One of the four prompts below asks, in so many words, for blacks that are not solid — for shadow that sits a little above black, the way ink does on old paper. It produced the most solid blacks of the four: 59.2 percent of the picture at a level where no detail can exist, against 9.5 percent for the run that did not mention blacks at all. The sentence asking for open shadow closed it.
The four asks
Everything below was generated on the AI image generator at 16:9. One paragraph of scene description stayed identical in all four runs — a woman in a dark coat in the left third of the frame, a plain wall behind her, a lantern held low with only part of it in shot. Only the sentence about light and finish changed.
The same scene four times. The percentage is how much of the picture sits below level 4 out of 255, where nothing can be seen however you grade it afterwards.
What asking for dark costs
Run 1 is the prompt everyone writes, and a version of it is the one circulating with this trend. It put 46.3 percent of the picture at dead black. Her coat is one shape rather than a coat; the right half of the frame is nothing at all. That is the complaint about this style, and it is not a grading problem — there is no signal to bring back.
It also came back as a picture inside a white page rather than a picture filling the frame, which is a different way of not getting a poster. That happened in this run and I would not read a rule into a single run of it.
Naming the light instead
Run 2 says nothing about how dark the image should be. It says where the light comes from, which side of the face it lands on, and that the coat's weave has to stay readable in the shadow. Dead black fell to 9.5 percent — about a fifth of run 1 — and the picture reads as lit rather than as dim.
This is the same shape of result as naming a light source rather than describing a mood elsewhere in this series. Darkness is a consequence of where you put the lamp. Asked for directly, it arrives as absence.
What grain does, and what it cannot do
Run 3 adds coarse film grain to run 2 and nothing else. Texture rises at every brightness: from 0.95 to 1.29 in the band around level 8 to 16, and from 1.20 to 2.99 in the band around 32 to 64. Dead black stays where it was, at 12.0 percent.
In this test, adding coarse film grain increased measured texture in each sampled brightness band. What it cannot do is put texture into a pixel that is already at zero. Grain decorates signal; it does not create it. Run 1 with grain on top would still be 46 percent nothing.
The sentence that closed the blacks
Run 4 adds one more clause to run 3, asking for blacks that sit above black with a faint green cast. Dead black went from 12.0 percent to 59.2 percent.
The green cast did arrive — you can see it in the fourth panel. The lifted blacks did not; the opposite arrived. My reading is that naming black four times in one sentence weighs more than the qualifier attached to it, but that is a reading and not a measurement, and it is one run. What is measured is the outcome: the clause written to protect the shadows is the clause that cost them.
The title comes out flatter than the picture
The finished poster takes run 3 into the editor, where the title goes on as a text layer. The type controls are real ones — a font list grouped by language, a size field, and a style section with Fill, Stroke, Background, Shadow and Glow, plus a Transform block with numeric scale, position and rotation.
Inside the letterforms the local variation is 0.461. On the lit wall, at almost the same brightness, it is 1.286.
That gap is the second half of the complaint about these posters. The picture carries 2.8 times the texture the type does, so the letters sit on the image rather than in it. Old printing put the same screen through both. No dedicated grain control for text was visible in the text panel used for this editing session. If you want them to match, one option is to take the type down to the picture rather than the other way round: a slight Opacity reduction, or a Shadow soft enough to break the edge, so the letters stop being the cleanest thing in the frame.
The finished piece. In this editing session, the export dialog displayed MP4 and MOV rather than a still-image export option, so the poster was saved from a video frame.
What to write instead
Where this stops
Four runs, one subject, one scene. Each condition was generated once, so the 59.2 percent result in particular is one draw and not a rate — the direction is large enough to act on and the exact number is not. The two conditions that matter most, dark against named light, differ by nearly five times, which is further than run-to-run variance has moved anything else measured in this series.
The measure itself is narrow on purpose. It counts pixels below level 4 and the local variation among pixels in each brightness band, so it says nothing about whether the picture is a good poster — only about whether there is anything left in the shadows to work with.
Light first, then grain
These four runs sort into an order rather than a list. Get the light right and the shadows stay open; add grain and it lands on something; add a sentence about the blacks themselves and it undoes the first step. Grain cannot be applied to a pixel that has already gone to zero, which means the sequence is not a preference — the second step depends on the first having worked, and no amount of finish will recover a picture that was asked to be dark before it was asked to be lit.
Configuration and cost. All four images: Image, Seedream 4.5, 16:9, 2K, Standard — the panel declared 1 credit per send, four sends. Total 4 credits, read from the send panel on each send, 2026-09-21. Editor work and export produced no charge. Credit costs are tied to that configuration on that date; other model, resolution or quality settings are priced differently.
Files.dark.png, lit.png, grain.png and print.png (runs 1 to 4), all 2560×1440; poster.mp4, the finished poster as exported at 1920×1080, and poster.png, a frame taken out of it.
The prompt lines are reproduced as they were sent, including their spelling.
How the numbers were taken. Luminance is Rec. 709 luma. Dead black is the share of the picture below level 4 out of 255. Texture is the median of a 5×5 local standard deviation among the pixels falling in each luminance band, which compares like with like and does not reward a picture simply for having more black in it. An earlier version of this measure looked at the darkest 40 percent of each picture and was discarded: it tracked how much empty black background each composition happened to contain rather than how the shadows were rendered. The band measure was self-tested first by blurring and by adding known noise to one of the images, which dropped and raised the figure in every band while leaving the band populations alone. The first run's white margin was detected and excluded before measuring. The type figures use the same local measure inside letterform interiors, eroded by nine pixels so that anti-aliased edges are not counted, against the lit wall eroded the same way; both sampling boxes are drawn in the figure.