How to match lighting across AI video shots so the character still looks the same

Every shot invents its own light

Put the same character in two shots of the same room and render them separately. In the first the light comes from the left and the face is soft. In the second the light comes from above and the cheekbones are suddenly sharp. Nobody changed the character. The engine guessed the light again.

Page 1 for this question agrees on the cause. A HackerNoon piece puts it in its headline: every AI-generated shot invents its own light. The advice that follows is right but thin: set lighting once, repeat it in every prompt, use a reference frame. Kling's own blog says to repeat the shared light source across multi-shot scenes. One guide gives "key light from the upper left at 45 degrees" as an example. What none of them says is where the light line goes in the prompt, what to do when it fails, or that fixing light after the fact is a full render.

Light belongs in the character block

We split every shot prompt in two. The character block describes who is on screen and is copied word for word into every shot. The scene block describes the place, the action and the camera, and is the only part we edit between shots. Our descriptor block guide covers how the character block is written.

Most people put lighting in the scene block, because light feels like part of the place. We moved it into the character block for anything inside one scene, because light changes how the face is drawn. A face lit hard from one side reads more angular than the same face under soft front light. If the light line changes between shots, the face changes with it, and the viewer reads that as a different person.

So the character block for a scene gets one fixed light line at the end:

Mara, early thirties, shoulder-length dark brown hair with a centre parting, round face, light freckles across the nose, olive canvas jacket over a cream knit jumper. Light: soft warm key from camera left at face height, gentle fill from the right, no hard shadows on the face.

That line answers three questions, and only three: where the light comes from (camera left), what colour it is (warm), and how hard it is (soft, no hard shadows). Those three are what flip between renders. "Moody lighting" or "cinematic lighting" answers none of them and leaves the engine to guess.

When the scene changes, the light line changes once, at the cut, and then stays fixed for every shot of the new scene.

Keep the direction on the same side

The most visible failure is the key light jumping sides between cuts: lit from the left in the wide, from the right in the close-up. Pick the side once per scene and write it in the same words every time. "Camera left" means the left of the frame. Do not switch to "her right" in the next shot, even though it is the same side.

If a shot turns the character around, keep the light on the same side of the frame. A cinematographer would move the lamp, and the engine has no lamp to move, so naming the side of the frame is what keeps the cut readable.

Light in the references

The reference images matter as much as the words. A reference sheet shot under even, neutral front light gives the engine a face it can relight in any direction. A moody portrait with half the face in shadow teaches the engine that half the face is dark, and it keeps doing that.

Our references are always flat-lit: front, three-quarter and side, soft light, plain background. The mood comes from the light line. How many references each engine takes:

engine	references	how the prompt names them
Kling O3 Pro	5 elements	@Element1 to @Element5
Seedance 2.5	30 images	@Image1 to @Image30
Seedance 2.0 and 2.0 Fast	9 images	@Image1 to @Image9
Veo 3.1 Fast	3 images	image 1 to image 3

On Seedance 2.5, with thirty slots, we add one still of the lit set as the last reference. The prompt can then say "in the room and light shown in @Image6", which gives the engine a picture of the light as well as the words.

When the last frame carries the light

The surest way we know to keep light between two shots is to start the second from the last frame of the first. The engine begins from a frame that already has the light it drew, so the light holds. We use it for continuous action within one setup: a character sits down, stands up, walks to the window.

It stops working when the camera angle changes a lot. A close-up started from the last frame of a wide will drift toward the wide's framing before it settles. For a real cut to a new angle, render fresh with the same light line and the same references. Our first and last frame guide covers when the carry works and when it morphs.

A relight is a full retake

No engine we sell offers a partial relight. If shot four comes back lit from the wrong side, fixing it means rendering the whole shot again at the full rate. That is why the light line is worth getting right before the first render, not after the fourth.

What one 8-second retake costs at 720p:

engine	silent	voiced
Veo 3.1 Fast	160	240
Kling O3 Pro	192	240
Seedance 2.0	360	360
Seedance 2.0 Fast	400	400
Seedance 2.5	560	560

Seedance bills by resolution, so silent and voiced cost the same. A six-shot scene on Kling O3 Pro, voiced, is 1,440 credits. If two shots come back with the light on the wrong side, the retakes add 480, a third on top. On Seedance 2.5 the same two retakes are 1,120.

The order we fix it in

When a shot's light does not match, change one thing per retake:

1. Check the light line was pasted exactly. One dropped phrase, such as "no hard shadows", is enough to change the face.
2. Name the side in frame terms. Replace any "her left" with "camera left".
3. Soften the light. Magic Hour's prompt cookbook makes the same point: flat, even light is the easiest to keep consistent from shot to shot.
4. Carry the last frame, if the shot continues the previous one.
5. Switch the reference set to flat-lit images, if the old ones were moody.

Then, and only then, colour-match the finished shots in your editor. A grade can warm or cool a shot to match its neighbours. It cannot move a shadow from one side of a face to the other.

In the studio the references are saved with the character and go into every shot in the same order, and each retake's credit cost shows before you render it.

FAQ

How do I keep lighting consistent across AI video shots?

Write one light line that names the direction, colour and hardness of the key light, such as soft warm key from camera left, and paste it word for word into every shot of the scene. At ReelDouble we keep it at the end of the character description, next to the face.

Why does my AI character look different when the lighting changes?

The engine draws the face under the light you describe, so hard side light makes a face read sharper and more angular than soft front light. If the light words change between shots, the face changes with them.

Can I fix the lighting of an AI video shot without rendering it again?

Not the direction of the light. A colour grade can match warmth between shots, but moving a shadow means a full retake. On ReelDouble an 8-second voiced retake at 720p is 240 credits on Kling O3 Pro and 560 on Seedance 2.5.

What lighting should AI character reference images have?

Flat, even front light on a plain background, from the front, three-quarter and side. Moody reference portraits teach the engine to keep half the face dark in every shot.

Does using the last frame of a clip keep the lighting the same?

Yes, for continuous action in the same setup, because the next shot starts from a frame that already has the light. For a cut to a very different angle, render fresh with the same light line and references.