB, SplatToFile3D) are registered.

2 Open Template Library

Click Templates in the ComfyUI top menu and search for TripoSplat: Image to Gaussian Splat.

3 Inspect Workflow Pipeline

The template routes a single LoadImage node through BiRefNet background isolation, into the TripoSplat Subgraph, and out to rendering and export nodes.

3. How TripoSplat Works (Simplified Pipeline)

Understanding the feed-forward mechanism helps you select better input images and tune settings effectively:

TripoSplat Feed-Forward Pipeline

Step 01

Image Encoding

DINOv3 vision encoder extracts spatial features.

Step 02

Triplane Generation

Decodes features into 3D triplane representation.

Step 03

Gaussian Prediction

Predicts position, scale, rotation, opacity, & color.

Step 04

Differentiable Splatting

Renders 3D splats to 2D previews or 3D files.

4. Model Installation & Directory Structure

Download and place all required TripoSplat, BiRefNet, DINOv3, and VAE weights into their designated ComfyUI folders:

📂 TripoSplat Model Component Directory Layout ComfyUI/models/

📂 ComfyUI/models/

📁 diffusion_models/ (or triposplat/)

📄 triposplat_fp16.safetensors (Core TripoSplat Model)

📁 clip_vision/

📄 dino_v3_vit_h.safetensors (DINOv3 Spatial Feature Encoder)

📁 vae/

📄 triposplat_vae_decoder_fp16.safetensors
📄 flux2-vae.safetensors (Spatial VAE Decoder)

📁 background_removal/ (or rembg/)

📄 birefnet.safetensors (High-Resolution Subject Matting)

5. Key Workflow Nodes Explained

Understanding the key nodes inside the template ensures you can tweak parameters for optimal results:

NODE 01 Core Subgraph

TripoSplat Subgraph

Encapsulates feature extraction, background removal via BiRefNet, and Gaussian generation. Key parameter: num_gaussians (default: 409,600).

NODE 02 Camera Orbit

CreateCameraInfo & RenderSplat

Defines 360° camera orbit angles (elevation, azimuth, distance) and renders differentiable splatting into 2D image preview frames.

NODE 03 Mesh Converter

SplatToMesh (Optional)

Converts volumetric Gaussian point clouds into standard textured polygon meshes (`.obj` / `.glb`) using Poisson surface reconstruction.

NODE 04 3D Exporters

SaveGLB & SplatToFile3D (.spz)

Saves final 3D assets to `.glb` format for game engines or `.spz` compressed format for real-time web Gaussian viewers.

6. Step-by-Step Guide to Generating a 3D Splat

Follow these four steps to convert your 2D image into a 3D Gaussian Splat:

Step 1: Load Your Input Image

Load a single 2D reference image into the Load Image node.

Best Inputs: Stylized character art, isolated product shots, 3D concept renders, or game props with clear silhouettes.

Step 2: Configure Gaussian Density & Background Removal

Double-click the TripoSplat Subgraph to adjust generation settings if needed:

num_gaussians: Set to 409,600 for high fidelity, or drop to 200,000 for lower VRAM setups.
remove_background: Keep enabled (true) to run BiRefNet automatically.

Step 3: Run the Workflow

Click Queue Prompt. ComfyUI will isolate your subject, extract features via DINOv3, and generate the 3D Gaussian Splat representation in 5 to 10 seconds.

Step 4: Preview & Export Your Asset

Review the 360° rendered preview video generated by RenderSplat + SaveVideo. Your exported 3D files will save directly into your ComfyUI/output/ directory:

.glb file via SaveGLB
.spz file via SplatToFile3D

7. Output Format Comparison

Select the right export format based on your downstream 3D pipeline:

Export Format Node	Output Extension	Best Used For
SaveGLB	`.glb`	Standard 3D software (Blender, Maya), web 3D viewers, Unity, & Unreal Engine.
SplatToFile3D	`.spz`	Highly compressed 3D Gaussian Splat format for real-time web viewers.
RenderSplat + SaveVideo	`.mp4` / `.webp`	360° rotating video showcase for social media & concept presentations.
SplatToMesh	`.obj` / `.glb`	Converting splat point clouds into polygonal meshes for sculpting or 3D printing.

8. Practical Tips & Troubleshooting

✂️ Importance of Clean Background Removal

If background pixels bleed into your subject, TripoSplat will attempt to render background clutter into 3D floaters. Keep BiRefNet enabled to isolate your subject cleanly.

⚡ Tuning Gaussian Density for Low VRAM

If experiencing VRAM slowdowns on 8GB to 12GB GPUs, reduce num_gaussians from 409,600 down to 150,000–200,000 inside the TripoSplat subgraph.

🎥 Adjusting Camera Orbit Preview

If the preview render cuts off your model, increase the distance value inside the CreateCameraInfo node to zoom out the camera radius.

9. Conclusion & Next Steps

TripoSplat makes single-image 2D-to-3D generation accessible right inside ComfyUI. By combining DINOv3 feature encoding, automated BiRefNet background matting, and fast Gaussian splatting, you can turn character sketches and product concepts into interactive 3D assets in seconds.

Next Steps to Explore

Import your exported .glb files directly into Blender or Unreal Engine 5.
Experiment with using rendered 360° 2D frames as conditioning input for video models like Wan 2.2 or LTX-2.5.
Explore related 3D guides on our ComfyHub Tutorials Directory.