Key takeaway — An AI video looks like it was "shot on different days" because people, light, and color drift a little from shot to shot. ⑪ Bind a person by the traits that must match ⑫ Call the camera by size, angle, and movement ⑬ Nail the light and the hour in one line ⑭ Cut down to one base color and one accent ⑮ Instead of "don't," say what fills the space.
This is part 4 of our AI ad video directing series. Part 3 split the prompt into slots; this time we look at what to write inside them so nothing drifts when the shot changes. This is the third group of the xbrush Academy's 20 Tips.
Tip 11. Bind a person by what has to match
Writing "the same woman" does not hold her together. The model doesn't carry a person as an identity; it carries a bundle of traits. So listing the traits that must match holds far better than asserting sameness — and when a shot drifts, you can point at which trait drifted.
Pick three or four things to lock: hair length, coat, glasses, skin tone.
Repeat those traits in full in every shot. Don't abbreviate.
If the face matters, an image beats any description.
Weaker | Stronger |
|---|---|
Shot 2 has the same woman in it. | Shot 2: same person as shot 1 — shoulder-length black bob, beige trench coat, gold-rim glasses. |
Real example — lock the position, not just the look
The seaside dialogue scene in the Higgsfield guide lists each character's traits — and builds into the woman's definition that she is always on the man's left. It's a guard against the two swapping places mid-conversation.
THE WOMAN, always on the man's LEFT side: mid-20s, slight build, long chestnut wavy hair
whipping loose in the wind, light freckles across the nose, hazel eyes, ...Source: Higgsfield Blog (higgsfield.ai)
Some prompts lock a person with text alone. A 30-second home-video prompt on Threads describes the hair, stubble, and stray strands in detail, then closes by listing what must hold: "the same face, hairstyle, clothing, body proportions, and appearance throughout the entire video."
Source: Threads @m_faixan43 (threads.com)
Tip 12. Call the camera by its name
"Make it look cool" is not an instruction. "Low-angle medium shot" is. Camera terms are vocabulary the model has actually learned. Naming narrows the field, and when the result is wrong, you know which name to change.
State size, angle, and movement separately — medium shot / eye level / locked off.
If nothing moves, write "locked off." Leave it blank and it drifts.
Can't recall the term? Find it in the Film Grammar area of the Concept Map and use that exact word.
Weaker | Stronger |
|---|---|
Show the product from a stylish angle. | Close-up, 15° low angle, locked off. Product label centered in frame. |
Real example — size, field of view, angle, movement in one line
A K-pop convenience-store lookbook prompt published on Notion opens every cut with shot size → field of view → angle → camera movement, always in the same order.
00.0s CUT 1 — EWS full body, 84° wide, low-angle fast dolly-in rushing toward her.
01.2s CUT 2 — WS full body, 63°, low angle, whip-pan catching her.
04.4s CUT 5 — ECU face, 18°, snap zoom in.Source: 존코바 on Notion (notion.site)
Higgsfield's noir example goes a step further and locks the lens field of view per segment (LENS LOCK SEGMENT 1: 84°, SEGMENT 2: 29° …). With the field of view fixed as a number, perspective doesn't shift between segments shot from the same distance.
Tip 13. Nail the light and the hour in one line
If the sun moves between shots, it stops reading as one film. Light is the strongest continuity signal there is. Let the hour drift and the same room with the same person looks like a different shoot day. Hold the light steady, though, and even a loose cut still reads as one piece.
Give the hour and the source together — "late afternoon, side light from the window."
Fix one shadow direction and carry it through every shot.
Moving between indoors and out? Give the transition its own shot.
Weaker | Stronger |
|---|---|
Warm and bright, please. | 4pm, side light from the window on the left. Shadows fall long to the right. |
Real example — even changing light gets one line per state
What about a film where time passes? Higgsfield's beach-party documentary example runs from afternoon to dawn across five segments — so it names one light source per segment. The light changes, but not by accident: every change is directed.
LIGHTING, five states, one coast: Segment 1, high afternoon sun, hard and almost overhead.
Segment 2, sun dropping, light warming and raking in low.
Segment 3, night, the bonfire is the dominant source, low, warm, orange, unsteady. ...Source: Higgsfield Blog (higgsfield.ai)
For scenes with many sources, like neon, you can bind them with a color-temperature number. The convenience-store lookbook above fixes the white balance with WB 4000K mixed neon.
Tip 14. Cut down to two colors
One base, one accent. Past that it's noise, not direction. Name several colors and the model keeps whichever is loudest and drops the rest — and which one survives changes every run. Two colors both survive, and every shot starts looking like it belongs to the same film.
Base color on the large surfaces, accent on exactly one small thing.
If there's a brand color, make it the accent.
Never write "colorful" — that decides nothing.
Weaker | Stronger |
|---|---|
A bright mix of blue, orange, and green | Base: muted blue-grey. Accent: orange — on the cup only. |
Real example — set the wardrobe color against the product
A 30-second fragrance commercial prompt shared on X picks the model's gown color and writes down why it was chosen: one warm color set against the cool tones of the bottle.
sunset-coral/burnt-orange halter gown — a color chosen to sit in warm contrast
against the bottle's cool blue-to-pink paletteSource: X @ShamiWeb3 (x.com)
A public-institution infographic example follows the same principle: restrained blue-and-teal palette ties the color to two families, and "restrained" keeps it from growing.
Source: Threads @mayjun_sydney (threads.com)
Tip 15. Say what fills the space, not what to leave out
A negative instruction invites the word onto the screen. Write "no text" and the model reads "text." Describing what occupies the space is far more reliable than describing an absence.
Whenever you want to forbid something, say what goes in its place instead.
If it truly must go, phrase it as a presence: "bare wall," "plain backdrop."
Something keeps showing up? Assign a different prop to that spot.
Weaker | Stronger |
|---|---|
No text or logos. And no people either. | Background is a bare white wall. One product sits in frame, nothing else. |
Real example — POSITIVE LOCKS instead of NEGATIVE
Among the published prompts, three used a separate NEGATIVE block, while eight used a POSITIVE LOCKS block to state in positive terms what must hold. Korean YouTuber 머니스웨거's pirate-ship one-take prompt is a good example. Rather than saying "don't draw the extras' faces," it says they "stay soft-focus and faceless."
POSITIVE LOCKS
One unbroken take, one camera path as written — track then rise, never orbit.
... Background sailors stay soft-focus and faceless. The fleet appears from 15s on.Source: 머니스웨거, Google Doc (docs.google.com)
For videos that start from a single image, you'll also see the target stated as a presence: First frame matches the provided image exactly (Threads @hamlog_ai).
Try it on xbrush
Add these five to the template from parts 2 and 3 and it looks like this. The only lines that changed are the locks.
[Subject] Woman in attached image 1 — shoulder-length black bob, beige trench coat, gold-rim glasses (same in every shot)
[Space] Café window table, white wall
[Camera] Shot 1 medium shot · eye level · locked off / Shot 2 close-up · 15° low angle · locked off
[Light] 4pm, side light from the window on the left. Shadows fall to the right (same in every shot)
[Color] Base: muted blue-grey. Accent: orange — on the mug only
[Background] Bare white wall; only the mug on the tableWhen a result drifts, find the slot that drifted and fix only that one. For hands-on practice, L8 (an ad video with Cinema) in the Academy's 10 Hands-On Steps is the closest match to this group.
Next up, the final group: sound, dialogue, and the ending.
Frequently Asked Questions
How do I stop a character's face from changing between AI video shots?
Instead of writing "the same person," choose three or four traits that must match — hair length, outfit, glasses, skin tone — and repeat them in full in every shot. If the face matters most, attaching a reference image is the most reliable option.
How should I write camera directions?
Name the shot size, camera angle, and camera movement separately, for example "medium shot, eye level, locked off." If the camera shouldn't move, write "locked off" rather than leaving it blank, or the camera will drift.
Why write the same lighting for every shot?
Light is the strongest signal that ties shots into one film. If the hour or light direction changes between shots, the same person in the same place looks like a different shoot day. If time is meant to pass, name the light source for each segment so the change is directed too.
How many colors should I specify?
Two: one base color and one accent. Name many and the model keeps only some, and which ones survive changes each run. If you have a brand color, make it the accent and use it on a small area only.
Should I avoid negative prompts?
Some prompts use them, but writing "don't" often brings that very word onto the screen. If you want something gone, describe what is there instead, in positive terms — for example, "a bare white wall" instead of "no text."