Open-Source Single-Image-to-3D Runner Guide
TripoSR Online Demo: Run Free Image-to-3D in Your Browser (No Local GPU Install)
TripoSR (jointly released under the MIT License by Tripo AI and Stability AI) reconstructs a 3D triangle mesh from a single photo in roughly 0.5 seconds on an A100 GPU. Below is the complete, practical guide to running the official stabilityai/TripoSR Hugging Face Space, spinning up a free Google Colab T4 fallback when queues stall, tuning foreground_ratio = 0.85, and exporting clean .obj or .glb files.
3 Ways to Run the Open-Source TripoSR Demo (Zero Cost Comparison)
Because TripoSR's weights (model.ckpt, ~1.4 GB) and code are MIT-licensed, you are never locked into a single server. Choose the execution path that fits your hardware and privacy requirements:
| Runner Option | Setup Time | GPU & Speed | Export Formats | Best When |
|---|---|---|---|---|
1. Official Hugging Face Space (stabilityai/TripoSR)
| 0 sec (instant browser) | Shared A10G / ZeroGPU (~0.5s–4s compute + queue) | Vertex-color .obj, .glb | Quick 1–3 image tests with zero login or install |
| 2. Google Colab Free T4 Notebook | ~3 min environment install | Dedicated Tesla T4 15 GB VRAM (~3s per mesh) | .obj, .glb + UV baked PNG via CLI | HF Space queue is full or you need batch CLI flags |
3. Local python app.py / ComfyUI Node
| 5–10 min local clone | 6 GB+ NVIDIA CUDA GPU (~0.5s A100 / ~2.5s RTX 4070) | .obj, .glb, UV maps, custom mc_resolution | Offline NDA assets & automated node pipelines |
Step-by-Step: Running the Official Hugging Face TripoSR Demo
The fastest zero-install entry point is the upstream Gradio Space maintained at huggingface.co/spaces/stabilityai/TripoSR (you can also test the live interactive Space embed on the TripoSR AI homepage ). Follow these five steps to avoid clipped geometry and melted backgrounds:
Prepare a clean single-object reference photo or PNG
Choose a well-lit front or three-quarter angle image showing a single subject without heavy motion blur, extreme perspective distortion, or cropped limbs. Transparent PNGs or high-contrast backgrounds produce the cleanest silhouette mask.
Open the stabilityai/TripoSR Hugging Face Space or Colab Gradio runner
Launch the official stabilityai/TripoSR Space in your browser, or start the free Google Colab T4 notebook if the public Hugging Face ZeroGPU queue is congested.
Enable background removal and set Foreground Ratio to 0.85
Leave Remove Background checked so rembg isolates the subject onto a neutral matte, and keep Foreground Ratio at 0.85 (adjusting down to 0.75 if elongated parts touch the frame border).
Run ~0.5s GPU reconstruction and inspect the 360-degree preview
Click Generate to run the DINOv2 image encoder, triplane transformer decoder, and 256-grid Marching Cubes extractor. Rotate the interactive 3D viewport to check back-side hallucination and silhouette thickness.
Export OBJ or GLB and convert locally for your target pipeline
Download the vertex-color .obj or .glb mesh. Use TripoSR.ai's local no-upload OBJ to GLB or PLY to STL converters for downstream apps, or upgrade to cloud 4K PBR or watertight STL generation when production fidelity is required.
Mastering TripoSR Demo Settings: Foreground Ratio 0.85, Background Removal & Marching Cubes
Under the hood, TripoSR is a feed-forward Large Reconstruction Model (LRM). A pretrained DINOv2 Vision Transformer encodes your 2D image into spatial feature tokens, a cross-attention transformer projects those tokens onto three orthogonal planes (XY, XZ, YZ triplanes), and a lightweight neural radiance MLP predicts occupancy density across a bounded 3D cube. Because there is no iterative diffusion loop at inference time, the entire forward pass finishes in ~0.5 seconds on datacenter GPUs — which also means input framing directly determines mesh sharpness.
1. Remove Background (rembg)
Always keep Remove Background checked when uploading raw smartphone photos or web JPEGs. TripoSR was trained on objects isolated against a uniform neutral matte. Leaving desk clutter, shadows, or floor planes in the image causes the triplane decoder to stretch background pixels into a flat wall fused behind your model. Only uncheck this box if you manually masked a clean alpha PNG in Photoshop or PhotoCraft.
2. Foreground Ratio (Default 0.85)
foreground_ratio controls how much of the preprocessed square canvas the object's bounding box occupies before entering DINOv2. At 0.85, the subject fills 85% of the frame with a 7.5% safety margin on each side. If sword tips, ears, or chair legs look sliced flat, lower the slider to 0.75–0.78. For compact spherical objects (fruits, helmets, rocks), raising it to 0.90 packs more Marching Cubes triangles onto surface details.
3. Vertex Colors vs. UV Baking
The browser Gradio UI extracts an isosurface at mc_resolution = 256 and writes RGB values per vertex into the .obj or .glb file. Standard OBJ importers in older tools sometimes ignore non-standard vertex color extensions — which is why converting with our local
OBJ to GLB Converter
or downloading the .glb directly is the safest route for Three.js, Godot, and Blender.
No Local GPU Needed
Fallback A: Run TripoSR on Free Google Colab (Tesla T4)
When the public Hugging Face Space shows a long queue or hits a ZeroGPU quota error, open Google Colab, switch Runtime → Change runtime type → T4 GPU, and run these commands to launch your own private Gradio demo URL or bake 2K UV textures:
!git clone https://github.com/VAST-AI-Research/TripoSR.git
%cd TripoSR
!pip install --upgrade pip setuptools wheel
!pip install torchmcubes
!pip install -r requirements.txt gradio
# Option 1: Launch interactive Gradio Web UI with public share link
!python app.py --share
# Option 2: CLI run with 2048px UV texture baking instead of vertex colors
!python run.py examples/chair.png --output-dir output/ --bake-texture --texture-resolution 2048 100% Offline & Private
Fallback B: Local python app.py Gradio & ComfyUI
Have an NVIDIA RTX card with at least 6 GB VRAM? Running the Gradio server locally removes every upload limit and lets you process proprietary game or industrial prototypes offline:
git clone https://github.com/VAST-AI-Research/TripoSR.git
cd TripoSR
python -m venv .venv && source .venv/bin/activate
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt gradio
# Start local Gradio server on http://127.0.0.1:7860
python app.py
# Low-VRAM CLI mode (6 GB GPUs): lower chunk-size from 8192 to 4096
python run.py input.png --chunk-size 4096 --model-save-format glb Prefer node-based image-to-3D pipelines? Generate a ready-to-import workflow JSON in our TripoSR ComfyUI Workflow Generator or customize CUDA wheels in the TripoSR GitHub Install Helper .
What to Do After Downloading Your TripoSR Mesh (Convert, View & Prep)
Raw TripoSR outputs are only step one of a real production workflow. Depending on whether you are heading to a web viewer, a game engine, or a 3D printer slicer, use these zero-upload browser tools right here on TripoSR.ai:
.obj meshes into a compact binary .glb file locally in your browser for Three.js, Babylon.js, Unity, or WebXR.
Convert PLY to STL →
Working with scanned or reconstructed .ply meshes? Convert to binary .stl locally with zero server upload for PrusaSlicer, Cura, or Bambu Studio.
ComfyUI Workflow Generator →
Chain SDXL/Flux image generation directly into TripoSRSampler with automatic RemBG+ masking and downloadable JSON presets.
Need another format route or quick geometry check? Browse all 15 private converters in the 3D File Converter Hub or inspect vertex and triangle counts in our GLB Viewer and STL Viewer .
Honest Assessment: Where TripoSR Shines vs. Where Modern Cloud Models Win
Released in March 2024, TripoSR set a milestone for open-source feed-forward 3D reconstruction: sub-second inference, permissive MIT licensing, and modest 6 GB VRAM requirements make it unbeatable for rapid greybox blocking, synthetic dataset generation, research baselines, and local ComfyUI experimentation.
However, single-pass 256-resolution triplane models have structural limits compared to 2026 multi-stage diffusion-transformer 3D architectures. Specifically, TripoSR bakes low-frequency vertex colors instead of crisp PBR material channels, softens fine typography or hard-surface bevels, and does not enforce watertight solid topology for additive manufacturing.
- Choose open-source TripoSR when you need unlimited free local generations, MIT commercial code freedom, or sub-second proxy meshes for layout blocking.
- Choose Marble 3D AI when you need production-grade Hunyuan 3.1 Pro + TRELLIS 2 geometry with sharp 4K PBR textures (18 free signup credits, $5 one-time Starter Pack, no subscription).
- Choose Ai3DSculpt when your goal is 3D printing a portrait, pet, miniature, or PNG cutout as a watertight manifold STL in 60 seconds (20 free signup credits = 1 free Fast Draft, no credit card required).
Frequently Asked Questions: TripoSR Online Demo
Where is the official free TripoSR online demo hosted?
The official open-source TripoSR Gradio demo is hosted on Hugging Face Spaces at stabilityai/TripoSR, collaboratively released by Tripo AI (VAST AI Research) and Stability AI under the MIT License. When the shared public Space hits GPU queue limits, you can also run the identical model weights in a free Google Colab T4 notebook or locally via python app.py.
What foreground ratio and background settings work best in the TripoSR demo?
Keep Remove Background enabled (powered by rembg) unless your input PNG already has a clean transparent or neutral gray background, and start with the default Foreground Ratio of 0.85. Lower the foreground ratio to 0.75–0.80 for tall characters, weapons, or furniture legs that clip at the edges of the triplane box, or raise it to 0.88–0.90 for compact round props to maximize marching-cubes voxel density.
How fast does TripoSR generate a 3D model and what formats does it export?
Once the model weights are warm in GPU memory, the feed-forward TripoSR transformer reconstructs a 3D mesh in roughly 0.5 seconds on an NVIDIA A100 GPU (or 2 to 5 seconds on consumer RTX GPUs, plus 5 to 20 seconds of background preprocessing and queue wait on shared cloud demos). The Gradio demo exports per-vertex colored .obj and .glb meshes, while CLI runs also support UV texture baking via --bake-texture.
Why do TripoSR demo textures look blurry in Blender, Unity, or Unreal Engine?
Standard TripoSR demo outputs store RGB color directly on mesh vertices sampled from a 256-resolution Marching Cubes grid rather than high-resolution UV-mapped PBR textures. If you need sharp 4K albedo, metallic, roughness, and normal maps for game engines or web viewers, generate the asset in a dual-engine cloud tool like Marble 3D AI (Hunyuan 3.1 Pro + TRELLIS 2) or bake UV maps locally at higher marching-cubes resolution.
Can I 3D print an OBJ or GLB exported from the TripoSR online demo?
Not always without mesh repair. TripoSR extracts an isosurface from a neural triplane field, which can leave thin walls, internal cavities, or non-manifold boundaries that cause Bambu Studio, PrusaSlicer, or Cura to warn about open edges. You can convert clean scans with our local PLY to STL converter, repair shells in Blender, or use Ai3DSculpt to generate a guaranteed watertight manifold STL directly from a PNG or photo in 60 seconds.
How much VRAM do I need to run the TripoSR Gradio demo locally?
Running python app.py at the default evaluation chunk size (8192) uses roughly 6 GB of NVIDIA CUDA VRAM. On 6 GB laptop GPUs or shared desktops, pass --chunk-size 4096 in CLI scripts or use our TripoSR ComfyUI Workflow Generator with the Low-VRAM preset to prevent CUDA out-of-memory errors.