Stability AI drops Stable Zero123 for image-to-3D generation

Stable Zero123 takes a single 2D image, figures out how that object looks from different camera angles, and turns it into a textured 3D mesh.

Built inside Stability AI, this model targets the hard problem of spatial consistency, predicting novel viewpoints without turning the backside of your subject into melted plastic.

How it works

Compared to its predecessor Zero123-XL, Stable Zero123 generates noticeably cleaner and sharper view variations thanks to three main upgrades.

First, data quality got a massive cleanup. The team aggressively filtered the open Objaverse collection, keeping only high-fidelity 3D assets and rendering them with much more realistic lighting.

Second, they fed the model estimated camera elevation angles during both training and inference. Having that elevation anchor keeps the perspective consistent instead of guessing wildly.

Third, an improved dataloader paired with pre-computed latents allowed for much larger batch sizes, yielding a 40x speedup in training efficiency over Zero123-XL.

You can also chain this into a full text-to-3D pipeline: spin up a concept image in SDXL, feed that frame into Stable Zero123, and let the 3D pipeline do the rest.

Hardware and licensing

Because Stable Zero123 is built on Stable Diffusion 1.5, generating a single novel viewpoint takes roughly the same VRAM as generating a standard SD1.5 image.

However, building an actual 3D asset is significantly heavier. The team patched the open-source threestudio repo to support Stable Zero123, running Score Distillation Sampling (SDS) to optimize a NeRF and export a textured mesh.

That reconstruction process takes noticeable time and memory, so Stability AI recommends a 24GB VRAM GPU.

The release splits into two flavors. The standard Stable Zero123 includes some CC-BY-NC training data and is strictly non-commercial, available now on Hugging Face.

The commercial release, Stable Zero123C, was trained strictly on CC-BY and CC0 datasets to clear licensing hurdles, but commercial rights require an active Stability AI membership.

Why it matters

Single-image 3D generation is moving from messy academic demos toward usable developer workflows. Stability AI showed that high-quality data filtering and smart camera conditioning deliver better spatial awareness without ballooning the base model size.

While the 24GB VRAM recommendation keeps full NeRF-to-mesh reconstruction out of reach for everyday laptops, technical artists and indie devs with proper rigs should definitely plug this into their pipeline.