In one CapCut Seedream 4.5 comparison, a prompt naming the door and front fender produced its strongest red color on the hood instead. That makes panel-by-panel inspection useful when creating a car visual; a single pair of images does not establish that the model cannot follow a panel instruction. Use your own or permitted portrait and vehicle materials, and avoid another brand’s badges or distinctive design.
What the paint does when nobody asks
The room was built to make a reflection easy to see. A deep red wall down one side with four fluorescent tubes on it, plain gray concrete on the other, an unbranded dark coupe parked between them at three-quarter angle.
The scene, identical in every still below. A wide editorial photograph of a car and a person in a concrete parking structure. The car is an unbranded dark grey coupe with a glossy paint finish… It carries no badges, no logos, no model name and no manufacturer grille pattern anywhere on the body… On the left of the frame is a deep red painted wall with four tall vertical fluorescent light tubes mounted on it in a row. On the right of the frame is a plain dark grey concrete wall with nothing on it. Overcast daylight from above.
For this comparison, I counted image pixels with brightness of at least 230 on a 0–255 scale. The light tubes themselves crossed that threshold, while sampled plain wall and floor patches did not. Bright pixels on car paint may indicate highlights, but this threshold alone cannot identify what the surface reflects or prove the absence of a reflection.
Three outputs without an explicit reflection request put 0.110%, 0.118%, and 0.154% of the sampled car pixels above that brightness threshold. These figures describe the tested images and sampling boxes, not a model baseline.
What it does when you ask
The same five panels on both cars. The percentage is paint pixels at light-source brightness; a* is how red the panel is, where 0 is neutral.
One output with a reflection request reached 1.452% bright pixels in the sampled areas. The hood also became redder, with its measured CIELAB a* value rising from 0.72 to 9.31 in the compared patches. That is consistent with more visible red in this output; it does not establish the physical origin of the color.
But look at which panels moved.
The door changed little in this pair, whereas the hood became much redder, despite not being named in the instruction. Inspect each generated panel before using the image in a car visual. If the placement matters, revise the prompt or edit the output and check it again.
Once it moves, is it a reflection or a decal
A still can show highlight placement but cannot show how it changes as the viewpoint moves. A short image-to-video clip offers another check, though generated motion cannot prove that the surface obeys real-world reflection physics.
Five seconds, viewpoint moving right, nothing in the scene moving. Generated from the still with the reflections in it.
I compared two regions over the hood during the opening part of the clip. They were aligned by tracking a nearby mirror, which approximated the movement only while the selected boxes visibly stayed on the panel.
The red window's patch of paint goes from 8.76% to 0.88% in a second and a quarter. The blue control stays at 0.00% the whole time.
One selected hood region fell from 8.76% to 0.88% bright pixels while the whole hood stayed within 1.4–3.0% across those frames. This is consistent with changing highlight placement in the generated sequence. It does not establish whether the model simulated a physical reflection or simply animated the highlights.
Past a second and a half I stopped trusting the windows: the car turns as well as translates, so a translation-only lock walks off the panel. The numbers above come from the interval where I could see both windows still sitting on the hood.
Test a portrait reference and scale
A generated portrait was saved and used as a reference for a later frame. The person remained recognizable in that example, and the camera appeared closer to eye level. If you want to use a real person’s likeness, use your own image or obtain their permission.
Same room, same car description, with a portrait attached as a reference. This is the only frame of the five that came out at eye level.
One quick visual check is where the car roofline appears against the person. In this frame it meets about 65% of the person’s image height. A simple real-world height ratio such as 1.30 m divided by 1.75 m is about 74%, but it cannot be used as the expected image ratio without accounting for perspective, pose, camera angle, and vehicle size. Treat the frame as a visual prompt to check scale, not a measured nine-percent error.
The roof sits at a different depth from the person, and the generated vehicle has no defined real-world dimensions. Compare several visible cues before deciding whether the scale needs correction.
This was also the only eye-level frame in this small set and the only one with an unrequested red hood. Because the portrait reference and camera angle changed together, the cause cannot be isolated. A later request for eye level did not produce it in one successful retry; that is an observation about this attempt.
Asking for black and white
The third complaint about this look is that it falls apart in monochrome. So I sent the plain prompt again with one word added.
The prompt was the color one plus “black-and-white”. Everything went neutral except the wall.
The red wall retained a positive a* value in the black-and-white request, while the sampled neutral areas remained near zero. In this result, the instruction did not remove the wall’s color. A direct desaturation adjustment in CapCut gives you a way to inspect and correct the final image.
If you need a monochrome deliverable, check the generated image and adjust saturation in the editor as needed.
What to do with this
Where this stops
This example includes one setting, a few stills, and one short clip. The reflection instruction was tested in one paired comparison, so it cannot estimate how often placement fails. Brightness and color readings support observations about these selected outputs; they do not establish how the image model represents physical reflections.
Lighting, paint color, scene composition, and camera position may change the result. Test the intended setup and inspect the final export.
Check the panel, then finish the image
The practical sequence is to describe the lighting and desired reflection, generate a draft in CapCut, inspect the named body panels, and correct the strongest mismatch before exporting. In this one pair, the red hood deserved attention even though the prompt named other panels. Finish by checking the color treatment and the person-to-car scale in the full frame.