Skip to main content
TripoSplat is an open-source model that generates 3D Gaussian splat representations directly from a single 2D image. It was developed by VAST-AI and released under an open-source license. Unlike traditional 3D reconstruction methods that require multiple views or generate meshes as the primary output, TripoSplat creates Gaussian splat representations โ€” a rendering technique where thousands of colored 3D Gaussians are placed in space to represent a scene. This approach enables fast, high-quality rendering with controllable density and budget. TripoSplat workflow
Make sure your ComfyUI is updated.Workflows in this guide can be found in the Workflow Templates. If you canโ€™t find them in the template, your ComfyUI may be outdated. (Desktop versionโ€™s update will delay sometime)If nodes are missing when loading a workflow, possible reasons:
  1. You are not using the latest ComfyUI version (Nightly version)
  2. Some nodes failed to import at startup

Download Workflow

Download JSON or search โ€œTripoSplatโ€ in Template Library

โ€‹
How it works

TripoSplat uses a feed-forward architecture that takes a single RGB image and directly predicts a set of 3D Gaussian primitives. The pipeline involves:
  1. Image encoding โ€” the input image is processed by a vision encoder (DINOv2)
  2. Triplane generation โ€” features are decoded into a triplane representation
  3. Gaussian prediction โ€” the triplane is sampled to produce Gaussian parameters (position, scale, rotation, opacity, color)
  4. Rendering โ€” Gaussians are rendered from arbitrary viewpoints using differentiable splatting

Learn about Subgraph

This workflow uses a Subgraph node for modular processing. Check out the Subgraph documentation to learn how to customize and extend the workflow.

โ€‹
Workflow node guide

โ€‹
LoadImage

  • Loads your input image (PNG/JPG)
  • Sample image: white-hotel-on-rocky-island.png (available in Template Library)

โ€‹
TripoSplat (subgraph)

The main subgraph node processes the image and generates the 3D Gaussian splat. Exposed parameters:

โ€‹
CreateCameraInfo

  • Defines the camera orbit for rendering the result
  • Parameters: orbit type, angle, distance, field of view, etc.
  • Default: orbit at 35ยฐ elevation, 30 distance, 2.5 zoom

โ€‹
RenderSplat

  • Renders the Gaussian splat into a 2D image from the defined camera angle
  • Parameters: output resolution (default 1024ร—1024), image quality settings

โ€‹
SplatToMesh

  • Converts the Gaussian splat to a mesh (optional)
  • Parameters: mesh density, smoothing, simplification

โ€‹
SaveGLB

  • Saves the result as a GLB 3D file

โ€‹
SaveVideo

  • Saves a video of the rendered 3D scene

โ€‹
SplatToFile3D

  • Exports the Gaussian splat in SPZ format

โ€‹
CreateVideo

  • Creates a video from rendered frames

โ€‹
Steps to run

  1. Load an image โ€” use the LoadImage node to load a single 2D image
  2. Run the TripoSplat subgraph โ€” the model will generate a Gaussian splat representation
  3. Choose output format โ€” export as GLB, SPZ, video, or render to mesh
  4. View results โ€” use the created 3D file or rendered preview

โ€‹
Output options

โ€‹
Model downloads

Download the TripoSplat model and required files. Place them in the corresponding models/ subdirectories.

TripoSplat diffusion

triposplat_fp16.safetensors โ€” TripoSplat diffusion model checkpoint

TripoSplat VAE decoder

triposplat_vae_decoder_fp16.safetensors โ€” VAE decoder

Flux2 VAE

flux2-vae.safetensors โ€” Flux.2 VAE for latent encoding

DINOv2 CLIP

dino_v3_vit_h.safetensors โ€” CLIP vision encoder (DINOv2)

BiRefNet bg removal

birefnet.safetensors โ€” Background removal model for preprocessing

โ€‹
Model storage location