The cakeify, squish, and Ghibli effects mostly run on one image-to-video loop, not a different magic app each time
Motion control copies movement from a reference clip, which is how the dancing-figurine trend scales to thousands of posts
I upscale every clip in Magnific before export so the effect survives the platform recompress
A 6 second effect clip takes me about 15 minutes end to end, and most of that is picking the start image
Every week a new AI video effect takes over my feed. One month a knife slices a sneaker open and it is sponge cake inside. The next, everything is a soft jelly version of itself that squishes and springs back. The effects look like magic, and they look like a different app every single time. They are neither. Almost all of the viral 2026 effects run on the same short pipeline. Once you can see the pipeline, you can make any of them in an afternoon.
The effect is the recipe, not the app
The biggest myth is that there is a "cakeify app" or a "squish app" you download. There is not. What you are watching is four things in a row: a start image, an image-to-video model, a tight prompt that describes the transformation, and an upscale pass at the end. The viral name is just packaging wrapped around that loop.
The models doing the heavy lifting are the ones you already know. Kling, Runway, Hailuo, Veo, and Magnific's video models all take a single still and animate it from a text instruction. I keep Magnific in the rotation because the same workspace does the image, the motion, and the upscale, so I am not shuttling half-finished frames between four browser tabs and losing quality at every hop.
The skill is not the tool. It is two things: the start frame and the instruction. A clean start frame with one clear subject on a plain background animates predictably. A busy frame turns to mush halfway through the clip. I learned that the slow way, after a dozen takes where the background melted and dragged the subject down with it.
There is a quiet third ingredient too, and it is motion amount. Most tools let you dial how much the model moves the frame. Crank it for a dramatic transformation, pull it back for a subtle one. People leave it on the default and then wonder why their clip either barely moves or thrashes around like it is possessed. It is the first slider I touch, before I even fix the prompt.
If you want the model comparison itself, I put real numbers on each one in best AI video tools for short-form content. This piece is about the part those comparison posts skip: what you actually type into the box.
Cakeify, squish, and the satisfying-destruction family
The "is it cake" effect, the squish effect, the inflate effect, the crush effect. They are one family. You are telling the model to apply a physics transformation to a solid object while keeping that object recognizable the whole way through.
The prompt skeleton I start from is always the same order. Subject, then transformation, then camera, then a physics word. For cakeify: "a realistic [object] on a plain plate, a knife slices through it revealing moist sponge cake layers inside, slow motion, macro, satisfying." For squish: "a [object] made of soft memory foam, a hand presses down and it compresses then slowly springs back, studio light, macro." Swap the object, keep the structure.
Two details decide whether it reads as viral or as a broken render. First, the start frame has to be a hero shot of one object, centered, plain background, good light. Second, the transformation needs a physical verb the model actually understands: slice, compress, inflate, crush, melt. Vague verbs give you vague results.
I run the same prompt three or four times with a fresh seed each time. The model is a slot machine, and you are paying for pulls. One pull in four is genuinely good, two are usable, one is cursed. Budget for that rather than expecting take one to ship. The people posting daily are not luckier than you, they just throw away more clips than they keep.
One more lever almost nobody pulls: keep these clips short. Three to five seconds is the sweet spot. The transformation is the whole point, so you want to land on it fast and loop clean. Longer clips give the model room to wander, the object drifts, hands sprout extra fingers, and the loop stops being satisfying. If an effect needs more time, I cut two short takes together instead of asking for one long render. The inflate variant is a good example: "a [object] slowly inflates like a balloon until it is round and tight, then holds, studio light, macro" loops beautifully at four seconds and falls apart at ten.
Ghibli, toy figures, and motion control
The painterly trends are a different move. For a hand-drawn look you style the still first, then animate the stylized still. Animating a raw photo and hoping it comes out painted fights the model the whole way. Paint the frame, lock it, then bring that frame to life.
The figurine trend is the most interesting one technically. You turn a person into a collectible figure inside the blister pack, then you make the figure move. The movement is the actual trick. Motion control, which Kling pushed hard this year, lets you upload your character image plus a short reference video, and it copies the motion from that reference onto your character. That is how thousands of near-identical dancing-figurine clips show up in a single week. One dance take, infinite characters.
Keeping the same face across a whole series is its own headache, and it is the difference between a one-off and a character people actually follow. I went deep on that in how to create consistent AI characters for your brand. The still-image cousins of these effects, the toy-box look and the rest, are moving just as fast on the photo side, which I broke down in viral AI photo trends of 2026.
My actual 15-minute workflow
Here is the whole thing, start to posted.
I pick or generate the start image first. Plain background, one subject, sharp. If I am generating it I make four options and choose the cleanest, because every flaw gets amplified the second it starts to move. A tiny artifact on a still becomes a flickering mess in motion.
One framing note that saves reshoots: build for vertical from the very first frame. I generate the start image at a tall ratio, or I generate square and leave headroom so the crop does not slice the top of the subject off. I also keep the action in the middle third, away from where the caption sits at the bottom and the profile buttons sit down the side. Nothing stings like a perfect reveal that happens directly behind a follow button.
I write the transformation prompt using the skeleton above and run it three times. While those render I draft the caption and the hook text. I pick the best take, then upscale it in Magnific to 4K. This step matters more than people expect. Every platform recompresses your upload on the way in, and a soft clip turns to porridge in the feed. A crisp 4K source survives that squeeze and still looks sharp on a phone.
Sound is half the effect. A satisfying clip with no audio is just a clip. I either pull a trending sound or build a quick layer myself, and for voice or custom effects I use ElevenLabs. The audio side is a craft of its own, which is exactly why I gave it a separate breakdown in the viral AI sound.
Then I batch. When one format is hot I make five at once and space them out with Buffer instead of dumping them in one clump. Trends have a window measured in days, and five shots inside that window beat one perfect clip posted a week too late.
Bottom line
A viral AI effect is a format, not a tool, and formats expire fast. The creators who win are not the ones sitting on a secret app. They are the ones who spot the recipe early, batch a handful of clips while the trend is still fresh, and move on before it goes stale. Learn the loop once, start frame, instruction, motion, upscale, sound, and every new effect becomes a variation you already know how to cook.
I keep a running list of which effects are actually landing, plus the settings behind them, in the Lab. If you want the current shortlist and the workflow notes, start at the Lab overview. The next effect is already on its way. Better to have the kitchen ready before it lands.
This article contains affiliate links. If you sign up through them, I may earn a small commission at no extra cost to you. (Ad)
Top comments (0)