Every AI video generator can produce an impressive paused frame. Here’s the uncomfortable truth, though: professional delivery is not about the impressive frames. It’s about the hundred frames nobody notices — the clean handoffs, the consistent faces, the physics that never draw attention. Because one bad frame at the wrong moment can cost a client’s trust. And trust? Trust is the only currency a freelancer has.
So quality control has to be a gate, not an afterthought. A gate is something every project passes through before delivery. Every time. This checklist is that gate. Run it in order and fix what fails before the file leaves your hands. Your reputation will thank you.
The QC Mindset: Catch Everything Before the Client Sees It
Think of QC as the final stage of production, not a safety net. Safety nets catch you after you fall; a production stage prevents the fall. Clients notice patterns faster than individual mistakes: one flickering face is a fluke, but flicker in three consecutive deliverables is your standard of work.
Build QC into your quotes — yes, charge for it. A five-minute clip needs twenty to thirty minutes of checking across the steps below; complex multi-character projects need more. And watch each shot once at half speed before signing off on it. Half speed. It’s boring, it’s slow, and it exposes flicker, shimmer, and morphing that full-speed playback hides. Save full-speed viewing for the final watch-through ritual in section nine. There’s a reason it’s last.
Visual Artifact Checks
AI-generated video fails visually in predictable ways. Predictable is good news — it means learnable. Once you know the failure modes, you can spot them in seconds.
Flicker and Temporal Shimmer
What it looks like: Fine detail — fabric texture, hair strands, foliage, chain-link fences — appears to vibrate or crawl between frames. Surfaces seem to shimmer like heat haze, even in static shots.
How to spot it: Pause on fine detail and step forward one frame at a time. If the texture rearranges itself rather than holding still, you have shimmer. Frame-stepping is your best friend here.
Severity: Fix. Shimmer reads as “something is wrong” even to non-technical viewers. They can’t name it, but they feel it. Regenerate with a simpler prompt — heavy texture detail in the wording tends to cause it.
Morphing Faces
What it looks like: A character’s face subtly changes across a shot — the nose shifts shape, eyes move relative to each other, or the whole face drifts toward a different person by the end of the clip.
How to spot it: Scrub slowly with attention locked on one face. Compare the first, middle, and last frames side by side — small changes compound over a long take. The drift you miss at full speed screams at you in a side-by-side.
Severity: Regenerate if the drift shows in the first-to-last comparison. Acceptable if it is detectable only frame-by-frame — short social clips get more slack than long-form client work. Detailed, consistent character descriptions reduce drift; see the anatomy of a great AI video prompt.
Extra, Melted, or Missing Fingers and Limbs
What it looks like: Hands with six fingers, fingers that merge into each other like wax, a forearm that dissolves into a sleeve, legs that bend at impossible joints, or limbs that appear and vanish behind objects. The classics. The memes write themselves.
How to spot it: Every time a hand is prominent — gesturing, holding something, close to the face — stop and count fingers. Check limb continuity wherever a limb crosses behind an object and re-emerges.
Severity: Regenerate. There is no acceptable level of malformed hands in professional delivery. None. If a shot keeps failing, reframe so hands are out of shot or in motion blur — a deliberate framing choice beats a deformed close-up every time.
Texture Swimming
What it looks like: Related to shimmer but larger in scale — entire surfaces (brick walls, tiled floors, patterned clothing) appear to flow or undulate as if seen through water, even when the camera is locked.
How to spot it: Watch static shots with patterned surfaces at half speed. If the pattern moves while nothing else does, that is swimming. The camera’s locked; the wall shouldn’t be dancing.
Severity: Fix by regenerating or replacing the shot. Acceptable only in fast motion where the audience cannot register the surface. Fast motion hides a multitude of sins.
Background Warping
What it looks like: Straight lines — doorframes, window edges, table edges, horizon lines — bend, wobble, or breathe as the shot plays. The background behind a moving character can ripple like it is reacting to them. Spooky. Not in a good way.
How to spot it: Trace the edges of architectural elements with your eye while the shot plays. Warping shows up most in shots with camera movement, where the generator struggles to keep geometry stable. Architecture is the lie detector of AI video.
Severity: Fix. Viewers subconsciously check architecture for reality — warping doorframes read as cheap instantly. Regenerate with a more locked-down camera or a simpler background description.
Object Pop-In and Pop-Out
What it looks like: An object — a cup, a bag, a piece of furniture — appears or disappears between frames without the camera moving past it or a character interacting with it. Now you see it, now you don’t. Magic, but not the fun kind.
How to spot it: This one hides at edit points and during character motion — the busy moments where your eye is elsewhere. Watch each shot’s start and end frames against the frames around them, and watch any moment a character’s body crosses an object.
Severity: Fix by regenerating or trimming — a pop in the first three frames can sometimes be fixed by trimming the head of the shot. Acceptable only in a blink-and-miss area of a fast-moving shot.
Lip-Sync Drift
What it looks like: The mouth moves but does not match the dialogue timing — syllables land while the mouth is still, or the mouth keeps moving after speech stops. Applies to AI lip-sync tools as well as natively generated talking heads. Uncanny valley territory.
How to spot it: Watch dialogue shots with the sound on at half speed, focusing on consonant sounds — P, B, and M need visible mouth closures. If closures land late or early by more than a frame or two, there is drift. A frame or two is the tolerance. Not ten.
Severity: Fix. Bad lip sync breaks the illusion faster than almost any other artifact — the brain is ruthless about mouths. Regenerate with a cleaner audio file, or cut dialogue into shorter segments and sync each separately.
Eye Direction Glitches
What it looks like: Eyes that drift to look at something off-camera for a few frames, pupils that shift independently, or a character who seems to glance at the camera when they should not. Small thing. Huge creep factor.
How to spot it: Scrub any shot where a character holds eye contact or looks at a target. Watch the pupils, not the head — the head can stay still while eyes wander. The pupils betray everything.
Severity: Regenerate for close-ups and hero shots. Acceptable in wide shots where the face is small in frame. Distance is forgiving; close-ups are not.
Motion and Physics Checks
AI movement looks right until it does not. Your job is to catch the moments where the illusion breaks — and there will be moments. There always are.
Unnatural movement: Watch how characters walk, turn, and gesture. Common failures: feet sliding instead of planting, arms swinging at the wrong rhythm, turns where the body rotates but the head lags behind. Scrub these moments frame by frame and regenerate anything that looks like the character is on ice. Ice-skating through a dialogue scene? Regenerate.
Gravity breaks: Objects falling too slowly or too fast, hair and clothing ignoring wind direction, liquids pouring in impossible streams. If a dropped object does not accelerate convincingly, the shot reads as weightless. Weightlessness reads as fake. The physics has to feel right even when viewers can’t articulate why.
Impossible object interactions: A hand that passes through a cup instead of gripping it. A character who sits on a chair but sinks through the seat. A door that opens without the handle moving. Check every interaction point: wherever two things should touch, verify they actually meet in space. Touch is where AI video most often lies.
Camera move smoothness: AI camera moves sometimes stutter or change speed mid-move — a little hitch you feel more than see. Watch at full speed with your attention on the background — background motion reveals stutter that subject tracking hides. A move that ramps oddly should be regenerated or replaced with a static shot. A clean static beats a queasy move.
Audio Checks
Video gets the attention, but audio is what clients actually judge first. Here’s a hard truth: bad audio makes good video feel cheap; good audio makes decent video feel professional. Clients may not know video, but everyone knows when something sounds wrong.
Dialogue sync: Beyond lip-sync drift, check that dialogue starts and ends cleanly. No clipped first syllables, no trailing noise after the last word. Listen to every line with headphones. Every line. Tedious? Yes. Necessary? Also yes.
Music levels versus voice (ducking): Background music must sit under dialogue, not beside it. Lower music by ear during every spoken section — roughly one-quarter of dialogue volume is a reliable starting point. Check on both headphones and a phone speaker; phone speakers reveal masking that headphones hide. Your audience is on phones. Mix for phones.
SFX placement: Every sound effect must land on its visual event — footsteps on foot plants, door sounds on door motion, whooshes on fast moves. A misplaced SFX is worse than none — it draws attention to the edit. Scrub to each event and confirm alignment. Each one.
Silence gaps: Check for unintended dead air — sections where dialogue, music, and ambience all drop out. Some gaps are intentional; unintentional ones feel like the export broke. If the viewer wonders “did my audio cut out,” you’ve failed this check.
Export audio specs: Confirm the delivered file has the audio the client specified. The professional default is 48kHz sample rate, stereo, at a healthy level — peaks around -3dB to -6dB for web delivery, never hitting 0dB (clipping). Verify the spec against the brief, not habit. The brief is the contract; habit is just habit.
Licensed versus AI-generated music documentation: Know the source of every track and sound effect. Licensed: keep the license or receipt on file. AI-generated: note which tool produced it and its terms. Document this in the delivery package — clients get asked about music rights more often than expected. Keep a simple log: track name, source, license type, date. Boring paperwork that saves careers.
Brand Consistency Checks
Brand rules are not suggestions. One wrong color in a deliverable tells the client you did not read their guidelines. And clients notice. They always notice the brand stuff — it’s their baby.
Colors and hex match: Sample brand colors from the rendered export, not the project file — compression can shift colors. Compare against the client’s hex values with a color picker, including gradients and background tints. The export is the truth; the project file is a hope.
Fonts: Every on-screen word must use the approved typeface at approved weights — captions, lower thirds, end cards, thumbnails. One substituted font in an end card is a full revision round. One font. Full round. Check them all.
Logo usage and clear space: The logo must appear at the approved size and position with the required clear space. Never stretch it, recolor it, or place it over a background that kills legibility. Check every placement against the client’s usage sheet. Logos are sacred; treat them that way.
Tone of captions: Read every caption out loud. Does it sound like the brand? A playful brand with stiff captions — or a serious brand with slang — signals the video was made without understanding the client. Reading aloud catches what silent reading misses.
Approved terminology: Check product names, slogans, and terms exactly against the client’s term list — capitalization, spacing, trademark symbols. A misspelled product name in a delivered video is a defect, not a typo. Defects get remembered.
Legal and Compliance Checks
This section protects you from consequences far more expensive than a revision round. Lawsuits cost more than re-renders. Rules change over time — always verify current policies before delivery. Always.
Likeness rights: Do not generate recognizable real people — celebrities, public figures, or private individuals — without proper consent and legal clearance. This covers faces, voices, and distinctive personas, and even “inspired by” versions carry risk. When in doubt, the answer is no. “Inspired by” is not a legal shield.
Music and SFX licenses documented: Every asset’s source and license status must be documented — royalty-free is not copyright-free, and “found it online” is not a license. Keep receipts or screenshots of terms with the project folder. Paperwork again. It matters again.
Platform AI-content labels: Major platforms — TikTok, YouTube, Instagram among them — require or encourage disclosure labels on realistic AI-generated content, and their specific rules change regularly. Before publishing, check the current policy on each platform, apply the required labels through the platform’s own tools, and confirm this with the client so the responsibility is shared. Do not assume last month’s rule still applies. It probably doesn’t.
FTC disclosure for sponsored content: In the United States, paid or sponsored content requires clear disclosure of the material connection between creator and advertiser. The general principle: clear, conspicuous, and placed where the audience will actually see it — not buried. Requirements evolve, so check the FTC’s current guidance on sponsored projects and put the disclosure decision in writing with the client before delivery. In writing. Before.
Technical Spec Checks
Technical specs are the least glamorous part of QC and the most common source of rejections. A perfect video still fails in the wrong format. Nobody applauds correct specs — but everyone notices wrong ones.
Resolution matches the delivery spec: Confirm the export resolution against the brief — 1080×1920 for vertical, 1920×1080 for horizontal, or whatever was specified. Also confirm no accidental pillarboxing or letterboxing from mismatched sequence settings. Stray black bars scream amateur.
Codec: Deliver the codec the client asked for. For most web and social delivery, H.264 in an MP4 container remains the safe default. If the client specified ProRes, DNx, or another mezzanine codec, verify the export actually used it — a renamed file is not a codec change. Renaming isn’t encoding.
Bitrate: Match the brief’s target when one is given; otherwise use a sane default for the resolution and note it in the delivery. Too low and motion falls apart; too high and the file may exceed upload limits. Somewhere sane in the middle.
Frame rate consistency: Every shot in one project must share the same frame rate. Mixed frame rates — a 24fps shot sitting in a 30fps timeline, or AI clips generated at different rates — cause judder, duplicated frames, or audio drift. Check the timeline settings and every source clip before export. This is one of the most common defects in AI video projects because generators often output varying frame rates. Check. Every. Clip.
File naming: Follow the client’s convention; if none was given, use project name, version, resolution, date. Never deliver “final_final_v2_REAL.mp4.” Version discipline in the filename prevents the client from publishing the wrong cut. We’ve all been tempted. Resist.
No black frames at head or tail: Scrub the first and last ten frames of the export. Unintended black frames look like a broken render on social platforms. Trim them unless the black is intentional — and if it is, make it long enough to read as deliberate. A deliberate pause; not a glitch.
Client-Specific Checks
Generic quality is not enough. The video must be the video the client asked for — not a beautiful video, the right video.
Brief compliance: Re-read the written brief — not your memory of it — and check every requirement: length, format, aspect ratio, number of versions, key messages, mandatory shots or lines. Mark each as met. Anything unmet gets fixed before delivery, not explained after. “I thought you meant…” is not a defense.
Revision notes all addressed: Go through every revision note from the last round and confirm each is resolved in the export. Keep a running log where each note gets a status: done, partially done, or intentionally not done with a reason. Notes “addressed” in the project file but lost in the export are not addressed. The export is the only thing the client sees.
Correct version number: Verify the delivered file is the current version with the latest revisions applied. Check the version in the filename, any slate or metadata, and your project log. Delivering v3 when v4 exists is a preventable embarrassment clients remember. They remember. Forever.
The Final Watch-Through Ritual
After every check above is done, perform one final ritual: experience the deliverable exactly as the audience will. Not as the editor. As the audience.
Full screen, with sound, no scrubbing: Watch the entire video from start to finish without touching the timeline. No pausing, no scrubbing back, no skipping. Scrubbing is an editor’s behavior; the audience never scrubs. If something feels wrong during a continuous watch, note the timestamp and fix it. Trust the gut feeling — it’s usually right.
On the target device: If the video is for vertical social, watch it on a phone. Desktop monitors forgive what phone screens expose — small text, muddy contrast, audio that only works on good speakers. The target device is the truth. Everything else is a preview of a lie.
The fresh-eyes rule: Step away before the final watch — fifteen minutes minimum, ideally an hour or overnight. Your brain stops seeing defects it has already decided are fine. Fresh eyes catch what tired eyes negotiated away. Build this break into your delivery schedule, not around it. It’s not wasted time; it’s the cheapest insurance you own.
If the video survives the fresh-eyes watch with no notes, it is ready. Ship it. For broader context on structuring prompts that survive QC in the first place, the AI video prompting system and the 2026 generator roundup cover the workflows that produce cleaner source material.
The AI Video QC Checklist — Summary
Print this section, keep it beside your desk, and run it on every project:
- ☐ Watched every shot at half speed for flicker and temporal shimmer — regenerated failures
- ☐ Compared first, middle, and last frames of every face for morphing
- ☐ Counted fingers and checked limb continuity on every prominent hand and limb
- ☐ Checked patterned surfaces in static shots for texture swimming
- ☐ Traced architectural edges for background warping during camera moves
- ☐ Verified no object pop-in or pop-out at shot starts, ends, and interaction points
- ☐ Watched all dialogue at half speed with sound for lip-sync drift
- ☐ Scrubbed eye contact and gaze targets for eye direction glitches
- ☐ Checked walking, turning, and gestures for foot sliding and unnatural movement
- ☐ Verified gravity, falling objects, hair, clothing, and liquid behavior look weighted
- ☐ Confirmed every object interaction — grips, sits, door handles — actually connects
- ☐ Watched camera moves at full speed with attention on the background for stutter
- ☐ Listened to all dialogue on headphones — no clipped syllables or trailing noise
- ☐ Set music levels under dialogue (ducking) — checked on headphones AND phone speaker
- ☐ Aligned every SFX to its visual event
- ☐ Listened through the full timeline for unintended silence gaps
- ☐ Confirmed export audio spec (48kHz, stereo, healthy levels, no clipping)
- ☐ Logged every music track and SFX: source, license type, date
- ☐ Sampled brand colors from the rendered export against client hex values
- ☐ Confirmed all on-screen text uses approved fonts and weights
- ☐ Checked logo size, position, clear space, and legibility against usage guidelines
- ☐ Read all captions aloud for brand tone match
- ☐ Verified product names, slogans, and terminology exactly against the term list
- ☐ Confirmed no recognizable real people generated without consent
- ☐ Documented all music and SFX licenses with the project files
- ☐ Checked current platform AI-label policies and applied required disclosures
- ☐ Confirmed FTC sponsorship disclosure (clear, conspicuous, agreed with client in writing)
- ☐ Confirmed export resolution matches the delivery spec, no stray bars
- ☐ Verified codec and container against the brief
- ☐ Checked bitrate against target or platform-appropriate default
- ☐ Confirmed consistent frame rate across all shots in the project
- ☐ Named the file per client convention with the correct version number
- ☐ Scrubbed first and last ten frames — no unintended black frames
- ☐ Re-read the written brief and marked every requirement as met
- ☐ Confirmed every revision note from the last round is resolved in the export
- ☐ Verified the delivered file is the current version, not an older cut
- ☐ Final watch-through: full screen, with sound, no scrubbing, on the target device
- ☐ Applied the fresh-eyes rule — stepped away, then watched once more