Lighting continuity across separate generations

Six shots of the same room, generated separately, assembled in order. The character holds. The wardrobe holds. The set holds. And the sequence still does not read as one place, because in shot one the key is coming from camera left, in shot three it is coming from camera right, and in shot five the room appears to be lit from above by nothing in particular.

Lighting continuity is the consistency failure that survives every fix aimed at the other ones. Lock the character with a reference and the light still drifts. Lock the set and it still drifts. It drifts because nothing in a normal prompt pins it, and each generation re-decides every property the prompt left open.

Key direction is re-rolled unless you write it down

The trap is that lighting language feels like it is doing continuity work when it is not. "Warm interior light" in all six prompts sounds like a constraint. It constrains colour temperature and nothing else. The key can be anywhere in the room and still be warm interior light, so the model puts it wherever that particular composition suggests, which is different in every shot.

Named setups have the same hole. "Rembrandt lighting" fixes the angle magnitude — roughly forty-five degrees off axis — and says nothing about which side. Two shots can both be textbook Rembrandt and be mirror images of each other. From the viewer's seat that is not two shots of one room, it is two rooms.

What holds is writing the geometry down and never varying the words. It is the same discipline that keeps a character recognisable across a run of shots — hold the prompt structure identical and vary only the action, the approach character consistency across a campaign works through in full. Identical text produces a narrower distribution of outputs, and lighting answers to that exactly as identity does.

The lighting note

Write one block. Use it verbatim in every prompt in the sequence. Change nothing in it, including word order, until the sequence is finished.

A lighting note needs five things:

1. Key source and side. A large window at camera left — the side is the load-bearing word.
2. Elevation. above her eyeline or at eye level or low, from the floor.
3. Quality. soft, wrapping or hard-edged shadows. This is source size in plain language.
4. Colour. cool overcast daylight, warm tungsten, or a stated split.
5. Falloff. What happens where the key does not reach, and what fills it. The far wall falls to deep shadow; no fill.

Assembled into one paragraph you can paste:

LIGHTING (identical in every shot): key is a large window at camera left,
above eyeline, soft and wrapping, cool overcast daylight. One warm practical
lamp on the desk at frame right. The far wall falls to deep shadow with no
fill.

Then the per-shot prompt is that block plus a shot line plus an action line, and nothing else moves:

Medium close-up, static camera. She looks up from the desk and turns
toward the door.
LIGHTING (identical in every shot): key is a large window at camera left,
above eyeline, soft and wrapping, cool overcast daylight. One warm practical
lamp on the desk at frame right. The far wall falls to deep shadow with no
fill.

Wide shot, static camera. She crosses the room and stops at the window.
LIGHTING (identical in every shot): key is a large window at camera left,
above eyeline, soft and wrapping, cool overcast daylight. One warm practical
lamp on the desk at frame right. The far wall falls to deep shadow with no
fill.

Two shots, one changed variable. The block is longer than it feels like it needs to be and that is the point — every property you leave out is a property that gets re-rolled.

What varies, and what must not

The rule is that camera and action vary, light does not. That sounds obvious until you hit the shot where it is awkward.

Element	Varies per shot?	Note
Shot size, angle, movement	Yes	This is the whole reason you have six shots
Action	Yes	One clear action per shot
Key side and elevation	No	Camera-left key stays camera-left even when the camera has moved
Source quality and colour	No
Practicals	No	Same lamp, same place, even if it is out of frame
Falloff description	No

The awkward case is a reverse angle. If the camera crosses to the other side of the subject, a window that was camera-left is genuinely camera-right in the new frame, and copying "camera left" verbatim would be wrong. Two fixes work: anchor the note to the set rather than the camera (the window is on the north wall, behind the desk), or write a second, explicitly-mirrored note for the reverse shots and reuse that verbatim. Improvising per shot does not work. Set-anchored notes are the better default past a handful of shots because they survive any camera move; camera-anchored notes are quicker and fine for a sequence shot from one side.

If the sequence has recurring subjects as well as recurring light, the note pairs naturally with reusable characters and products, which hold identity by key while the note holds the light. They solve different halves of the problem and do not compete, as the four consistency mechanisms lays out in full.

The shot-by-shot review check

Run this before assembly, on stills pulled from each clip. It takes about a minute for six shots and it catches things the eye misses at speed.

1. Name the key side in every shot, out loud. Left, right, front, behind. Write the six answers in a column. Any disagreement that is not an intentional reverse angle is a reshoot.
2. Check the shadow direction on one fixed object. Pick something present in multiple shots — a chair, a mug, a door frame. Its shadow should point the same way relative to the set in all of them.
3. Check elevation via the terminator. On a face or any rounded object, where does lit turn to shadow? A high key puts the terminator low; a key at eye level puts it near the equator. Inconsistent elevation is subtler than inconsistent side and reads as "something is off" rather than as an error.
4. Check the practical. If a lamp is on in shot one it is on in shot six, and it is the same colour.
5. Check ambient level. Hold two shots side by side. If one has a visibly lifted shadow floor and the other does not, the model added fill to one of them.

Steps one and two catch most of it. Step three separates a sequence that looks assembled from one that looks shot. This is the lighting-specific slice of a broader problem: the four distinct things people mean by "these shots don't match" are separated in consistency collapse, and it pays to know which one you have before rerolling.

When the note is not enough

Chained clips. If you extend a sequence by taking the last frame of one clip and seeding the next, the light carries automatically and so does every accumulated artefact. The widely-repeated fix is to never feed a dirty frame forward: re-mint or upscale the extracted frame in image-to-image before it becomes the next clip's input.

Endpoint-critical shots. Where both the start and end state matter, first and last frame generation pins both ends with actual images, which pins the lighting at both ends too. That is a stronger constraint than prompt text, because the model is matching pixels rather than parsing a description.

Assembly. Colour-match every clip at assembly time. Even a perfectly-executed note leaves small exposure differences between independent generations, and matching them at the timeline is far cheaper than regenerating. The pre-publish check is the natural place for the final pass. Because the editor is EDL-based the timeline is re-renderable, so a reshot clip can be swapped in and re-exported without rebuilding the edit, and a single charge applies to the export regardless of clip count.

FAQ

Should the lighting note go at the start or the end of the prompt?

End, in most cases, with the shot line first. Cinematography-forward prompt structures put the camera at the front and it is the block most likely to get dropped if it is buried. The lighting note is long and distinctive enough that position matters less to it. What matters much more is that its position is the same in every prompt, since you are trying to keep the whole input string as stable as possible.

Does this work when I'm mixing models across a sequence?

Partly. The note still constrains each generation, but different models have different priors about ambient fill and contrast, so you will end up with more residual drift than a single-model sequence and lean harder on the colour match at assembly. If the sequence has to feel like one location, keeping every shot on one model is the cheaper choice even when another model would have been marginally better for one particular shot.

How long can a lighting note be before it starts hurting?

The note competes with the shot and action lines, so one that runs longer than the rest of the prompt combined tends to flatten the camera instruction. Five clauses is a good ceiling. Past that you are describing a setup complex enough to establish as a reference image instead.

What if one shot genuinely needs different light?

Then it is a different scene and should look like one — a deliberate change reads as a cut to somewhere else. The failure mode is accidental variation inside what is meant to be one continuous space. Write the second note out in full, mark which shots use which, and keep both verbatim.