Generating B-roll That Actually Cuts Together With Your Real Footage

The problem with AI B-roll is rarely that it looks bad. Runway Gen-4, Kling 2.5, Veo 3, Sora 2 all produce clips that hold up on their own. The problem shows up on the timeline: you drop a generated insert between two shots you actually filmed, and the cut announces itself. The eye snags. Something about motion, grain, or color says this came from a different place, and now the viewer is thinking about your footage instead of your story.

I cut a lot of interview and brand work, and I've spent the last year figuring out why generated inserts refuse to sit down next to real cameras. Almost none of it is about the prompt being clever. It's about matching the physical characteristics of the camera you already shot on, and about generating far more clip than you'll use.

Match the lens before you match anything else

The single biggest tell is focal length and how the scene moves through space. If your A-roll was shot on a 35mm-equivalent at f/2.0 with a shallow, creamy background, and your generated B-roll has deep focus and a wide, security-camera field of view, no amount of color grading will rescue the cut. The two clips describe different physical spaces.

So I write the lens into the prompt explicitly, and I match it to the real rig. If I shot an interview on a full-frame body with a 50mm, my B-roll prompt starts with the optics:

Medium shot, 50mm lens, full-frame sensor, shallow depth of field,
subject in focus with soft background falloff, slight handheld drift,
interior office with warm practical lighting from the left, f/2.0,
24fps, natural motion blur

The words 24fps and natural motion blur matter more than people expect. A lot of models default to an over-sharp, slightly-too-smooth motion cadence that reads as video-game-ish. Asking for the frame rate you actually shot, plus motion blur, pushes the clip toward a filmed cadence. If you shot 60p for slow motion, say so, and conform it the same way you'd conform real 60p footage.

Generate long, generate many, then throw most of it out

Real B-roll editing has always been about coverage. You shoot four minutes of a barista to use six seconds. AI doesn't change that math; if anything it makes it worse, because the failure modes are unpredictable. A hand will morph on frame 70. A reflection will do something impossible halfway through. Text on a laptop will melt.

My working ratio is roughly ten generations to one usable clip, and I generate the maximum duration the model allows so I have room to find a clean two-second window. Kling and Runway both let you extend clips now, but extension compounds artifacts, so I'd rather generate several independent 8-to-10-second takes and hunt for the stable stretch inside each. Treat every generation as a take. Log the good windows. Discard ruthlessly.

A practical trick: scrub every generation at quarter speed before you commit to it. Artifacts that are invisible at full speed, a flickering earring, a background sign that breathes, become obvious. If you'd cut around it in real footage, cut around it here too.

Degrade the generated clip to meet your camera

Generated video tends to be too clean. No sensor noise, no chromatic aberration, no rolling-shutter wobble, no compression character. Real cameras have all of these, and your brain reads their absence as wrongness even when it can't name it. The fix is to add the imperfections back, matched to your source.

My finishing chain for a generated insert usually includes:

• Grain matched to the A-roll. I'll pull a frame of my real footage, sample the grain structure, and apply a matching grain plate. In Resolve I use the built-in Film Grain with the ISO dialed to roughly what I shot; if I shot at ISO 1250, I don't leave the insert at a phantom ISO 100.
• A touch of the same optical softness. If my lens is soft in the corners, I'll add a subtle vignette and a hair of edge blur so the AI clip doesn't look sharper than everything around it.
• Matched motion cadence. If there's a mismatch in shutter feel, a light directional blur or a frame-blend can bridge it.
• The same codec path. I render the insert and re-ingest it through the same transcode my camera files went through, so it carries the same compression fingerprint.

This is the step most people skip, and it's the one that does the heavy lifting. You are not trying to make the AI clip look better. You are trying to make it look like it came out of the same camera on the same day.

Color is a match, not a grade

Grade your real footage first and lock it. Then bring the generated clip to that look rather than grading it independently. I'll drop the insert next to a hero shot, throw a temporary split-screen up, and match by eye: black point, white point, skin tone, the color of the shadows. Generated clips often have a slightly plasticky, over-saturated midtone that you'll want to walk back. If your project has a distinct look, say a teal-shifted shadow, apply it as the same node structure you used on the A-roll instead of trying to prompt the model into the look. Prompted color is unreliable; a matched grade is repeatable.

Where AI B-roll genuinely earns its place

I'm not precious about this. There are shots that are expensive, dangerous, or simply impossible to reshoot, and that's exactly where a generated insert pays off: an aerial you didn't budget for, an abstract texture, a product detail you no longer have on hand, a location you can't get back to. For literal, in-scene B-roll of your actual subject or your actual product, real coverage still wins almost every time, and it's usually faster than fighting ten generations.

The honest trade-off: AI B-roll saves you a reshoot but costs you an hour of generation, selection, and finishing per usable clip. When the reshoot would cost a day, that's a great deal. When you could have grabbed the shot with your phone in the moment, it isn't.

A checklist before you commit a generated insert

1. Does the lens, focal length, and depth of field match the shots on either side of it?
2. Did you scrub it at quarter speed for morphing, flicker, and impossible physics?
3. Is the motion cadence, frame rate and blur, consistent with your A-roll?
4. Did you add grain and optical character matched to your camera?
5. Did you grade it to your locked look rather than grading it on its own?
6. Does it survive being watched three times in context without pulling focus?

Get those six right and the cut disappears, which is the entire goal. Nobody should ever watch your video and spot the AI shot. If they can, it wasn't finished.

One last note on rights and disclosure: model terms and platform rules around synthetic footage keep shifting, and requirements vary by where and how you publish. This is workflow guidance, not legal advice, check the current terms for the tools you use and the platforms you post to before you deliver client work.