Grain, flare and vignette against the AI sheen

Put a generated clip next to something shot on a real camera and the difference is rarely resolution. The generated frame is often sharper. It is evenly lit corner to corner, clean in the shadows, free of fringing, and every plane is in focus at once. That combination does not occur in photography, and a viewer who has watched ten thousand hours of footage recognises it as wrong long before they can name a single thing that is wrong with it.

The gap is not missing detail. It is missing error. A camera is a stack of physical compromises — a sensor with noise, a lens with falloff and aberration, a shutter with an angle — and every one of those compromises leaves a signature that we have learned to read as "photographed". Generated frames arrive with none of them.

Putting a few back is the last pass in a finish, and it is the one that most reliably moves a clip from "obviously generated" to "you would have to be looking".

What a camera adds that a model does not

Sensor noise. Real footage has grain, it varies with ISO, and it lives mostly in the shadows and lower midtones. Highlights are comparatively clean. Generated footage tends to have either no noise at all or a smooth, uniform dithering that reads as compression rather than sensor.

Lens vignette. Every lens loses light toward the corners, more so wide open. The falloff is gradual, radially symmetric around the optical centre, and usually subtle enough that you notice it only in a side-by-side.

Chromatic aberration and corner softness. Colour fringing on high-contrast edges near the corners, and corners that resolve slightly less than the centre. Generated frames are typically perfect to the edge.

Flare and veiling glare. When a bright source is in frame or just outside it, light scatters inside the lens barrel: streaks, ghosts, and a lift in the blacks near the source. Crucially it is motivated. It exists because something bright is there.

Motion blur consistent with a shutter. Real motion smears by an amount tied to shutter angle and frame rate. Generated motion is sometimes crisper than the speed it depicts would allow, which is part of why fast movement can look uncanny.

That is the whole vocabulary. You do not need all five, and the last two are the ones most often overdone.

Order matters more than the settings

The finishing order that survives contact with delivery:

1. Clean and deflicker first. Shimmer and frame-to-frame flicker get amplified by everything downstream. If your source is flickering, fix that before you enlarge it. Flicker is also partly a generation-side choice: clips generated in the 24 to 30 range tend to come back steadier than the same prompt pushed to 60. Versely's default sits at 25.
2. Step the resolution, do not jump it. Clean to 1080p, then 1080p to 4K if you need 4K. Going straight from a small source to 4K amplifies every artifact you could have removed cheaply at the intermediate step. Upscaling is the mechanism; upscale a video to 4K is the tool route. Be aware that upscalers tuned for live-action restoration can be aggressive on generated footage, inventing detail that was never in the render — worth checking a still frame at full size before committing a whole sequence.
3. Grade and match across clips. Every clip in a multi-clip sequence needs matching to one reference look, because separate generations do not agree on contrast or colour temperature. Why your generated clips do not match in the timeline covers why, and what a grade can and cannot rescue covers the limits of doing it heavily.
4. Texture last. Grain, vignette, flare, aberration. All of it after the upscale and after the grade.

Step four's position is not arbitrary. Upscalers treat grain as noise and remove it, so grain applied before an enlargement simply disappears and you have paid for nothing. Grading after grain redistributes the noise unevenly, which is how you end up with grain that is heavy in the shadows of one shot and absent in another. Apply texture to the finished, matched, correctly sized image and it stays where you put it.

Amounts, which is where this usually goes wrong

Grain. Fine, monochrome, weighted toward midtones and shadows, kept largely out of the highlights. The test is not whether it looks good — it is whether you can see it at all. Pause on a still frame at normal viewing distance on the device people will actually watch on. If you can identify grain without being told it is there, it is too much. Correct grain is only obvious when you toggle it off and the frame suddenly looks plastic. Colour grain, or grain that swims independently of the image, reads as a filter immediately.

Vignette. Centre it on the frame's optical centre, not on your subject. Radially symmetric, feathered across most of the frame radius, and shallow — a fraction of a stop at the extreme corner, not a dark ring. If the vignette has a visible edge anywhere, the radius is too small. A vignette that draws attention to the subject is a design element; a vignette that imitates a lens is one nobody notices.

Flare. Motivated only. If there is no bright source in or near the frame, a flare is a sticker, and everyone can tell. Anchor it to something in the shot — a window, a practical lamp, a low sun at the frame edge — and match its position to that source. A flare that stays put while the camera moves is worse than no flare at all.

Chromatic aberration. Almost nothing, and only near the corners. This is the effect that most quickly becomes a costume.

Consistency across the sequence. This matters more than any individual amount. The same grain profile, the same vignette, the same aberration on every clip in the cut. Inconsistent texture is a stronger tell than no texture, because a viewer registers the change at the cut point even when they cannot describe it.

Push what you can upstream

A finish is a correction. It is generally better to get the look into the generation, where it exists in the pixels rather than being painted on afterwards, and where it has not yet been through a compression pass that will fight it.

The trap is that the words that sound like they should do this are the ones that do the least. "Cinematic", "moody", "atmospheric" mostly bias the grade toward contrast and desaturation; they do not put a light anywhere or specify a lens. The phrasing that works names the physical setup:

Shot on 35mm film, visible fine grain, slight falloff toward the corners.
85mm lens, shallow depth of field. A warm practical lamp at frame left,
ambient room fill, subject at 45 degrees to the source.

One concrete lighting or optical descriptor is worth more than five style adjectives. That is true across model families and it is the single highest-leverage habit change in prompt writing for anyone chasing a photographic result.

What this pass cannot do, and what it costs

Texture does not repair content. A hand with six fingers under film grain is a hand with six fingers. Anatomy, prompt-adherence misses, morphing objects and motion incoherence all survive any amount of finishing, and reaching for grain to solve them is an expensive way to spend an hour discovering that.

It also costs bitrate. Grain is close to random, which makes it the most expensive thing in your frame to encode, and platform re-encodes will smear a heavy grain pass into blotchy patches — the opposite of the effect you wanted. Deliver conservatively for social, where the compression budget is tightest, and check your export on a phone rather than on a grading monitor. Why colours shift between screen and export covers the wider version of that problem.

On the assembly side, the Versely editor is an EDL-based timeline that handles clips, transitions, text, music, voiceover, overlays and captions, and the texture pass belongs in your grading tool rather than in it. One practical note if you are checking work there: the editor's preview: true pass renders at 480p and carries a short per-user cooldown, which makes it useful for confirming timing and text placement but useless for judging grain — at 480p a correct grain pass is invisible and an excessive one may still look fine. Judge texture on the final export. How previews and the final export are charged sets out how that works.

FAQ

How much grain is too much?

If you can see it on a still frame at normal viewing distance without looking for it, too much. The working target is that removing the grain makes the frame look wrong, while adding it produces no conscious reaction. Nobody should be able to point at the effect.

Should I add motion blur to generated video?

Only if the footage looks unnaturally crisp during fast movement, and then sparingly. Generated motion is sometimes already blurred, and stacking a second blur on top produces a smeared look that is its own kind of tell. Check a fast-moving frame first, and treat this as a targeted fix rather than a standing part of the recipe.

Does the finishing pass make footage harder to identify as AI-generated?

Not in any way you should rely on. The pass addresses the visual signature that a human viewer reads — uniform sharpness, clean corners, absent noise — and it does nothing about provenance metadata or any of the disclosure obligations that attach to synthetic media. Treat it as craft, not as concealment.

Is it worth doing this on short-form vertical video?

Vignette and flare, marginally. Grain, often not: the compression budget on a fifteen-second vertical clip is tight enough that a grain pass mostly becomes encoding noise, and the viewing surface is small enough that nobody sees the intended effect. Consistency across a series matters more than texture at that size, which is closer to the argument in making a faceless channel look authored.