Why AI Still
Needs VFX

A generation arrives fully formed and entirely convincing, and for a few seconds it is easy to believe the work is done. Then you cut it against the shot before it, or hold it for four seconds instead of one, and the illusion begins to migrate. A model does not render a thing. It renders an image of a thing, freshly guessed each frame. The job of traditional VFX is to persuade that guess to hold still, sit in a plate, and survive a broadcast delivery spec. I did not come to AI from prompting. I came to it from twelve years of compositing, tracking and cleanup, and that is the lens through which I keep finding the same truth: AI produces material, not finished frames.

01 - The seamMaterial is not a shot

A finished shot is a contract. It agrees to the frame before it and the frame after it. It agrees to a lens, a grain structure, a colour space, an eyeline, a light direction. A generation agrees to none of this by default, because it was never modelling continuity in the first place. It was modelling plausibility, one frame at a time.

This is the distinction that decides whether a project ships. On the Gigaclear commercial, the striking part was never the environment coming out of the model. It was reliable. The difficult part was the hundred-character crowd rebuilt in Cinema 4D and the real actors composited into a world that had no idea they were coming. AI gave us the raw plausibility. Everything that made it a national television spot happened afterward, in the parts of the pipeline that predate AI by decades.

02 · TimeTemporal consistency and flicker

The most honest way to see the limitation is to slow a clip down and watch a single surface. A brick wall shimmers. A jacket weave crawls. A patch of skin breathes at the pore level. None of this is motion in the scene. It is the model re-deciding what the surface is, frame by frame, with no memory of what it decided a sixteenth of a second earlier. Because the model holds no persistent geometry, it has nothing to be consistent to.

The fix is unglamorous and entirely traditional. You stabilise, you work in patches, you rebuild a clean plate and re-project it, you deflicker, sometimes you frame-blend or run optical flow to average away the crawl, and where a surface is truly load-bearing you replace it outright. What you are really doing is imposing a continuity the model could not supply, by hand, until the eye stops noticing the boil.

03 - The actor problemCompositing generated worlds with real footage

The moment a real person enters an AI environment, every weakness in the generation is put under a spotlight, because the actor is a model of a thing. They obey gravity, hold their proportions, and cast a consistent shadow. The plate behind them does not. So the seam is not the actor and it is not the environment. The seam is the relationship between them, and the relationship is what a compositor exists to build.

That relationship is assembled from ordinary craft, none of it new:

• Edges and mattes. A hard AI-blur edge reads as a sticker. Real edges have motion blur, spill, and softness that changes with depth. You roto, you refine, you decontaminate the fringe.
• Interaction. Contact shadows, bounce light, and reflections that the environment should be casting onto the actor and never does. You paint and light them in.
• Depth. Atmosphere, defocus and a shared depth cue so the actor sits in the world rather than in front of a screen of it.

04 · AnchoringTracking, match-move and the plate

Add an element to a moving shot and the question is immediate: does it belong to the camera, or does it float on top of it? AI-generated camera motion is a particular problem here, because there is no real camera behind it. There is no consistent focal length, no true parallax, no rigid relationship between foreground and background. The move looks right and solves for nothing, which means anything you add has nothing trustworthy to lock to.

So you reconstruct what was never there. You track the shot, you solve for a camera where you can, and where the generated motion is too incoherent to solve you stabilise, add your element into the calmed plate, and re-introduce the movement afterward. A logo, a product, a set extension only sits in the plate when it shares the plate's motion. Match-move is how a two-dimensional guess is given a place to stand in three dimensions.

05 - The artefactsCleanup, and why the model cannot help

The familiar failures are easy to spot: hands that change between frames, a logo that morphs as the camera passes, or text that resolves into approximate letters. They show the difference between generating a plausible appearance and maintaining a stable object through a shot. A hand has a fixed structure; an image model may produce the appearance of one without preserving every rule from frame to frame.

The repair is old-fashioned paint and tracking. You lock text to a tracked surface or replace it with real type. You rebuild a logo as a clean graphic and track it back on. You patch a warping hand across frames, borrowing from cleaner ones. This is the work that decides deliverability, and it is the work no amount of prompting removes. Which is why "95% AI" is an honest description of the material and a misleading description of the hours. The last five per cent is the finishing, and finishing is most of the craft.

06 · Continuity and the earColour, grain, and sound

A single spot is rarely one model. One shot comes from here, another from there, and each arrives with its own idea of black level, contrast and colour temperature. Cut them together untouched and the edit flickers between worlds. So you grade for continuity, not just for look: you neutralise, match, and carry a consistent response through the sequence so the shots agree they were photographed by the same instrument.

Then you make it photographic. Generated frames are often suspiciously clean, or carry the wrong grain, or lack the lens signature the eye expects, the slight softening at the edges, the chromatic fringing, the vignette, the bloom. You add matched grain, a real lens character, and defocus so every shot shares one optical fingerprint. And underneath all of it, sound. A convincing image with dead audio reads as fake; the right footstep, room tone and low-end weight will hold an imperfect frame far longer than another render pass ever could. The ear finishes the picture the eye was still doubting.

07 - The pointThrough VFX, not around it

None of this is an argument against AI. It is a fast way to generate useful material. The model’s output then needs to join the rest of the production process: temporal consistency, compositing, roto, tracking, cleanup, grade, grain and sound. These are not obstacles around an AI pipeline; they are part of the pipeline. Generation provides a first answer. Finishing tests whether that answer can hold through an edit and a delivery.