Glimpse3D
Turns a single photo into a 3D Gaussian splat model, with the heavy models running on a free Colab GPU.
- Kind
- AI pipeline
- Year
- 2025
- Built with
- Python
- PyTorch
- TripoSR
- SyncDreamer
- SDXL Lightning
- gsplat
- 1 more
The question
How little do you need to build a 3D model of an object?
Glimpse3D starts from one photo and invents the missing views with a chain of image models.
Why it exists
A 3D model usually takes many photos or a scan. Glimpse3D chains several models (TripoSR, SyncDreamer, SDXL Lightning, gsplat) so a single image is enough.
How it works
From a photo to a 3D model
- 1
Photo
The image is checked for size, aspect ratio and blur before any GPU time is spent.
- 2
Coarse mesh
TripoSR builds a rough 3D shape from the photo.
- 3
Views
SyncDreamer produces 16 views around the object, which can be sharpened with SDXL Lightning and ControlNet.
- 4
Splat
gsplat optimises a cloud of Gaussians against those views, and the MVCRM step refines it for consistency.
- 5
Export
The model leaves as .ply, .glb or .obj.
Decisions
A failed stage degrades the result, not the job
Each stage falls back to the previous output and adds a warning, so a failure in the optional steps still leaves a usable model.
One model on the GPU at a time
Each stage loads its model, runs, unloads and clears the cache, which is how a chain of large models fits a 15 GB T4.
An honest demo mode
In demo mode a clearly labelled placeholder can stand in for a failed run, and the run is never reported as a success.
Trade-offs and limits
- The Colab backend lives only as long as the Colab session, because the tunnel closes with it.
Details
- README estimate, not from a saved run: about 30 minutes on a free Colab T4, about 12 GB of VRAM at peak
Built with
Languages
- Python
AI and models
- PyTorch
- TripoSR
- SyncDreamer
- SDXL Lightning
- gsplat
- MiDaS