Guides
Stable Diffusion to 3D: getting a real mesh out of an open image model
Stable Diffusion makes images; Stable Fast 3D and SPAR3D make meshes. How to prompt a clean input, convert it, and check the licence.
2 min read
Quick answer
Can Stable Diffusion make 3D models?
Stable Diffusion itself makes images. Stability AI publishes separate image-to-3D models, Stable Fast 3D and SPAR3D, that turn one image into a textured mesh. Any image-to-3D generator works the same way: render a clean, centred picture of one object in Stable Diffusion, then convert it and export GLB, OBJ or FBX.
Images first, mesh second
Stable Diffusion is a family of open image models you can run on your own GPU, in tools like ComfyUI or Automatic1111. None of them write a 3D file. For a mesh you add an image-to-3D step, and there are two kinds:
- Stability's own 3D models. Stable Fast 3D (2024) and SPAR3D (2025) each take one image and return a textured mesh. Both are on Hugging Face under the Stability AI Community Licence, which allows commercial use below an annual revenue threshold.
- Any other image-to-3D model: TRELLIS, Hunyuan 3D, Tripo and so on. They don't care which tool drew the picture.
Rendering a good input in Stable Diffusion
Stable Diffusion gives you more control than a chat tool, so use it:
- Prompt for one object, "full object in frame, centred, front view, plain white background, soft studio lighting".
- Negative prompt: "cropped, shadow, text, watermark, scenery, multiple objects, reflections".
- Aspect ratio to match the object: portrait for a character, landscape for a car.
- ControlNet with a sketch or a depth map, if you already have the shape you want. This is the main reason to use Stable Diffusion over a chat tool: the silhouette stays yours.
- Background removal before export. A clean alpha edge helps the 3D model separate the object.
Best photos for image to 3D explains why each of these matters.
Running the 3D step
You have three choices:
- Locally. Stable Fast 3D, SPAR3D and TRELLIS all have code on GitHub and need a CUDA GPU. TRELLIS asks for at least 16 GB: what is TRELLIS.
- In a Hugging Face Space, free but with a daily GPU quota: image to 3D on Hugging Face.
- In a hosted app. In Vi3W, upload the image and pick a 3D model: TRELLIS on every plan, Hunyuan 3D and Tripo on paid plans. The mesh is retopologised and exported as GLB, FBX or OBJ.
Licences, twice
Two licences apply: the one for the image model you used, and the one for the 3D model. Stable Diffusion versions differ (some are under the Stability Community Licence, older ones under CreativeML OpenRAIL). Check both before you sell anything.
Try it
Have a render you like? Upload it to Vi3W and get a mesh back without setting up a second model.
Last reviewed