Most creators keep one long negative prompt and paste it into every generation. Feels efficient. But it works against you. A close-up of a face doesn’t fail the way a wide landscape fails. A product shot on a white table doesn’t need the same protection as a street scene full of signs. Protect everything equally, and you protect nothing well — the model spreads its attention across twenty warnings and half-ignores each one.
The fix I use in my own Seedance 2.5 and Veo work is simple: negatives attach to the shot type, not the project. My storyboards already label every panel — face shot, product shot, wide, text-risk — so the right negative set rides along with the board.
One Negative List Does Not Fit Every Shot
Think about what actually ruins each kind of footage. A portrait dies from extra fingers, drifting eyes, skin turned to plastic. A product shot dies from a warped logo, a bent bottle, reflections that invent a second object. A landscape dies from wobbling power lines, melting trees, a horizon that breathes. Three different failure families. They need three different warnings.
A universal string like “morphing, blurry, deformed, extra limbs, text, watermark” covers a little of everything and nothing deeply. In practice I run a short base string everywhere — flicker, jitter, warp, watermark, low quality — then add a shot-specific module of four to seven terms. That keeps every negative prompt under a dozen terms, which is where models still listen to each word. The base handles the model’s global habits; the module handles the shot in front of you.
Second cost of the one-list habit: contradictions. Negating “text” in a shopfront scene is fine. But negating “people” in a café scene where your character waits tables actively fights your positive prompt. Shot-matched negatives avoid self-sabotage — because you only ban what this frame can plausibly get wrong. Want the broader theory first? My guide to negative prompts covers the universal artifact list. This article splits that list by job.
People Shots: Anatomy And Identity Negatives
People are where models spend most of their failure budget — because faces and hands are the hardest things to keep stable across frames. For any shot where a person is recognisable — close-up, medium, dialogue coverage — my negative module focuses on anatomy and identity drift.
A working people module looks like this: extra fingers, fused fingers, asymmetric eyes, crossed eyes, distorted teeth, face morphing, identity change, plastic skin, waxy skin. Notice what’s not in there: I don’t negate “wrinkles” or “pores,” because real skin needs texture — and banning it is how you get the mannequin look covered in my piece on over-negating. I also avoid negating emotions. “No smiling” style terms confuse expression control. The positive prompt should carry the emotion; the negatives only carry defects.
Full-body and walking shots get a movement anatomy swap: extra limbs, bent joints, sliding feet, leg merging. Hands deserve their own attention whenever they touch anything — holding a cup, opening a door, counting money. Contact points are where fingers multiply. If hands are central to the shot, I say so on the storyboard and budget an extra generation for that panel. No negative string fully solves hands in motion.
One more people-specific habit: keep the module identical across every shot of the same character. Changing negatives shot to shot introduces its own inconsistency — the model renders slightly different skin and features each time. Consistency of negatives is part of character consistency — the face you locked in shot one should meet the same rules in shot nine.
Product Shots: Shape, Label And Reflection Negatives
Product work fails differently. The bottle’s the right colour but the wrong shape. The label is almost readable — which is worse than unreadable. A reflection in the glass shows a room that doesn’t exist. Clients notice all three instantly. The product is the one object in frame they know by heart.
My product module: deformed shape, warped logo, misspelled text, gibberish letters, melted edges, duplicate product, distorted reflection. The shape terms matter most — “deformed shape” and “melted edges” catch the slow bend that bottles, cans, and boxes suffer during camera moves. Keep camera moves modest on product shots for the same reason: a slow push-in holds geometry far better than an orbit.
Labels need an honest decision before you generate. If the audience must read the label, generation is the wrong tool for that element. Shoot the real packshot or composite it in the edit, as covered in my clean-plate method for text and logos. If the label only needs to look plausible at a glance, negate “misspelled text” and “gibberish letters,” keep the product small enough in frame that nobody can read it, and never let it sit centre-screen in a close-up. Plausible beats almost-readable every time.
Reflections are the stealth failure. Glass, chrome, glossy packaging invent windows, hands, studio lights that were never prompted. Add “distorted reflection” on glossy products, and describe the environment simply in the positive prompt — “on a matte wooden table, soft daylight” gives the reflection something sane to reflect.
Landscape And Architecture Negatives
Wides fail quietly. No single frame looks broken — but across four seconds the treeline ripples, the building leans, the road markings crawl. These shots are usually backgrounds or establishing frames, so they get the least attention. Then they undercut every good shot cut against them.
My landscape module: warping horizon, melting trees, wobbling buildings, bent power lines, flickering sky, crawling textures, distorted perspective. Straight lines are the tell, so architecture gets special care: “bent lines, leaning walls, warped windows” for city shots. Water and foliage are the other repeat offenders; “rippling distortion” helps on lakes and rivers — though the stronger fix is a gentler camera move and a shorter clip.
Because establishing shots play first, I generate them at the same negative settings as the scenes they introduce, even though the failure modes differ. The module changes, the base string doesn’t. That keeps grain, sharpness, and flicker behaviour consistent across the cut — which matters more to the viewer’s eye than any single artifact. The audience reads the whole sequence, not individual frames. Consistency across the cut is what makes separately generated shots feel like one afternoon in one place.
Text-Heavy Shots: The Hardest Negatives
Shop signs, book covers, phone screens, menus — anything with lettering is where AI video is most exposed. Models produce alphabet soup: letters nearly right, words dissolving mid-shot, signage changing language between frames. Negatives reduce the frequency of this. They don’t fix it. Pretending otherwise burns credits.
The text module: gibberish text, fake letters, misspelled words, changing text, unreadable signage, warped logo. Use it when text is unavoidable set dressing — a street of shopfronts behind your character, a café menu board out of focus. Depth of field is your ally here: slightly blurred background text reads as atmosphere; sharp background text reads as errors.
When text carries information — a price, a title, a name — don’t negotiate with the model. Decide on the storyboard that the generated shot will be a clean plate: a blank sign, a blank screen, a blank label. Generate that, then add the real words in Premiere where they stay sharp, spelled correctly, and identical in every frame. That workflow gets its own full treatment later in this series, but the negative-prompt side is simple: for clean plates, negate “text, letters, words, logo” outright so the surface arrives truly blank and easy to track.
Assembling A Shot-Type Negative Kit
Here’s the kit structure I recommend building once and reusing. One base string, four modules:
| Shot Type | Negative Module Focus | Typical Terms |
|---|---|---|
| People / close-up | Anatomy, identity, skin | extra fingers, face morphing, plastic skin, asymmetric eyes |
| Product | Shape, label, reflection | deformed shape, warped logo, gibberish letters, distorted reflection |
| Landscape / wide | Geometry, texture stability | warping horizon, wobbling buildings, crawling textures |
| Text-heavy / clean plate | Lettering control | gibberish text, fake letters, changing text — or a full text ban for plates |
Assembly takes ten seconds per shot: base string, plus the module matching the storyboard label, plus at most two shot-specific terms for unusual risks — “duplicate people” for a crowd, “sliding feet” for a walk cycle. If your total passes twelve terms, cut the least likely failure rather than adding more. Dilution is a real cost; every extra term weakens the ones that matter.
Store the kit inside your reusable prompt template so it’s part of the document you already open for every project — not a note you have to remember to find. And prune it quarterly: terms earn their place by catching failures you actually saw in QC, not by sounding thorough.
A Worked Example: One Scene, Three Shot Types
Let me walk through a short café scene the way I actually board it — because this is where shot-type negatives stop being theory. Three shots: a wide of the café interior, a medium of my character sitting down with a coffee, a close-up as she reacts to something off-screen. Same location, same character, same afternoon light. Paste one negative string across all three and I’d be lying to myself that they carry the same risks. They don’t.
Shot one, the wide, gets the landscape-and-architecture treatment even though we’re indoors. My worry: geometry and background life. Tables that wobble, pendant lights that bend, other patrons who duplicate or melt into each other. So the wide carries terms like warped furniture, wobbling background, duplicate people, merging faces, flickering lights. I’m not negating fingers here. At wide distance nobody can count fingers — and spending terms on invisible failures just dilutes the ones I can see.
Shot two, the medium, is a people shot with a prop problem attached. She sits, sets the cup down, hands stay in frame — so anatomy moves to the front: extra fingers, fused fingers, sliding hands, face morphing. The cup gets one product-style term — deformed cup — because a bending cup at the moment of contact is exactly the small wrongness viewers feel without naming. I keep the café menu board in the background deliberately soft, so I don’t need a single text term in this shot.
Shot three, the close-up, is all face. The module tightens to asymmetric eyes, crossed eyes, distorted teeth, plastic skin, identity change — and I drop the crowd and furniture terms entirely. Same character, same scene, three different negative prompts. Each one short, each one aimed. When I review the three clips in Premiere, the failures that remain are honest ones I can judge — not a random scatter from one bloated string fighting itself. That, to me, is the real payoff. You stop guessing which of your thirty negatives helped, and you start knowing.
Common Mistakes When Splitting Negatives By Shot
A few traps show up every time someone adopts this system. The first: building modules so large that each one becomes the old universal list in disguise. If your people module has fifteen terms, you haven’t built a module. You’ve rebuilt the problem with a new label. Four to seven terms per module is the range that’s consistently worked for me.
The second trap: mixing defect terms with taste terms. “Ugly, boring, dull, low quality people” — the model has no shared taste with you, and those words just add noise. Ban defects you can point at in a paused frame. Nothing else.
The third trap: forgetting motion changes the shot type. A medium shot where the character stands and walks to the door is secretly an anatomy shot and a background-stability shot at once. When a panel involves travel across the room, I borrow two terms from the landscape module — usually warping background and sliding feet — and drop two face terms I no longer need. Modules are a starting kit, not a cage. The storyboard note for that panel is where I decide the borrow, before a single credit is spent.
Last one: don’t tune negatives per generation. Pick the module for the shot, run your planned takes, judge properly at the end. Rewrite the negative string after every single output and you’ll never learn whether the module or the luck did the work — and your library turns into superstition. Change modules between shots. Change them between projects when QC teaches you something. Mid-shot, hold your nerve.
The takeaway is plain: stop writing one negative prompt per project and start writing one per shot type. Your boards already know what each shot is. Let the negatives match. It’s a small organisational change, but it removes the most common reason good prompts still ship bad footage — protection aimed at the wrong failure, spread so thin it protects nothing.