Stable Video Diffusion guide
Stable Video Diffusion (SVD) turns a single still image into a short clip. The output hinges on a handful of parameters — motion bucket ID, augmentation strength, frame count and FPS. Pick your desired motion level and subject and the guide recommends a starting configuration plus an explanation of each control.
What each parameter does
Motion bucket ID (1–255) is the headline motion dial. It was set during training by binning video clips by measured optical flow — clips with low motion got low IDs, clips with dramatic camera or subject movement got high IDs. At inference, low values produce subtle, stable motion (hair gently moving, water rippling); high values produce dramatic movement but increase the risk of warping and limb flicker.
Augmentation (noise augmentation) strength (0.0–0.3) controls how much the model perturbs the conditioning image before using it to guide generation. Near zero it stays very faithful to the source frame; higher values introduce more creative interpretation and allow larger motion arcs but can cause the subject to drift from the original image identity.
Frame count (14 or 25) is fixed by the checkpoint you load, not a parameter you set at inference. SVD-XT (the 25-frame variant) produces longer clips and is better for smooth continuous motion; the 14-frame model generates faster and works well for short loops.
FPS (6–30) is a micro-conditioning hint to the model. It affects how the model distributes motion across frames, not the final playback speed. Lower FPS hints produce slower, more deliberate motion for the same number of frames; higher FPS hints produce smoother but more compressed motion. Most renders use 6–12 FPS for the clip and then optionally interpolate to 24 FPS for smooth playback.
Recommended starting configurations
| Subject type | Motion bucket ID | Augmentation | Notes |
|---|---|---|---|
| Portrait / face | 50–90 | 0.0–0.02 | Keep low; face warping is very noticeable |
| Product / still life | 60–100 | 0.0–0.03 | Subtle orbiting motion or shine effects |
| Nature / landscape | 100–160 | 0.02–0.05 | Tolerate more motion; clouds and water benefit |
| Abstract / texture | 150–220 | 0.05–0.15 | High motion works when there is no rigid subject to warp |
Diagnosing bad output
Warping or smearing: motion bucket ID is too high or augmentation too strong for the subject. Lower both and try again.
No visible motion: motion bucket ID is too low (below ~40) or augmentation is exactly 0.0. Raise motion bucket to at least 60.
Subject identity drift: augmentation strength is too high. Reduce to 0.0–0.02 for portraits or objects that must remain recognisable.
Flickering in the background: common with busy, high-frequency textures in the background of the source image. Simplify the source image or reduce motion bucket.
Source image tips
SVD was trained on natural photographs rather than synthetic images, so it produces cleanest results from photographic inputs:
- Use a clear subject against a relatively uncluttered background.
- Avoid text, logos, or fine grid patterns in the image — these warp noticeably.
- Square or landscape-orientation images at 1024×576 or 1024×1024 are natively supported; other ratios require cropping or padding.
- A well-lit, sharp source image generates more stable motion than a blurry or high-contrast one.
Seed variance
SVD motion varies substantially between seeds at the same settings. Generate 3–5 variations and pick the cleanest before concluding that settings need adjusting. What looks like a parameter problem is often just seed variance.