Stop faking realism: when a declared style beats a bad photoreal
Photorealism is the hardest bet you can make with a generator and the one where failure is most visible. A declared style moves the target somewhere the model can reach, and the viewer forgives everything because you told them it is a drawing.
What is in here
Photorealism is the hardest bet you can make with a generator, and the one where failure is most visible, because your viewer has spent a lifetime looking at real things. A declared style moves the target to somewhere the model can land: papercut, cutout collage, type-led graphic, illustration, miniature, editorial still-life. The viewer forgives every departure from reality because you told them in frame one that this is a drawing. Our rule: faked realism is a kill, a declared treatment is legal, and the declaration happens on paper before anything renders.
- Why photoreal is the most expensive bet in generated video, and what your viewer is grading it against
- The kill that wrote our physical-plausibility law, and the narrowing that came later
- Six declared styles that survive a feed, and what each one buys
- How to declare a treatment so it reads as a decision rather than a failure
- The two-minute test, and why a declared style is the cheapest way to pass it
Four numbers this post runs on
Measured in this studio01How do you fix the uncanny valley in AI video?
Stop competing with reality. The uncanny valley is a comparison problem: the closer a synthetic image gets to real, the more brutally the eye grades what is left over. Move the work into something openly not real, drawn or cut or printed or built at 1:24 scale, and there is no gap left to grade.
What each bet has to get right
Two bets| Dimension | Photoreal | Declared style |
|---|---|---|
| What the viewer compares it against | A lifetime of real hands, real light, real objects | Yes. The rules your own frame one set |
| Cost of one wrong frame | No. The whole ad. The eye finds it in under a second | A drawing that got looser for a moment |
| Where failure shows up | Hands, contact, weight, reflections, labels, continuity | Yes. Inconsistency inside a system you defined |
| What the model is being asked to do | Simulate a world it has never touched | Yes. Repeat a look it can hold |
| How ownable the result is | Every brand in your category has the same photoreal | Yes. The style belongs to one brand and travels with it |
| Time to replicate it with a phone and Canva | No. About two minutes, if it is a photo with type on it | Longer than the client has, which is the point |
Masahiro Mori described the curve in 1970, about robots, and the shape is the part worth stealing. Affinity climbs as a thing gets more humanlike, falls off a cliff just short of real, then recovers on the far side. Nearly everything a generator makes lands in the ditch. There are two ways out and only one of them is cheap. You can climb out the far side, which means solving hands, contact, weight, reflection and continuity at once, on a Tuesday, on budget. Or you can walk back down the near slope on purpose and stand somewhere the eye is not measuring.
Four things people believe about stylized ads
Flip them02The kill that wrote the rule
We once built two halves of a photographed loaf pulling apart on screen, cut out properly and color matched and passing every automated gate the engine had, and it was killed on sight because that is not really how two pieces of bread separate. The eleven physics tests that came out of it are a page of their own.
A photograph holds exactly one state. Ask it to perform an event and somebody has to invent the geometry the camera never recorded: the torn edge, the crumb, the way a crust gives before it splits. The eye reads invention immediately, even when it cannot say what it just saw. Our whole failure atlas of generated footage starts at that one note.
One kill, and the law it turned into
Click along itEverything measurable was correct
Clean alpha, a shadow derived from the element's own silhouette rather than a stock blur, matched grade, one grain plate over the whole composite. Every automated check passed.
Which is the useful part. No metric in the system had any opinion about whether the event was possible.
Cost of a change here: one renderA material-behavior note, not a craft note
The kill reason was about how bread behaves, not about how the shot was made. Nobody mentioned the shadow, the grade or the grain.
Cost of a change here: the whole beatA still may move as a card and nothing else
Push, settle, drift, a light patch traveling across the surface. Anything that looked like a physical event was banned outright.
It held the line and it was too blunt. It also banned papercut, collage and every other treatment that never claimed to be real in the first place.
Cost of the over-correction: every stylized ideaThe kill is faked realism, not motion
A still may split, tear or pour when the treatment is deliberately stylized, reads as designed animation, and is declared in the direction lock before render.
Realism still comes from real footage or true video generation only. Card moves stay the safe default.
Cost of declaring it: one line, written before anything rendersThe kill was never the motion. The kill is a photograph pretending the event was real.
our physical-plausibility law, after it was narrowed
A declared composite, seen from the side
Layer anatomy03What counts as a declared style?
A treatment you commit to in writing before you render, that a stranger can identify inside one second, and that holds for the whole film. Papercut, cutout collage, type-led graphic, illustration, miniature set, editorial still-life. The test is simple: would somebody call it a drawing, or would they call it a bad photograph?
Six declared styles, and what each one buys
Switch between themLayers of paper, lit and moved
Elements read as cut sheets with visible edges and honest drop shadows. Because paper genuinely tears and folds, this is the one style where a split or a peel is not a lie.
- Buys you: physical events without faked realism
- Costs you: product texture, which has to come back in a photograph somewhere
- Watch for: edges too clean to read as paper
- Good for
- Process stories, before and after, anything that transforms
- Bad for
- Food appetite, skin, sparkle
Real photography, arranged as objects
Photographed elements kept whole, laid on flat grounds, moved as cards. The one we reach for most, because it keeps real product texture and never asks a photograph to become something it is not.
- Buys you: real product, plus complete control of composition
- Costs you: the sense of a place, unless you build one
- Watch for: elements floating with no contact and no shadow
- House edge spec
- 1.6px feather, no keyline, nothing else
- Grounding
- Shadow offset as a fraction of the element's own width
The words are the picture
One word per cut, footage as ground rather than subject. Because each cut is a word, the cutting acquires grammar, which is exactly the fix for a chain of clips that do not match.
- Buys you: coherence out of heterogeneous footage
- Costs you: nothing visual, and everything if the writing is weak
- Watch for: weight. An ultralight face at small sizes has a sub-pixel stem and dissolves
- Legibility floor
- Cap height at least 5% of frame height
- Dwell floor
- 1.2s, or 0.35s per word, whichever is longer
Drawn, and drawn consistently
The furthest from the valley and the hardest to hold across shots. Generators drift between renders, so the discipline is a small locked vocabulary: one line weight, one palette, one horizon.
- Buys you: total permission. Nothing has to obey physics
- Costs you: consistency, which becomes the whole production problem
- Watch for: per-render drift between shot three and shot four
- Also read
- Per-render drift
- Best used with
- One real photograph at the payoff
A tiny world, declared as one
Sets, models, tilted focus, oversized hands entering frame. It reads as made rather than captured, which is the permission. The scale cue has to be deliberate in every shot or the illusion inverts.
- Buys you: charm, and an excuse for impossible camera positions
- Costs you: credibility for anything that has to look clinical
- Watch for: an object rendered at the wrong real-world size
- Related failure
- Scale anchoring
- Rule we carry
- Every landmark carries its real dimension beside the person
Assembly, not a photo with type on it
Deconstructions, ingredient arrangements, multi-element compositions on a designed ground. The only one of the six that stays fully photoreal, and it survives because the craft is in the arrangement rather than in a simulated event.
- Buys you: premium, and real product truth
- Costs you: real assembly time, which is the cost that shows
- Watch for: the version that is one photo with a headline over it
- The bar
- Visible craft a client cannot repeat in two minutes
- Banned output class
- Photo plus typography, at any level of polish
Notice what they have in common. Each one gives the viewer a rule in the first second, and every frame after that is judged against the rule instead of against the world. That is the entire trade, and it is why the choice belongs to the brief rather than to the grade.
The words we use for this, defined
Search the vocabularyDeclared treatmentDirection
Card moveCraft
Faked realismKill class
Deviation registerProcess
The two-minute testStandard
True-depth parallaxCraft
04Declaring it is a production step, not a mood board
A style declared after the footage exists is a filter. A style declared before it changes what gets generated at all: which shots are worth asking for, which are impossible, which the treatment makes unnecessary. We write it into the direction lock, next to the beat durations and the committed soundtrack, and a build that departs from it without saying so is killed regardless of how it looks.
The declaration, as it actually appears in a direction lock
Copy this shapeTREATMENT: papercut, hard-edged, visible paper grain on every element DECLARED AT: direction lock, before any render LEGAL EVENTS: tear, fold, peel, slide - because paper does these STILL ILLEGAL: any event that reads as photographed reality CONSISTENCY: same edge treatment, same 3 paper stocks, all 14 shots DEVIATION REGISTER: any departure written down BEFORE it renders, as what / why the precedent does not cover it / risk accepted
The deviation register is the part that does the work. Every kill we have logged, the loaf included, was an undeclared departure from something the reference material already demonstrated. Writing the departure down before rendering costs a minute and turns an argument about taste into an argument about a decision somebody made on purpose. That is also the whole case for approving paper before pixels.
Where the drawing works and where the photograph works
Second by secondRead it as a list
| At | Channel | What happens |
|---|---|---|
| 0.0s | Declared layer | Paper tears open the frame |
| 2.2s | Declared layer | Type assembles, one word per cut |
| 5.6s | Real footage | The event, photographed for real |
| 8.4s | Declared layer | Cut-paper transition carries the turn |
| 11.0s | Real footage | Product, held, no treatment on it |
05Why does my AI ad look cheap even when nothing is broken?
Usually because it is a photograph with type on it. Ten technically clean statics of ours came back scored 50 out of 100 against a bar of 85 and were killed in a single verdict, on the grounds that a buyer who believes they could rebuild the thing in two minutes will not pay thousands for it.
The rule it produced is the two-minute test: every pitch creative has to contain visible craft the client could not replicate in two minutes with a phone and a free design tool. Composition assembly counts. Deconstructions, ingredient arrangements, multi-element still-lifes, an integrated design system running across a set. A declared style is the fastest way to clear that bar, because the treatment itself is craft that shows in the first frame, and the beauty ads everyone has already seen is the category where that bar bites hardest.
Assembly, not a photo with a headline over it
Ours, made this way06When photoreal is still the right bet
When the thing being sold has to be verifiable. A hotel room has to be that room. A serum has to be that bottle. Food has to look like it would taste of something, and no drawing has ever made a stranger hungry the way a good macro does. In those jobs the answer is not a style, it is real footage of the real thing plus generation for coverage around it.
The split we use: photograph the twenty seconds holding the physical event and the product truth, then generate the environments and the variants nobody was going to pay to shoot. If neither is possible, declare a style. Never split the difference. A half-stylized photograph is the worst of both, and it is the frame the eye goes to.
Four films, four different answers to the same question
Ours, built this wayPhotoreal or declared: which bet is yours?
Answer four07A declared style becomes a fingerprint
A declared style buys you a lot and it is not a free pass. It becomes a fingerprint: run one treatment across a set and the pieces start reading as one piece, which is how we once retired eight finished ads that differed on cut rhythm and matched on argument, world and point of view. The full account is in what makes an ad resist fatigue. A treatment also hides weak writing behind a nice surface for about four seconds, and the type floors do not relax because the ground is now a drawing.
The part we cannot prove
We have no performance data showing a declared style outperforms a good photoreal in market. What we have is a kill log. Fewer than half the films that clear every automated gate here survive a human eye, and the ones killed for looking fake were overwhelmingly the ones trying hardest to look real.
Declare a style before your next render
Tick as you go - it remembersQuestions people actually ask
Open what you needHow do I fix the uncanny valley in my AI video?
Move away from realism rather than toward it. The valley is a comparison effect, so the fix is removing the comparison: commit to a treatment that reads as designed, declare it before you render, and hold it consistently across every shot. If you cannot leave realism, then do not generate the physical event at all. Shoot the two seconds where something actually happens and generate the rest around it.
Do stylized ads convert as well as realistic ones?
We do not know, and neither does anyone selling you an answer. There is no published study we could find that measures visual treatment against in-market performance for short-form ads.
What is measurable is failure. A photoreal frame that gets one physical event wrong loses the viewer in under a second, and that is a cost you can see in a hook rate. A drawing cannot make that particular mistake.
Why does my AI ad look cheap even though nothing is obviously wrong?
Most often because it is a single photograph with type over it, which any buyer can picture themselves making in two minutes. Cheapness reads as absence of decisions rather than absence of polish. Add assembly: build the composition instead of dressing an image, and make the treatment visible in the first frame so the viewer knows a person chose something.
Can I mix a declared style and real footage in one ad?
Yes, and it is the most reliable structure we use. Keep the border clean: the stylized layer carries the argument and the transitions, the real footage carries the product and any physical event. What fails is the blurred middle, where a photograph is nudged just far enough to look processed but not far enough to read as designed.
What do I write in a brief to declare a style?
Name the treatment, name the events it legally permits, name what stays illegal, and state the consistency rule that makes it hold across shots. Four lines. Then add a place to write down departures before they render, because an undeclared departure is the failure mode, not the departure itself.
Does a papercut or collage treatment look dated?
It can, if it arrives with the whole kit of tape, staples, torn edges and halation. We deliberately leave those out. The treatment that survives is usually the restrained one: one edge quality, two or three materials, and a palette borrowed from the brand rather than from a craft supply shop.
If your last three ads all tried to look real and all felt slightly wrong, the cheapest move left is to decide on paper what kind of object your ad is going to be, and then to be that object without apology for the full fifteen seconds.
Where the numbers came from
- IEEE Spectrum. The Uncanny Valley: The Original Essay by Masahiro Mori - the 1970 curve, in the authorized translation
- Wikipedia. Uncanny valley - for the history of the effect outside robotics
- W3C. Understanding Success Criterion 1.4.3: Contrast (Minimum) - the 4.5:1 ratio our own legibility gate is built on
Every figure above links to the place it was published. Numbers marked as ours are measured inside this studio and we say so where they appear. We do not print a statistic we cannot point at.
Send a link. Get one finished ad back.
One finished cut built from your own product, inside three days, free and yours to run whether or not we ever work together. Tell us if you want it photoreal or declared, and we will tell you honestly which one your product should be.
Replies within a day. Ad within three.