SV3D_Conditioning_motorway_edition

Stable Video 3D (SV3D) is Stability's video sibling of Zero123: instead of one rotated view, it generates a whole orbiting video of your object from a single input image. SV3D_Conditioning is the node that turns the init image, CLIP vision, VAE, and camera elevation into the conditioning + latent for that video. This is the Motorway edition - inputs come out of the pipe by key, outputs go back in.

Mechanically it's the Zero123 recipe with video-shaped knobs. The image goes through CLIP vision to extract the subject, the VAE encodes it into the latent space, and the elevation plus a video frame count tell the model "this is a 21-frame orbit." Then the sampler generates the frames. SV3D is still the best pure "give me a turntable of this object" toy in ComfyUI - the results are jittery at the seams like all early video models, but as a rotating preview of a 3D-ish render it's a fun thing to have in the toolbox.

What to set

• INPUT_clip_vision_key / INPUT_init_image_key / INPUT_vae_key (defaults clip_vision / init_image / vae) - the usual trio, all from the pipe by key.
• width / height (default 576 × 576) - SV3D renders at 576, notably bigger than Zero123's 256.
• video_frames (default 21) - the number of frames in the orbit. 21 is the sweet spot the model was trained around.
• elevation (default 0, -90 to 90) - camera tilt. Keep it near 0 for a level orbit.
• OUTPUT_positive_key / OUTPUT_negative_key (defaults positive / negative) - conditioning output keys.
• OUTPUT_latent_key (default latent) - the video latent output key.

Output is the MOTORWAY with positive, negative, and latent stored, ready for a KSampler and then a video decode.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/agilly1989/ComfyUI_agilly1989_motorway

Restart; it's under agilly1989 Nodes → Motorway-ed → conditioning → 3d_models. Manager has the pack. No pip dependencies.

Where people get burned

VRAM, mostly. 576×576×21 frames is a chunky latent, and SV3D is not small. If you're on 8GB, be ready to decode the video in chunks or drop to fewer frames.

The model-pairing rule from Zero123 applies: CLIP vision, VAE, and checkpoint all have to be the SV3D family - mix in a Zero123 encoder and the conditioning silently doesn't align, giving you mush. And the standard Motorway trap: keys are case-sensitive and exact, and 'MotorwayClass' object has no attribute 'hash_' means a key isn't in the pipe. Beta pack, motorway-ed nodes come and go across releases - a vanishing node after an update is upstream churn, not a broken install.