Reflections and mirrors in generated footage

Put a character in a bathroom and generate. The room is right, the light is right, the face holds — and the reflection in the mirror is a person standing at a slightly different angle, in a room with one fewer cabinet, wearing a shirt whose collar does not match. Do it with a shop window and the street in the glass belongs to a different street. Do it with a puddle and the building it reflects has more floors than the one above it.

This is one of the more reliable tells in generated footage and one of the least discussed, because it survives a casual watch. Anyone who pauses catches it immediately.

Why the reflection disagrees with the room

A reflection is not decoration. It is a second, geometrically constrained rendering of the scene: everything in it is determined by what is in front of the surface, where the surface sits, and where the camera is.

Generative models match appearance rather than solving geometry. They have learned very well what a mirror looks like — a bright rectangle containing a person-shaped thing, roughly the right size and pose. What they have not learned is that the content of that rectangle is a strict function of everything else in the frame. So you get a reflection that is statistically plausible and specifically wrong.

The failure is worse in video, because a reflection has to stay consistent with the scene and with itself over time. As the camera moves, a real reflection shifts by a different amount than the surface it sits on — that parallax is what makes it read as depth rather than as a picture on a wall. Models rarely get it right, so the reflection often tracks with the mirror frame instead of behind it, which reads as a screen.

Mirrors also flip handedness, and text in a mirror is reversed. Text is already the weakest surface in generated imagery — models learn glyph shapes rather than letterforms, which is why background signage is unreliable — and reversed text adds a transformation on top of that. A mirrored logo is close to a guaranteed artifact.

Compose so the reflection is not load-bearing

The most economical answer is to arrange the shot so that no reflection has to be correct. Five tactics, roughly in order of how often they apply.

Angle the surface out of frame. A mirror cropped at the edge of frame, showing a sliver, gives you the production value with almost none of the risk. A mirror centred and facing camera is the maximum-exposure version of the same object.

Point it at off-screen space. A mirror angled to return part of the room the camera never sees is unfalsifiable — nobody can compare the reflection to a source that is out of shot.

Break up the surface. Steam on a bathroom mirror, grime on a shop window, rain distortion on glass, ripples across a puddle. Each gives a physical reason for the reflection to be indistinct, and an indistinct reflection cannot be checked. Prompt them as material conditions rather than effects: "condensation across the lower half of the mirror," "rain running down the glass."

Put it out of focus. A reflective surface behind the subject at a shallow depth of field is soft by construction — the same logic as the tactic for hands and other high-risk detail, set out in hands, teeth and signage.

Choose a darker surface. A black glossy tabletop, a smoked-glass panel, a dark screen. These read as reflective while returning very little checkable content. Product work leans on this constantly.

The one arrangement to avoid unless the shot demands it: a subject facing a large, clean, front-on mirror with both subject and reflection sharp and comparable. That composition invites the exact comparison the model is going to lose.

When the reflection matters, prompt it as a second subject

Sometimes the reflection is the shot. A character regarding themselves, a product on a polished surface, a city in a rain-slicked street. In that case, stop treating the reflection as something the model will fill in and describe it explicitly.

Three things to state:

1. The surface, with its material and its condition. "A large clean wall mirror," "a black polished acrylic surface," "still water, no ripples."
2. What is reflected, described as concretely as you described the subject. Not "her reflection" — "her reflection showing the same navy jacket and the same tied-back hair, seen from the front."
3. The geometry, in plain language. "The mirror is directly behind her and the camera sees her back and her reflected face."

That third line is the one people leave out and it carries the most information. "Mirror behind her, camera sees her back and her reflected face" is a specific, checkable arrangement. "She looks in a mirror" is a category of image.

Two further levers when the reflection is load-bearing:

• Bind it with an image. A reference image of the scene, or an image-to-video path starting from a still you have already approved, constrains the reflection far more tightly than any sentence. If the still is right, the video has somewhere correct to start.
• Lock both ends. First-last-frame generation lets you specify opening and closing frames and leaves the model only the motion between them. The reflection is then correct at both ends by construction and only has to survive the interval — a much smaller ask than maintaining one across an open-ended generation.

In product work the reflection on a glossy surface is often what sells the material, so it is worth setting up deliberately. The product video generator path benefits from the same discipline: name the surface, the reflected content, and the geometry.

The three-question check

Run this on stills pulled from the render, at full resolution, before you look at anything at speed. At 25 fps a wrong reflection is on screen too briefly to catch in playback and plenty long enough to catch in a pause.

1. Content. Does the reflection contain what is actually in front of the surface — the same person, the same wardrobe, the same objects, in the same arrangement? Count things. A reflected room with a different number of doors is the common form.
2. Handedness. Is the mirror image flipped the way a mirror flips? A parting on the left should read on the right. Any text should be reversed. A reflection that is a straight copy rather than a mirrored one is wrong even when everything in it is correct.
3. Parallax. Across the clip, as the camera or subject moves, does the reflection shift by more than the surface does? Reflected content sits behind the surface plane and therefore moves differently from the frame around it. A reflection locked to its frame reads as a poster.

Question three only exists in video, and it most often survives a still-frame review and then quietly makes the shot feel wrong. Check it by comparing the first and last frames side by side rather than by watching. The editor's pre-publish check covers the broader list, and the diagnostic tree for generations that come back wrong is a structured route when a shot is failing in ways you cannot name.

The masked repair path

When the shot is otherwise good, the reflection is bounded, and re-shooting risks losing a performance, repair it in place.

The stills route is more reliable. Extract the frame, mask the reflective region, and regenerate only that area with inpainting — describing what should be there rather than what is wrong. The AI photo editor handles the still-image side, and if that frame seeds a chained continuation, it is now a cleaned handoff as well.

The video route needs the mask to travel with the surface, which is what video segmentation is for. It works and it is meaningfully more effort. Repair when the clip contains something you cannot easily get again; re-shoot when it does not.

A third option is underrated: replace the reflection with a practical. Regenerate with steam, grime, a lower-key surface, or the mirror pushed to the edge of frame. You are not fixing the reflection, you are removing the requirement that it be correct — usually the fastest path to a shot that ships.

Previews are the cheap way to check an assembly before committing. The editor's 480p preview pass is free and carries a short per-user cooldown, and the final export is charged once regardless of clip count — though 480p is for timing and continuity of action, not for judging whether a reflection is correct. That needs full-resolution stills.

FAQ

Are reflections getting better with newer models?

Rendering quality has improved broadly and reflections benefit from that. The mechanism has not changed: models still match appearance rather than solving the geometry that determines a reflection's content. Expect better-looking reflections that are not necessarily correct ones, and keep checking.

Does a character reference fix the person in the mirror?

Partly. A reference constrains appearance, so the reflected figure is more likely to be the right person wearing the right thing. It does not constrain geometry, so the reflected pose and angle can still be wrong. References help with the content question and do nothing for handedness or parallax.

What about reflections in eyes and on skin?

Those are specular highlights rather than image-forming reflections, and they behave much better, because a highlight only has to be in a plausible position rather than contain correct content. Catchlights are worth prompting deliberately — naming a light source with its size and direction gives you a catchlight shape that matches the scene.

Is it worth avoiding reflective surfaces entirely?

No, and the shots would be poorer for it. Reflective surfaces carry real production value. The rule is not to avoid them but to avoid making them checkable: off-axis, out of focus, broken up, or pointed at space the camera never sees.