SVD_img2vid_Conditioning

SVD_img2vid_Conditioning is the conditioning node for Stable Video Diffusion - Stability's late-2023 image-to-video model, the last significant open release before its research bench walked out. It's old enough that most current video work has moved on to Wan and LTXV, but you'll still meet it: it ships in ComfyUI core, it's in a mountain of archived workflows, and it's genuinely instructive because it's the cleanest example of how image-to-video conditioning works in this whole family.

SVD takes one image and animates it. This node does the entire front half: it encodes the image with CLIP vision for semantics, VAE-encodes it as the starting frame, and packages both into positive and negative conditioning plus the empty latent. Everything you see in modern video conditioning nodes - concat latent images, clip vision outputs, motion parameters - is here first, in the most transparent form.

How it works

Under the hood it runs the init image through CLIP vision to get pooled embeddings, VAE-encodes it as the concat_latent_image, then builds the conditioning pair. The negative side gets zeroed pooled output and a zero latent - the standard "unconditional" counterpart.

The inputs that matter:

• clip_vision / init_image / vae - the vision model, your starting frame, and the VAE. The triple that powers every image-to-video node.
• width / height - defaults 1024×576 (16:9). SVD's native territory.
• video_frames - default 14. SVD's native length is short; 14–25 frames is its comfort zone.
• motion_bucket_id - 1 to 1023, default 127. SVD's motion dial. Higher = more motion, but push it too far and motion turns to warping. This is the one parameter people actually tune.
• fps - default 6, for interpretation/timing.
• augmentation_level - default 0. Adds noise to the input image before encoding; higher values make the model "see" the starting frame more loosely, which can reduce the frozen-first-frame look but costs fidelity.

Outputs: positive, negative, and latent - the last being a zeros latent sized to your frames, which the KSampler fills.

What people actually hit

The default motion quality. SVD's famous weaknesses - slow pans, occasional warping, a first frame that can look pasted on - are all motion_bucket_id and augmentation_level territory. Start at the defaults, then nudge augmentation up toward 0.3–0.5 if your first frame looks too frozen.

It's superseded, and that's fine. The honest verdict: Wan and LTXV leave SVD's quality in the dust and run locally on comparable hardware. Keep SVD in your toolkit for what it's still genuinely good at - the "living photo" effect and the depth-parallax workflows that the depth-estimation crowd still runs - not as your main video engine.

Model files. SVD checkpoints go in the models folder like any other; it needs its own VAE. Old workflows usually reference them by exact name, so if a workflow won't load, the missing file is the first thing to check.

Ships with ComfyUI core. It's the "read this to understand every other i2v node" of the group.