You generated a clip and the motion is not smooth — limbs pop between frames, the subject seems to teleport a few pixels every fifth frame, body parts judder at hand-off moments. This is “motion jitter,” and it is distinct from background flicker (which is environmental) and drift (which is identity loss). Jitter is the motion path itself failing to interpolate smoothly. Fix it by reducing motion ambition, providing clearer pose anchors, or rendering at higher fps.
Common causes
Ordered by what trips creators most often.
1. Multiple actions stuffed into one prompt
Running while waving, then turning to look back, then jumping — three distinct actions in a 3-second clip. The model tries to fit all of them and produces a jerky path that abruptly transitions between them with no in-between frames.
How to spot it: Count action verbs in your prompt. Over 1 primary action per 3 seconds and you are in jitter territory.
2. Start pose unspecified
The model picks a random starting pose, then tries to animate toward whatever the prompt implies. The random start often does not align with the prompted action, producing a snap correction in early frames.
How to spot it: Watch frame 1 vs frame 5. If frame 1’s pose makes the prompted action impossible (subject facing wrong way, wrong limb position), the model is fighting its own start.
3. Low generation fps
Some models default to 12 or 16 fps internal generation, then interpolate up to 24fps or 30fps for delivery. Interpolation does not invent missing details — it cross-fades. Result: judder, especially at fast motion moments.
How to spot it: Check the tool’s default fps. Runway is 24fps native; Pika offers 24-30; Kling allows up to 30; Sora varies. If you got 24fps from a 16fps internal generation, jitter is built in.
4. End pose implies a discontinuity
Standing still, then suddenly running — the model has to invent a transition. It often “teleports” the limbs into running pose rather than smoothly accelerating.
5. Multiple subjects with independent motion
Two characters dancing, two cars racing — each subject has its own motion path, and the model often jitters one to keep the other consistent.
6. Camera motion at the same speed as subject motion
Tracking shot at the same speed the subject moves — to the model, the subject appears static while the background flies by. The model jitters the subject trying to “show” it is moving.
7. Motion strength too high for the subject
A clip of a person calmly drinking coffee should be at low motion. Default Runway motion strength is sometimes too aggressive for static-to-mild motion shots and adds unwanted jitter.
Before you change anything
- Save the prompt, model, and the jittery output clip.
- Identify whether jitter is at hand-off moments (start, middle, end) or continuous throughout.
- Note the model, motion settings, and clip length.
- Decide whether the clip needs the specific action sequence, or if a simpler action would serve.
- Commit or back up the prompt template before changing it.
Information to collect
- Full prompt, motion strength, model, version.
- Clip length and aspect ratio.
- Specific timestamps where jitter is worst.
- Whether the jitter pattern is consistent across multiple regenerations (structural prompt issue) or varies (close to random).
- The intended motion type (calm, moderate, fast).
Shortest path to fix
Step 1: One primary action per clip
Rewrite the prompt to feature one verb:
Before:
a woman running through a city, looking back over her shoulder,
then jumping over a fence, motion blurAfter:
a woman running through a city at moderate speed, slight head turn,
smooth continuous motionIf the use case truly requires multiple actions, split into multiple clips and cut between them.
Step 2: Specify the start pose
Add an explicit starting pose to anchor frame 1:
starting in a standing pose, then beginning to run, smooth acceleration
from still to motionFor image-to-video, the reference image is your start pose. Choose a reference image where the pose makes the upcoming motion physically plausible (e.g., subject already leaning forward for a running clip).
Step 3: Use the highest native fps the tool supports
- Runway Gen-3: 24fps native, no fps choice
- Pika 1.5: can request 30fps
- Kling 1.6: can request up to 30fps
- Sora: tier-dependent
Pick the tier that delivers natively at your target fps, not the cheapest tier that interpolates up.
Step 4: Add motion descriptors that match the action speed
For slow motion:
slow continuous movement, fluid motion, soft transitionsFor fast:
sharp athletic motion, motion blur, dynamic accelerationThe descriptor helps the model decide how much temporal smoothing to apply.
Step 5: Use image-to-video for precise control
Text-to-video gives the model freedom to start anywhere. Image-to-video pins frame 1 to your reference. For any clip where the starting pose matters, use image-to-video.
Step 6: Add explicit no-jitter cues
smooth motion, no jitter, continuous frame-to-frame coherence,
no pose discontinuityThis is a soft hint but does measurably help on some models.
Step 7: For multi-subject scenes, separate and composite
Render each subject in its own clip on a simple background, then composite together in post. Trying to make the model coordinate two independent motion paths in a single generation is the highest-risk approach.
How to confirm the fix
- Watch at 25% speed; limbs should move in continuous paths, not snap.
- Pause at every 6th frame; pose should evolve smoothly, not skip.
- Three clips at the same settings should all be smooth, not just one lucky output.
- A teammate watching the clip should not flinch at any motion moment.
If it still fails
- Reduce to a single second of motion. If even 1s jitters, the prompt has structural issues.
- Switch to image-to-video with a strong reference pose.
- Try a different model — Kling has stronger motion coherence in many cases, Runway is stronger on simple subjects.
- Render at the slowest speed your prompt allows, then time-remap in post if you need faster motion.
- Package the prompt, model, output, and jitter timestamps before asking community help.
Prevention
- Storyboard before generating — one action per clip is the rule.
- For multi-action sequences, plan as multiple clips concatenated, not single-shot.
- Default to image-to-video for any clip where the start pose matters.
- Standardize on the model’s native fps; avoid mixing fps within a project.
- Use motion descriptors that match the action’s actual speed.