SVD_img2vid_Conditioning_motorway_edition

Stable Video Diffusion (SVD) turns a still image into a short video - the classic "image to video" base model. SVD_img2vid_Conditioning is the node that takes your image, the CLIP vision encoder, and the VAE, and builds the conditioning + latent that the sampler turns into frames. This is the Motorway edition, so the three inputs come out of the pipe by key and the outputs go back in under keys.

The important part of this node that trips people up: SVD doesn't use a text prompt. The conditioning it builds is image-conditioned - the CLIP vision output carries "what this picture shows," and the knobs like motion_bucket_id and fps are how you actually steer the video. motion_bucket_id is the movement amount (higher = more motion, but more artifact risk) and fps sets the playback rate of the generation. That's why the outputs are named positive and negative but there's no prompt anywhere - they're image-derived conditioning objects, not text.

What to set

• INPUT_clip_vision_key / INPUT_init_image_key / INPUT_vae_key (defaults clip_vision / init_image / vae) - the trio, all read from the Motorway by key.
• width / height (default 1024 × 576) - SVD's native aspect: wide, not square.
• video_frames (default 14) - frame count. 14 at 6fps is about 2.3 seconds, which is what SVD was trained for.
• motion_bucket_id (default 127, 1–1023) - the motion slider. 127 is a middle ground; push toward 200+ for more movement.
• fps (default 6) - generation fps. Keep it low; SVD is trained low-fps.
• augmentation_level (default 0, 0–10) - noise added to the init image. 0 is a strict still; a small value can help short videos look less frozen.
• OUTPUT_positive_key / OUTPUT_negative_key (defaults positive / negative) - the conditioning output keys.
• OUTPUT_latent_key (default latent) - the video latent output key.

Output is the MOTORWAY with positive, negative, and latent stored, ready to feed a KSampler.

Installing it

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

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

Where people get burned

SVD is picky about the image. It wants a clean, reasonably sharp 1024×576 input - feed it something smaller and the model upscales it badly, giving you wobbly video. And unlike a lot of models, SVD's fps isn't just metadata; generating at 24fps when it was trained at ~6-14 produces weird temporal artifacts.

The Motorway trap is the usual one: keys are case-sensitive and exact, and 'MotorwayClass' object has no attribute 'hash_' means a key isn't in the pipe - make sure clip_vision, init_image, and vae were written upstream. Beta pack, motorway-ed nodes toggled across releases - a node vanishing after an update is upstream, not a broken install.