Video
How to fix flicker and morphing in AI video

Key takeaways
Flicker is fast change from one frame to the next in brightness or texture. Morphing is slower change in shape, where hands, faces, and logos shift as the clip runs.
Most of both starts at generation. High motion settings, long clips, prompts that pack in several actions, and soft or cluttered source images all give the model room to redraw things between frames.
Fix the source first. Shorter clips, a sharp start image, first and last frames, one simple motion per shot, and a fixed seed for retakes do more than any filter.
In the edit, deflicker filters even out brightness, frame interpolation smooths stutter, and stabilization removes camera jitter. None of them can redraw a hand that grew a finger, so cut around those frames.
A clip that left the model clean can still pick up flicker from a frame-rate mismatch, frame-by-frame upscaling, or a low-bitrate export.
How do you fix flicker and morphing in AI video?
Regenerate the shot shorter and simpler, from a sharp start image, then clean up what's left in the edit with a deflicker filter, frame interpolation, or stabilization, and cut away from any frames that still morph. Generation settings cause most of the problem, so changes there go further than anything you can do afterward.
Work out which problem you have before you reach for a fix. A filter that evens out brightness does nothing for a face that slowly becomes someone else, and a fresh generation is wasted if the flicker came from your export.
Flicker and morphing aren't the same problem
Flicker is fast and local. Brightness pulses, skin texture crawls, and fine patterns like hair, fabric, or leaves shimmer even when the subject holds still. VBench, a benchmark suite for video generation models, scores it as its own dimension, temporal flickering, measured on clips with no real motion so that any frame-to-frame change counts against the model.
Morphing is slower and structural. A hand gains a finger, a logo rewrites itself, a jacket's buttons wander, or a face drifts toward a different person over a few seconds. It's the video form of identity drift, where each small error becomes the starting point for the next frame.
Both have the same root. The model decides the content of every frame, and anything it isn't held to can change. The VBench authors found a tradeoff between steadiness and movement across the models they scored, noting that "Models strong in temporal consistency often have a lower Dynamic Degree." Put simply, the clips that move the most tend to be the ones that flicker and morph, and the steadiest clips often barely move.
What causes flicker and morphing
Start with motion strength. Many models expose a setting for how much movement to generate. In Stable Video Diffusion, the diffusers guide says raising "motion_bucket_id" "increases the motion of the generated video," and that a higher "noise_aug_strength" makes the clip resemble the source image less while adding motion. More motion gives the model more license to redraw.
Clip length counts too. Long clips give small errors time to stack up, and models that build a clip in segments condition each new segment on frames they already made, so mistakes carry forward. The FramePack paper by Lvmin Zhang and colleagues proposes "drift prevention methods to address observation bias (error accumulation)" in this kind of model.
Prompts that ask for too much cause their own trouble. Google's Veo best-practices page warns that "Trying to chain multiple distinct events (A then B then C) in one prompt for a short video often leads to muddled or incomplete videos." Three moving subjects, a camera move, and a lighting change in five seconds is a lot for a model to hold steady.
In image-to-video, the start frame sets everything that follows. The same Google page says "A sharp, clear, and well-composed photograph yields a more coherent and higher-quality video." A soft, compressed, or cluttered still leaves the model guessing at detail, and it can guess a little differently on each frame. Lace, small text, chain-link fences, and crowds give it the most to guess at.
Some local settings trade quality for memory. The same diffusers guide says that decoding one frame at a time, with "decode_chunk_size" set to 1, uses the least memory, "but the video might have some flickering."
Frame rates vary from model to model. The diffusers Wan documentation saves Wan 2.1's 81-frame clips at 16 fps, about five seconds, while the Wan 2.2 repository lists its 5B model at 24 fps. Put a 16 fps clip on a 30 fps timeline and the editor has to repeat frames unevenly, since 30 isn't a multiple of 16, so motion stutters in a regular rhythm that's easy to mistake for flicker.
Upscaling can add flicker after the fact. A still-image upscaler run on each frame separately invents fresh detail per frame, with no memory of the frame before, so textures can crawl. Diffusion upscalers are the hardest case. The Upscale-A-Video paper says temporal consistency in video super-resolution "is complicated by the inherent randomness in diffusion models," and research methods such as StableVSR add a temporal module to a single-image model to deal with it.
Then there's compression. Grain, noise, and fine texture take a lot of bitrate to reproduce, so at low bitrates the encoder throws some of that detail away, and not the same detail on every frame. If a clip looked clean in the generator's preview and shimmers after upload, check the export before you regenerate. Our guide to export settings for Reels, TikTok, and Shorts covers the numbers.
Symptom, cause, and fix at a glance
Symptom | Likely cause | First fix |
|---|---|---|
Brightness pulses on a still subject | Frame-to-frame luminance changes from the model or the decoder | Deflicker filter, then check decode settings |
Hair, fabric, or leaves shimmer | Soft or busy source image, frame-by-frame upscaling, or low bitrate | Sharper start image, a video upscaler, a higher export bitrate |
Hands, faces, or logos change shape mid-clip | Too much motion, or a clip that runs too long | Lower motion, a shorter clip, a cut around the bad frames |
The character looks different by the last second | Errors building up over the clip | Shorter clips, each started from the same approved still |
Motion stutters in a regular rhythm | Clip frame rate doesn't match the timeline | Edit at the clip's native rate, or interpolate up |
The whole frame jitters | An unsteady virtual camera | Stabilization |
The action falls apart halfway through | A prompt with several events | One action per clip, joined in the edit |
Fixes during generation
Make shorter clips. Generate a shot in short pieces and join them in the edit. Short shots are normal in edited video, and each new clip starts fresh, so it doesn't inherit the last one's drift. Start every piece from the same approved still. Chaining each clip from the last frame of the one before repeats the error build-up you were trying to avoid.
Give the model a clean start. Use a sharp, well-lit still with an uncluttered background, at or above your output resolution. If the model accepts reference images, they tell it what a face or product looks like without fixing the composition, which helps a recurring character hold across several shots.
If you know where the shot should finish, give the model a first and a last frame so the subject has somewhere fixed to land. Google's Veo documentation on Vertex AI includes a guide to generating from first and last frames, and ComfyUI publishes a Wan 2.2 first-last-frame workflow for local use. Keep the two frames close in pose and framing. Our first and last frame explainer goes into how the method works.
Simplify the motion. Ask for one subject, one action, and one camera move per clip. Lower the motion setting if your model has one, and describe slow, continuous movement ("she turns her head toward the window") instead of fast choreography. A locked-off or slow camera also leaves less background for the model to redraw.
Seeds help with retakes. A seed fixes the starting noise, so you can bring back a take that mostly worked and change one thing at a time. Google's Veo guide says to "use the same seed parameter" for consistent output across scenes. A seed won't stop flicker inside a clip, but it keeps you from chasing a moving target while you adjust motion or wording.
Render at your delivery rate. Where a model lets you pick the frame rate, choose the one you'll edit and publish at, so nothing downstream has to invent or drop frames.
Fixes in the edit
Post tools work on pixels. They can even out brightness, fill in frames, and steady the camera, but they can't tell that a hand has six fingers.
Deflicker filters: FFmpeg's deflicker filter is documented to "Remove temporal frame luminance variations," averaging brightness over a moving window of frames, 5 by default and up to 129. Blackmagic lists de-flickering among the image restoration tools in DaVinci Resolve Studio, the paid edition. FFmpeg's filter works on brightness alone, so texture crawl and shape changes get through it.
Frame interpolation: RIFE, an MIT-licensed model presented at ECCV 2022, generates in-between frames and can multiply a clip's frame count by two, four, or more. FFmpeg's minterpolate filter does motion-compensated interpolation without a separate model, and by default it swaps in duplicate frames at detected scene changes. Interpolation fixes stutter from a low native frame rate, but it builds new frames from the ones you have, so any morph comes along with them.
Stabilization: FFmpeg's vid.stab filters run in two passes, vidstabdetect to measure the motion and vidstabtransform to correct it, and they need an FFmpeg build compiled with libvidstab. The simpler deshake filter handles small horizontal and vertical shifts. Both steady a jittery camera, not a subject that wobbles inside a steady frame.
Cutting around bad frames: When a morph lasts only a few frames, trim it, cut to another angle, or cover it with a cutaway and come back. A well-placed cut hides more than any filter.
If you stack these, conform the frame rate first, then stabilize and deflicker, and interpolate last, so the interpolator builds from the cleanest frames you have. Then export at a bitrate high enough to keep the texture you just fixed.
Flicker and morphing FAQ
Why does my AI video flicker?
The model is redrawing brightness and fine texture slightly differently on each frame. High motion settings, soft or busy source images, memory-saving decode settings, frame-by-frame upscaling, and low-bitrate exports all make it worse.
Why do faces and hands morph in AI video?
Small errors build on each other as the clip runs, especially in long clips with a lot of motion. Hands, faces, and text have precise structure, so any drift there is easy to see. Shorter clips started from the same approved still, plus first and last frames, keep them closer to the original.
Can a deflicker filter fix morphing?
No. FFmpeg's deflicker filter, for one, averages out brightness changes between frames. It doesn't track shapes, so a finger that appears or a logo that rewrites itself passes straight through. Regenerate the shot or cut around those frames.
Does frame interpolation reduce flicker?
It reduces stutter, which can look like flicker, especially in clips generated at 16 fps. It doesn't remove real flicker or morphing, because the new frames are built from the existing ones, problems included.
Should I make one long clip or join shorter ones?
Join shorter ones. Each short clip starts fresh from your approved still, so drift doesn't have time to build, and the cuts between them read as ordinary editing.
