Matching generated footage to real camera footage

Cut a generated shot into a sequence of real camera footage and the audience finds it before you do. Not because it looks fake in isolation — on its own it may look excellent — but because it sits next to material that was recorded by one physical sensor, through one physical lens, under one set of lights, and it disagrees with all three at once. A shot only has to be wrong on one axis to stick out. It is usually wrong on four.

The four are grain, grade, focal length and motion blur. They are not equally fixable, they do not fail in the same way, and the order you attack them in decides whether you spend twenty minutes or an afternoon. Here is the order, and the reason it goes that way.

The four axes, and why generated shots miss all of them

Grain. Real footage has a noise floor. A sensor at a given ISO produces a specific amount of chroma and luma noise, and that noise is present in every frame of every shot you recorded that day. Generated footage arrives clean — often unnaturally clean, with flat skies and smooth skin that have no grain structure at all. Clean next to grainy reads as plastic, and the eye catches it in shadows first.

Grade. Colour, contrast and black level. Generated clips tend to arrive pre-graded toward whatever the model thinks "cinematic" looks like. That is partly a prompting artefact: mood adjectives such as "cinematic," "moody" or "dramatic" bias the grade rather than the lighting, so you get a look you didn't ask for and no control over the light position you actually wanted.

Focal length. The lens changes the geometry of the picture, not the mood of it. A 24mm shot of a face has a different nose-to-ear relationship than an 85mm shot of the same face, and the background is a different size behind them. Wide lenses separate foreground from background; long lenses compress them. If your real coverage was shot on an 85mm and your generated cutaway looks like a 24mm, the two shots are describing different rooms.

Motion blur. Real cameras have a shutter, and the shutter angle sets how much a moving subject smears in one frame. Generated clips are inconsistent here — some arrive with almost no blur, which reads as staccato, and some smear more than the movement justifies. It's the axis people diagnose worst, because it only shows up when something moves.

Fix the lens first, because it is the only one you cannot fix later

Grain is additive. Grade is a transform. Motion blur can be softened or hidden by a shorter cut. Focal length is geometry baked into the image at generation time, and no downstream work reconstructs it. Reframing a wide shot into a tighter crop does not give you compression; it gives you a cropped wide shot with less resolution.

So the first thing to change when a generated shot sticks out is not a slider. It is the prompt, specifically the lens token, followed by a regeneration.

Lens vocabulary genuinely moves the output. Terms like 35mm anamorphic, 85mm portrait, 24mm wide-angle, macro and shallow depth of field shift field of view and depth of field in current video models rather than acting as vibe words. Name the one your real coverage was shot on:

Medium close-up on an 85mm lens, shallow depth of field, background
falls off softly. A woman in a grey knit turns from the window toward
camera. Kitchen, mid-morning. Window light entering from camera left,
subject at 45 degrees to the source, warm practical lamp behind her
on the counter.

Note what that prompt does not contain: no "cinematic," no "beautiful lighting," no "moody." One concrete lighting descriptor — source, direction, angle — is worth more than five style adjectives, and it leaves the grade alone so you can set it yourself.

Second-order benefit: prose is a bad way to specify a repeatable move. Speed and easing are re-estimated on every generation, so "slow dolly" gives you a different move each time. If you need the same push across three shots, camera control as a parameter beats camera control as an adjective.

The matching order

Work from the mechanical to the cosmetic. Anything you fix out of order, you fix twice.

Conform the format. Frame rate and resolution come first because they are gates, not adjustments. A clip at a different rate to your camera master jolts at every cut no matter how well it's graded. Versely timelines default to 25fps and most models offer 24 — see 24fps clips on a 25fps timeline for the three ways to conform and the one that wrecks lip sync, and why your generated clips don't match in the timeline for the broader version.
Fix the lens. Regenerate with the focal length and depth-of-field behaviour named. Do not proceed until the geometry agrees, because every step below sits on top of it.
Fix the motion. If the blur is wrong, the usual cause is a generated move faster or jerkier than anything your camera did. Slow the move at generation time rather than retiming after, and prefer a locked-off frame — models bias hard toward adding movement, so you have to say the camera is entirely motionless to get a static shot.
Grade to the plate, never the reverse. Pull a still from the nearest real shot, put it side by side, and match black level first, then mid-tone contrast, then saturation, then hue. Colour-matching every clip at assembly is the standard counter to drift across a multi-clip sequence.
Add grain last, on top of everything. Grain goes above the grade in the stack, or the grade re-contrasts your grain and it crawls. Match the amount to your real footage's noise floor, not to a preset labelled "35mm." Don't denoise the real plate down to meet the generated shot — you'll lose detail across the whole sequence to fix one insert.

If you are upscaling: clean and deflicker before you scale, then go to 1080p, then step to 4K. Jumping straight to 4K amplifies artefacts you could have removed more cheaply lower down. Upscaling happens after grain matching, not before.

What to do when the shot still won't sit

Sometimes the geometry is right, the grade matches, and it still reads as an outsider. Two blunt tools:

Cut it shorter. Mismatch is a comparison problem, and comparison takes time. A generated insert that runs 1.5 seconds gives the eye far less to work with than the same shot at four seconds. This is why generated material works best as b-roll and inserts rather than as the shot the scene rests on.

Put it behind something. A generated shot under a dissolve, behind a title, or as the second half of a whip gets a fraction of the scrutiny of a hard cut on a static frame. That's not a cheat; it is how mismatched material has always been integrated.

Or commit to the mismatch. The thing that never works is a half-match. A sequence that is clearly stylised throughout reads as a deliberate choice. A sequence that is 90 percent real with one nearly-matching insert reads as a mistake, because the audience has a reference for what the rest of it looks like.

A practical loop

Assemble the sequence with the generated shots in place at the length you actually intend to use. Run a low-resolution pass to look at it as a whole — the editor's 480p preview is free and carries a short per-user cooldown, and it is the right fidelity for judging whether a shot sits, because it strips out the detail that distracts you from the shape. Fix in the order above, then export once; a single charge applies to the final export regardless of clip count.

For sourcing the inserts, the AI B-Roll Generator is the surface built for this job, and the video editor is where the conform, grade and grain work lands.

FAQ

Should I match the generated shot to the real footage, or grade everything together?

Match the generated shot to the real footage. Your camera material is the anchor because it is internally consistent already — every shot from that day shares a sensor, a lens set and a lighting setup. Regrading the whole sequence to accommodate one insert means touching dozens of shots to fix one, and you will usually end up compromising the look of the footage you paid to shoot.

Which of the four axes do audiences actually notice first?

Grain in shadows and motion blur on movement, in that order — but focal length is the one that makes a shot feel wrong without anyone being able to say why. Grain and grade errors read as "that shot looks different." A focal length error reads as "that isn't the same room," which is a much harder note to recover from.

Can I add grain and call it done?

Only if the other three already match. Grain is a texture applied on top of a picture; it does nothing about geometry, contrast or motion. A perfectly grained shot with the wrong lens still announces itself. Grain makes a matched shot invisible — it does not make an unmatched shot acceptable.

Do I need to match motion blur if nothing in the shot moves?

No, and that is a good argument for generating locked-off inserts wherever the edit allows. A static frame with a static subject removes the motion axis entirely, leaving three things to match instead of four. All it takes is a line in the prompt saying the camera does not move.