TripoSplat vs SHARP vs TRELLIS: Single-Image 3D Compared
Three open models now turn a single image into 3D, and people keep asking which one to use: TripoSplat (VAST-AI), Apple's SHARP, and Microsoft's TRELLIS. They look similar from the outside, but they make different things, carry different licenses, and are built for different jobs. This guide compares all three so you can pick the right one in a minute.
- One object → a clean Gaussian splat, free and commercial-OK: use TripoSplat.
- A whole scene from one photo, fast, research only: use Apple SHARP (it cannot be used commercially).
- A textured mesh or text-to-3D, multi-format: use TRELLIS (needs a 16 GB GPU).
The three tools at a glance
| Tool | Maker | Input | Output | Commercial use? | Best for |
|---|---|---|---|---|---|
| TripoSplat | VAST-AI / TripoAI | Single image | Gaussian splat (up to 262,144), .ply / .splat |
Yes — MIT (code and weights) | Object-centric splats: products, characters, game assets |
| Apple SHARP | Apple | Single image | Scene-style Gaussian splat, .ply |
No — research-only license | Whole-scene view synthesis from one photo, in under a second |
| TRELLIS | Microsoft + Tsinghua / USTC | Image or text | Gaussian splats, radiance fields, or textured meshes (GLB) | Yes — MIT (weights and most code) | Mesh / game assets, text-to-3D, multi-format output |
What each one actually makes
The biggest difference is not quality, it is the kind of 3D you get.
TripoSplat is object-centric. It converts one image into a standalone object you can orbit fully, made of up to 262,144 Gaussians, and exports .ply or .splat. Think a single product, prop, or character — not a room. For a step-by-step walkthrough, see our TripoSplat tutorial.
SHARP is scene-centric. It regresses a metric-scale Gaussian splat of the whole scene in the photo, so you can move the camera to nearby viewpoints. It is the speed champion (under a second) but it reconstructs what the camera roughly saw rather than generating a clean, fully-rotatable object.
TRELLIS is multi-format. Its Structured LATent (SLAT) representation can decode the same input into Gaussian splats, radiance fields, or textured meshes, and it accepts text prompts as well as images. If your pipeline needs a GLB mesh with PBR textures for a game engine, TRELLIS is the only one of the three that produces it directly. It became the top-ranked open-source model on Hugging Face's 3D Arena leaderboard. The trade-off is weight: it needs an NVIDIA GPU with 16 GB or more of VRAM to run locally.
Can you use it commercially?
This is the question most comparisons skip, and it is the one that can get you in trouble. The three models are not equal here.
- TripoSplat — yes. The repository states that "TripoSplat code and weight models are released under the MIT License," which permits commercial use.
- TRELLIS — yes. Its model weights and the majority of the code are MIT-licensed, so commercial use is permitted.
- Apple SHARP — no. SHARP's model license restricts use to research purposes, and states research purposes "does not include any commercial exploitation, product development or use in any commercial product or service." If you put a SHARP-generated splat into a client project, a game, or a product page, you are outside its license.
So if the work is commercial, your realistic choices are TripoSplat (for object splats) or TRELLIS (for meshes). SHARP is the impressive demo you cannot ship. In every case you are still responsible for having the rights to the input image you upload.
Which one should you use?
- You have one photo of a single object and want a usable, sellable splat: TripoSplat. It is purpose-built for objects, MIT-licensed, and the fastest to start (no install needed in the browser).
- You want to "step into" a photographed scene and it is for research or personal use: SHARP. Fastest, scene-aware, but non-commercial.
- You need a textured mesh for Blender/Unity/Unreal, or you want to generate 3D from a text prompt: TRELLIS, if you have a 16 GB GPU or use its hosted Space.
- You are not sure and just want to see your image as 3D right now: start with TripoSplat in the browser, then branch out.
How to try each one free
All three are free to try. SplatDrop is an independent product built on the open-source TripoSplat model by VAST-AI / TripoAI, running on fal.ai. Not affiliated with VAST-AI / TripoAI.
- TripoSplat: the quickest path is to convert an image to a gaussian splat online free in the browser — no signup, instant
.plydownload. You can also run the official Hugging Face Space or the native ComfyUI template (v0.23.0). - Apple SHARP: run it from the official GitHub repo or a community Hugging Face Space. Remember the weights are research-only.
- TRELLIS: use the official Microsoft/TRELLIS Hugging Face Space, or install locally (NVIDIA GPU, 16 GB+ VRAM). ComfyUI support is community-built, not native.
FAQ
- Can I use a model made with Apple SHARP commercially?
- No. Apple SHARP's model weights are licensed for research purposes only, and the license excludes commercial products and product development. For commercial work, use TripoSplat or TRELLIS, which are both MIT-licensed.
- Is TripoSplat free for commercial use?
- Yes. TripoSplat's code and weight models are released under the MIT License, which permits commercial use. You are still responsible for having the rights to the image you upload.
- Does TRELLIS make meshes or Gaussian splats?
- Both. TRELLIS decodes one input into Gaussian splats, radiance fields, or textured meshes (GLB export), so it suits mesh and game-asset pipelines as well as splats.
- Which single-image tool is fastest?
- Apple SHARP is the fastest, generating a scene-style Gaussian splat from one photo in under a second.
- Which should I use for a single object like a product or character?
- TripoSplat is built for object-centric Gaussian splats and is MIT-licensed, so it is the simplest free, commercial-friendly choice for turning one object photo into a splat.
- Do any of these run in a browser without install?
- Yes. TripoSplat runs in the browser via SplatDrop or its official Hugging Face Space, and SHARP and TRELLIS each have official Hugging Face Spaces. Local installs need an NVIDIA GPU, and TRELLIS needs 16 GB or more of VRAM.