Articles

Why Does My AI Video Look Bad?

Skin, grade, and the log footage trap

By Raymond Karsten · Published 12 August 2026 · Last updated 12 August 2026

You fix one thing and something else breaks. Here is what is actually going wrong.

Problem one — the camera block is a shopping list

Most prompts name a camera, a lens brand, a sensor and a codec. None of that is visual information. A model has never seen a badge; it has seen images.

Describe the photograph, not the equipment.

Here is the block almost everyone writes:

ARRI Alexa 35, Signature Prime 40mm, ARRIRAW,
LogC4, shot on set, cinematic, 8K, ultra detailed,
professional colour science, film look

And here is the same intent written as an image. Same length, entirely different result.

40mm, eye level, half a step back from her.
Hard side light from a window camera-left, the far
side of her face falling into shadow. Background
two stops darker than her skin. Fine haze in the
air catching the light.

Cut: body names, codecs, resolution claims, “professional colour science”. Keep: focal length, height, light direction, contrast between subject and background.

Problem two — the grade is missing, not the detail

Before you regenerate anything, put the clip on a timeline and add contrast and a black point. Newer video models deliver a deliberately flat image, and flat reads as cheap even when the detail is all there.

Do this first

Set the black point so the darkest part of the frame is actually black. Lift contrast until skin separates from the wall behind it. Five minutes. Half the time the “plastic” feeling lifts and you never touch the prompt.

Problem three — you asked for log footage

A video model cannot produce log. Log is a recording curve that preserves range you can pull back later. The model has no range to preserve — it paints the appearance of log, which is flat and grey, and hands you none of the latitude.

shot in LogC4, flat profile, soft highlight
roll-off, wide dynamic range, ungraded, ready
for colour grading

Every line above lowers contrast in the delivered frame. “Soft highlight roll-off” is the worst offender — the model cannot roll off highlights it never captured, so it drains contrast out of the whole image instead.

Ask for the finished look instead:

Deep true blacks, clean bright highlights on the
window frame, warm skin against a cool grey wall.
Contrast already in the image. Final grade, not
a flat profile.

Problem four — nothing in the prompt describes skin

Say nothing about skin and the model averages toward the middle of everything it has seen: smooth, even, poreless. That average is what your eye calls plastic.

Describe the surface, not the person

Visible skin texture. Pores across the nose and
cheeks, fine lines at the eyes, uneven tone with
faint redness on the cheekbones. Oil sheen in the
highlights only. Matte forehead. Tiny hairs catching
the side light along the jaw.

Evenly matte reads as plastic. Evenly shiny reads as sweat. You want the sheen in the highlights and nowhere else. Add the negatives too:

No airbrushing. No beauty retouching. No smoothing.
No plastic sheen across the whole face.

Model version matters here as well — flatter, newer checkpoints remove the micro-contrast that reads as living tissue. I covered that trade-off in Seedance 2.5 vs 2.0.

See how this fits the full pipeline

Problem five — everything is equally sharp

Real lenses fall off. When the wall behind the actor is as crisp as her eyelashes, the brain reads render, not photograph.

Shallow depth of field. Focus on her near eye,
her far ear already softening, the background
dissolved into shapes. Foreground edge out of
focus in the bottom of the frame.

Problem six — the air is empty

Light needs something to travel through. Without it you get clean, dead space between subject and background.

Dust in the air. A shaft of light crossing the room
from the window, visible where it passes through the
haze. Faint smoke low in the frame. Grain in the
shadows, not the highlights.

If you cannot see the light, you cannot feel the room.

The lesson

Plastic skin and flat grey footage are prompt problems, not model problems. The camera block is where the damage happens, because it is the one block people fill with words they never had to define.

Write what the image looks like. Grade before you regenerate. Name the surface, the light and the falloff, and let the badges go.