StableZero123_Conditioning

StableZero123_Conditioning is the node that turns a single photo into the conditioning for a 3D model - specifically Stability AI's Stable Zero123, the 2023-era model that took one image and synthesized the other views around it. It's the grandfather of this whole corner of ComfyUI: every multiview-conditioning node that came later, from SV3D to Hunyuan 3D, is walking the path this one laid down.

The mechanism is the part worth learning, because it's the same recipe they all use. The node takes your init_image, extracts what the object looks like (via CLIP vision), and decides where the camera is looking from (via elevation/azimuth angles), then bakes both into the conditioning. That camera-angle embedding is the secret: the model isn't just told "here's the object," it's told "here's the object seen from this viewpoint," which is what lets it render novel angles coherently.

How it works

Concretely: the CLIP vision encoder produces the image embedding, the camera embedding is computed from elevation and azimuth (in degrees, both default 0), and the two are concatenated into the positive conditioning. The init image is also VAE-encoded into the conditioning as a latent ("concat_latent_image") so the renderer has the actual pixels to work from. The negative gets zeros - nothing here to avoid. A fresh empty latent at your width/height and batch_size (both default 256) is created for the sampler to denoise. That latent is the output that actually becomes your multiview frames.

Inputs and outputs

• clip_vision - the CLIP vision model paired with Stable Zero123.
• init_image - your photo.
• vae - encodes the image into the conditioning.
• elevation / azimuth - the camera angle of the input relative to the object (default 0, range ±180).
• width / height / batch_size - output size, default 256.
• Outputs: positive, negative, and the empty latent to sample.

Where it fits

Ships with ComfyUI core, part of the stable3d node family (with SV3D and the batched variant sitting beside it). Model weights - the Stable Zero123 checkpoint plus its CLIP vision - go in your ComfyUI models folder. By 2026 it's a legacy pick: Stable Zero123's outputs are rough by modern standards, and Hunyuan 3D or Tripo do far better. But it's still the lightest, simplest introduction to the image-to-3D pipeline, and it runs on hardware that can't sniff the newer models.

Common issues

The input image does most of the work - a cluttered background or a subject that isn't centered produces a blob of a 3D model, so crop and isolate your subject first. The elevation/azimuth defaults assume the input was shot at eye level facing the object; if your reference is a top-down or angled photo, set them to match or the model gets confused about where the camera was. And don't expect fidelity: Stable Zero123 invents plausible backs rather than reconstructing true ones - that's the model's nature, not a settings error.