Stablezero123

Feed it one picture of an object and it hands you back six views of that same object from different angles - front, side, three-quarter, the works - all in a single tiled image. That's the whole trick. It's the ComfyUI on-ramp to "I have one photo and I want to turn this thing into 3D," which usually means generating a bunch of consistent angles and then handing them off to photogrammetry or a gaussian-splatting pipeline downstream.

First, a naming thing that trips people up. Despite the pack being called StableZero123, this node doesn't run Stability AI's "Stable Zero123" model. Its defaults point at sudo-ai/zero123plus-v1.1 - that's Zero123++ from the SUDO-AI team, a different (and honestly more capable for this multi-view job) model. So if you came here looking for Stability's checkpoint, that's not what's under the hood. The README is upfront that it's a wrapper around SUDO-AI-3D/zero123plus.

How it works

Zero123++ is a multi-view diffusion model. Instead of denoising toward "a nice image," it's been trained to denoise toward a fixed layout of six novel camera angles of whatever object you gave it, using reference attention so the six views stay consistent with each other - same object, same lighting, same colors, just rotated. The output isn't six separate files; it's one image with the six views tiled in a grid (three rows, two columns). You split that grid apart afterward - which is exactly what its sibling node, SDZero-ImageSplit, exists to do.

Two things it really cares about, both from the README and how the model was trained: your input should be a square image, and it works best on a single, centered object on a clean background. This is not a "reangle my crowded living room" tool. Product shots, a single character, a toy, a prop - that's the lane.

The inputs that matter

• images - your source image (IMAGE). Square. One object. Clean background if you can manage it.
• inference_steps - default 75, range 1–100. This is your quality/speed dial. 75 is already generous for a diffusion pass; if you're just checking whether an object works at all, drop it to speed things up, then crank it back for the final.

The other two inputs, ckpt_name (sudo-ai/zero123plus-v1.1) and pipeline_name (sudo-ai/zero123plus-pipeline), are just HuggingFace repo IDs. Leave them alone. They're strings rather than dropdowns because the node pulls those repos from HuggingFace directly rather than reading a local checkpoint folder.

The single output is an IMAGE - the tiled six-view grid. Wire it into a Save Image if you just want to eyeball the result, or into SDZero-ImageSplit (columns 2, lines 3) to break it into the six individual frames you'll actually use downstream.

Installing it

Nothing fancy. Either search "StableZero123-comfyui" in ComfyUI Manager and install, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/deroberon/StableZero123-comfyui

Then restart ComfyUI. The first time you run the node it downloads the Zero123++ model and pipeline from HuggingFace (those two repo-ID defaults), so the very first execution needs a live connection and will sit there a while pulling weights before anything happens. After that they're cached.

Where people get burned

The big one: it's explicitly experimental. The author's own words - "this node is under development, so use it at your own risk," and the interface "will probably change a lot," which can break saved workflows on a future version. Treat it as a fun capability, not a production dependency.

Beyond that, most disappointment comes from the input. A non-square image, a busy scene, or multiple objects and the consistency falls apart or the views look mushy - this class of model was always aimed at clean single objects. And manage expectations on the output: six diffusion-generated angles are consistent-ish, not a laser-accurate turntable. They're good enough to feed a reconstruction step, which was always the point; going from one photo to actual 3D still means a photogrammetry or splatting stage after this, not this node alone.