The photo you already have is the weakest layer in the poster you are about to make. It is the thing you own, it is the reason you started, and it is the one element that has no say in how the finished picture is lit — because the light in a poster is decided by whatever gets rendered last, and that is usually the part you did not bring.
Which is why posters made from a selfie have a particular look. The face is fine, the ruined city is fine, and the person still reads as cut out and dropped in. What follows is six stills and one clip working out where that comes from, and what actually moves it.
The input, and why it is worth measuring
The starting picture was generated deliberately flat: an ordinary portrait, indoors, lit head-on from a window, no side in shadow. That is what most people are working from.
The two boxes sit on the same garment at the same height. The difference between them is the number to watch.
Measuring "which side is lit" needs a number that survives a change of material. Brightness does not: a concrete block reads brighter than asphalt whatever the light is doing. The blue-yellow axis of CIELAB does, because it moves with the color of the light falling on a surface rather than with the surface. Warm light pushes it up, blue shadow pulls it down. Sampled on the same jacket on both sides, the input reads +15.8 on the left, +10.7 on the right, a difference of +5.1. Close to nothing, which is the correct answer for a portrait with no side light.
Three ways to put that person in a burning street
All three routes were given the same scene: a ruined street at dusk, a heavy orange fire glow low on the left, deep blue shadow filling the right.
The scene, shared across every route. A ruined city street at dusk: collapsed concrete, overturned cars, rubble, ash drifting in the air, a heavy orange glow low on the horizon on the left, deep blue dusk shadow filling the right side of the street. Cinematic grade, heavy film grain, no text, no letters, no numbers, no logos.
Same scene, same measurement, four outcomes.
Locating the scene's own light was done the same way in every frame: mask off the vertical band the person occupies, take the warmest two percent of the remaining pixels, and ask where their center of mass sits. In all of these it lands between 7 and 17 percent of the way across the frame — the fire is on the left, as asked, in every one.
The composite line is the one to sit with. Cutting the man out and placing him in a street with a hard warm-cool split moved his lighting from +5.1 to +4.1. In other words, nothing happened to him at all. He arrived carrying the light from the room he was photographed in, and a cut-out has no mechanism for changing that.
Generating him inside the scene solved the light and cost the face. The prompt described a man in his early thirties with short dark hair and light stubble in an olive jacket, and it produced exactly that — a different person. That route makes a good poster of someone who is not you.
The sentence that moved the light
Which leaves the middle route, and this is where the useful part is. The same reference photo, the same scene, the same model, twice — with only the lighting sentence rewritten.
First wording. Relight him to match that scene: a hard warm rim light along his left cheek and left shoulder, the right side of his face falling into cool shadow.
Second wording. A large orange fire glow fills the LEFT EDGE of the frame behind him; the RIGHT SIDE of the frame is deep blue dusk shadow. Light him from that fire: everything on the LEFT HALF of his body as we see it — the left side of his face, his left shoulder, his left sleeve — is warm orange, and everything on the RIGHT HALF of his body as we see it is cool blue shadow. The warm and cool sides of his body must match the warm and cool sides of the street behind him.
The first wording put the warm light on the side of him away from its own fire: −7.9, the wrong sign. The second put it on the fire's side both times it was run: +55.3 and +11.7. Same picture in, same scene out, the face kept in all three.
The reason is not mysterious once it is written down. "His left" names a side of a body; the render has to decide whether that means the body's left or the viewer's left, and there is no way to tell from the sentence. "The left edge of the frame" names a side of the picture, which has exactly one meaning. The failure was not a limit on relighting. It was a pronoun with two readings, sitting in the one sentence that had to be unambiguous.
This applies past lighting. Any direction word in a prompt has the same fork in it — <em>behind him</em>, <em>to his right</em>, <em>over his shoulder</em>, <em>facing away</em>. Each is a place where the render can be right about the words and wrong about the picture.
What the cut-out route costs on the web
The composite route is the one people reach for when the face has to be exact — a real person, a real event, an actual likeness. It is also the route with the light problem, so the obvious question is whether the editor can relight a cut-out after the fact.
Smart tools lists Relight. The panel behind it, on the web, is one sentence and one button.
Background removal itself worked cleanly — Smart tools, Remove background, Auto removal, one switch, and the edges came back tidy. Relight is listed in the same panel, and on the web on 17 September 2026 it opened to a description and a single control reading <em>Open in desktop app</em>. There were no sliders, no direction, nothing to set.
So on the web the sequence matters more than it otherwise would. A cut-out placed over a scene keeps the light it was shot in, and the built-in repair for that is somewhere else. If the exact face is the requirement, the working order is to relight in a generation pass first and composite second, rather than composite and hope to fix it after.
Putting the title on it
Words in the picture are a separate job with its own reasons, covered in the walkthrough on setting names and dates over a generated frame and again in the piece on checking poster headlines. The short version holds here: Text in the left rail, <em>Add heading</em>, edit on the canvas, drop the size until the line fits, drag it into the lower third.
The finished frame, exported at 3840 by 2160. The film name is invented.
One thing worth knowing before you export a still. A text layer lands at the playhead, not at the start of the timeline. The playhead was three seconds in, so the first frame of the export had no title on it at all, and a grab of frame one would have looked like the type had failed. Park the playhead at zero before adding text, or pull the grab from a frame the text actually covers.
What happens once it moves
A poster that has been talked into matching its own light is a good candidate for a short loop, so the matched version went through image to video with a conservative brief: hold the framing, keep the fire on the left and the blue on the right, a very slow push in.
1280 by 720, 24 frames per second, 121 frames, 5.04 seconds.
The light holds. The shot size does not.
The good news first: the split survives. Measured on the jacket with boxes placed separately for each end of the clip and checked on the picture first, the difference reads +53.5 at the opening frame and +22.1 at the last one. It weakens as the fire moves out of shot, and the warm side is on the left at both ends and in every frame of the strip above. Whatever the video model re-renders, it did not throw away the lighting it was handed.
The camera instruction is the part that did not land. "A very slow push in" produced a change of shot size: the head measures 190 pixels across at the first frame and 405 at the last, a factor of 2.13 in five seconds. That is not a drift, it is a move from a waist-up frame to a tight close-up. If the poster composition is the point of the clip, the move has to be budgeted for — either by asking for less than you want, or by accepting that the last second is a different shot from the first.
What it cost
Thirty-one credits, and the split inside that number is the lesson about where to spend. Six credits of stills answered the whole question — which route keeps the face, which one keeps the light, and what the sentence has to say. The single clip cost more than all six together and returned one finding about the camera. The figures are what the panel showed for these configurations on 17 September 2026; the balance moved from 23,697 to 23,666 across the session, which matches the six single credits and the twenty-five.
Rewrite every direction word into something the camera can see
The fix in this article was a sentence, and it generalizes into a habit that takes about ten seconds per prompt.
Read back what you wrote and find every word that points somewhere. Then ask of each one: is this a direction the camera can see, or a direction only the subject knows? <em>His left</em>, <em>her right</em>, <em>behind him</em>, <em>over her shoulder</em> are the second kind. <em>The left edge of the frame</em>, <em>the top right corner</em>, <em>nearer the camera than the car</em>, <em>the half of his body we can see</em> are the first.
Anything in the second group gets rewritten into the first group before you send. It costs nothing, it is not a longer prompt, and on the evidence here it is worth the difference between a person standing in a scene and a person standing in front of one.
Written 17 September 2026. Every still and clip here was generated in CapCut for this article; no photograph of a real person was used, and the film title, the scene and the poster are invented and are not connected to any released work. Stills are 2560 by 1440 PNGs from Seedream 4.5 at 16:9 and 2K; the clip is 1280 by 720 at 24 frames per second from Seedance 2.0 Mini; the finished poster is a frame from a 3840 by 2160 editor export. The lighting figures are the mean of the blue-yellow CIELAB channel inside boxes placed on the same jacket on each side of the body, and the scene's light direction is the horizontal center of mass of the warmest two percent of background pixels with the subject's vertical band masked out. Every box and mask was rendered onto the picture and checked by eye before any number was taken from it; two earlier versions of the scene measurement were built and discarded first, one sampling ground at a fixed height and one sampling sky, because the two disagreed about which side of one frame was lit and neither could be trusted after that. One image credit was spent by a run that errored after sending; its output was the same prompt as the second wording and is reported above as the second of the two tries rather than discarded. The head-width figures are hand-marked on two frames and the marks are published with them.