EasyLoader (Zero123)

Zero123 is the "novel view synthesis" model - you give it a single image of an object and a camera angle, and it imagines what that object looks like from that new angle. It's the ComfyUI on-ramp to single-image 3D-ish rotation, and easy zero123Loader is the one-node way to set it up: load the model, hand it your starting image, dial in elevation and azimuth, get a pipe you can sample.

Setting expectations: this is a niche corner. Zero123 (from the Zero-1-to-3 research line, later refined by Stability as "Stable Zero123") barely registers in community discussion - under 80 corpus threads all-time, most of them clustered around release. It's a specialist tool for turning product shots or a single character render into a few alternate angles, not a general image node. If that's your use case, this loader saves you real wiring. If it isn't, you can skip it.

How it works

The node loads the Zero123 checkpoint and its VAE, takes your init_image as the reference view, and encodes the camera transform you request into the conditioning. Under the hood Zero123 was trained on rendered 3D objects across many viewpoints, so it has a learned sense of how objects rotate - it's not doing real geometry, it's predicting a plausible new render. The node packs all this into a PIPE_LINE for the sampler to finish.

The inputs that matter

• init_image - your source image. Zero123 works best on a single, cleanly isolated object on a plain background. Busy scenes with multiple subjects confuse it.
• elevation and azimuth (both -180 to 180, default 0) - the camera move. Azimuth swings the camera around the object horizontally; elevation tilts it up or down. Small moves (say 30-45 degrees of azimuth) hold together far better than trying to spin all the way to the back, where the model has to invent detail it never saw.
• empty_latent_width / empty_latent_height (default 256) - Zero123's native resolution is small. The default 256 is not a typo; this model was trained at low res and pushing it much higher tends to fall apart.
• batch_size - generate several angle variations at once.
• ckpt_name and vae_name - pick the Zero123 model and its VAE.

Outputs: pipe for the sampler chain, plus model and vae broken out if you want them raw.

How to install it

Install the pack through ComfyUI Manager ("ComfyUI Easy Use") or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

Then install.bat / pip install -r requirements.txt, restart. It's already on comfy.icu. You'll also need the Zero123 model itself - the stable-zero123 checkpoint from Stability, placed in your checkpoints folder - before the ckpt_name dropdown will show anything.

Common issues

The ckpt_name dropdown coming up empty is the usual first snag: no Zero123 model downloaded, nothing to pick. Grab stable_zero123.ckpt from Stability's Hugging Face and drop it in.

The bigger reality check is quality. Zero123 is genuinely impressive for what it is but it's a 2023-era model at 256px - expect soft, sometimes warped results, especially at extreme angles. It shines for small rotations of a clean single object and struggles with everything else. If you crank the resolution up expecting sharp output, you'll get artifacts instead; keep it near native and upscale afterward if you need more pixels. And like every Easy-Use loader, the pipe it produces is meant for the pack's own sampler nodes downstream.