AI Video Artifacts: A Troubleshooting Checklist

A clip that looks good in its first frame can still fail in motion. Use this checklist to isolate flicker, changing faces, distorted products and unwanted camera movement before spending credits on another attempt. These are diagnostic starting points, not guaranteed fixes or benchmark results.

Inspect before regenerating

Play the whole clip, then pause at the start, middle and end. Record the first time the problem appears. Compare the input image with the output: if the flaw already exists in the source, repair or replace the source before adjusting motion. Confirm the chosen duration, aspect ratio and resolution in the studio; writing different settings in a prompt does not change the controls.

Match the symptom to one experiment

Symptom	Check first	Next experiment	Accept only if
Face changes during a turn	Blurry face, extreme pose or occlusion in the source	Keep the same image and reduce the turn to a small glance	Identity stays recognizable throughout
Product shape bends	Tiny edges, complex reflections, competing actions	Try a stationary product with one slow camera move	Silhouette and required details remain accurate
Background or lighting flickers	Busy texture, shifting shadows, conflicting lighting instructions	Use a simpler scene and steady lighting	No distracting changes across adjacent frames
Camera moves when it should hold	Motion cues in the image or several camera commands	Ask for one continuous stationary shot	Framing stays within the intended composition
Movement feels rushed	Several actions squeezed into the chosen duration	Split the sequence into separate shots	Each action is readable at normal playback speed
Text or logo becomes unreadable	Small lettering in the source	Add verified text and logos in your editor	Final exported lettering is correct

Changing one variable makes a comparison easier to interpret. Repeating the same request can produce variation, but a single better clip does not prove the prompt caused the improvement.

Rewrite a crowded prompt

Crowded draft: “The bottle rotates, the camera circles, water splashes, a hand picks it up, the logo stays sharp, dramatic changing lights.”

First diagnostic draft:

One continuous shot. The bottle remains stationary on the table.
The camera slowly moves closer. Soft studio lighting stays steady.

This original example reduces simultaneous demands. It has not been tested as a model comparison. If rotation is essential, test that action separately before adding a camera move. For composition and action planning, use the AI video prompt guide.

Keep a small experiment log

Attempt	Source	Model and selected settings	One change	First flawed time	Decision
A	Same reference file	Record from studio	Baseline	Record timestamp	Keep / revise
B	Same reference file	Same settings	Reduce subject movement	Record timestamp	Keep / revise

Set a credit ceiling before starting. Record the credit amount displayed for each selected configuration and calculate total credits spent divided by accepted clips. A technically completed clip can still be unusable for your project. See the pricing and accepted-clip cost guide.

Stop, edit or change the input

Stop repeating a setup when the same important defect persists within your budget. Consider another reference image, a simpler shot or a different supported model. Trimming a flawed ending can work only if the remaining footage still communicates the intended action. A higher output resolution is not a dependable repair for incorrect anatomy, identity or product geometry.

For commercial product footage, review the entire clip against the real product. For recurring people, use the character continuity guide. Export, then check the finished edit again: captions, crops and overlays can introduce new problems.

Choose the right repair before another generation

Decision	Appropriate when	Check before delivery
Edit the clip	A flawed tail can be removed without losing the intended action	The new ending and adjoining cut still make sense
Stabilize in an editor	Unwanted camera shake is the problem	The crop does not cut off the subject; geometry is still correct
Replace the starting image	The defect is visible before motion begins	Face, hands, edges and required product details are already correct
Regenerate one simplified action	A required action fails during motion	Complete playback plus inspection around the recorded defect timestamp

Stabilization addresses camera motion, not a different face or a warped object. A still image can be more useful than repeated video attempts when the required product detail cannot be verified. Record whether the next step changes the source, the prompt or the edit, so the result answers a specific question.

Inspect at the actual delivery size

Review the downloaded file and the exported edit, not only the small browser preview. Check at the intended display size and normal playback speed, then pause around the defect. Inspect text after captions and crops have been applied. A clip accepted in landscape may lose its required detail in a vertical crop; use the vertical video guide to plan framing before another generation.

What is sourced and what is a proposed workflow?

Runway’s official image-to-video prompting guide explains that the image establishes visual information and the prompt describes motion; it also warns that source-image artifacts can carry into motion. That advice is model-specific. The diagnostic table and experiment log above are VidVid editorial suggestions for organizing your own checks, not promises that all models follow the same behavior.

Next step

Choose one defect and one experiment. Open the VidVid studio, confirm the available controls and displayed cost, and inspect the complete result before trying another change.

Before another attempt, use the reference image preparation guide to check source defects and calculate your delivery crop.