skip to content
Navid Mesbah
فا
← Back

Deep Motion: Editable Motion Design With AI

/ 9 min read

Two Kinds of Motion

Generative video has become remarkable. New models produce cinematic camera moves, believable physics, and light that behaves like real light. For atmosphere, texture, and exploring ideas quickly, they open possibilities that didn’t exist a few years ago, and I use them with real pleasure.

Motion design asks for something different. A motion designer usually knows exactly what they want: this logo should scale from 92% to 100% over 400 milliseconds with a soft overshoot, land on the beat, and hold for one second before the tagline enters from the left with a slight delay. Then the client asks for the tagline to come in a little earlier, and the designer wants to change that one thing while everything else stays exactly as it was.

That kind of work depends on structure: elements you can select, timings you can adjust, curves you can reshape, all independently. Video generation and structured animation solve different problems, and both are valuable. Deep Motion at Charsook is our answer to the second one: AI that works directly with keyframes, timing, and relationships, so every result stays open to the designer’s hand.

What Motion Designers Actually Do

To understand the problem, it helps to understand the work.

Motion design is the craft of making things move with meaning. Title sequences, interface animations, explainer videos, logo animations, data visualizations, transitions between scenes. It sits somewhere between graphic design, animation, and filmmaking.

A motion designer’s daily work is mostly about time. Not just what happens, but when, for how long, and in what relationship to everything else. They think in milliseconds and frames. They care whether a movement accelerates or decelerates, whether it overshoots and settles or arrives precisely, whether three elements enter together or one after another with a slight stagger.

These decisions are small individually but enormous collectively. The difference between an animation that feels premium and one that feels cheap is almost never one big choice. It’s hundreds of tiny timing and easing decisions, each slightly better.

And critically, motion design is iterative. Nobody gets it right the first time. The designer makes a version, watches it, adjusts, watches again, adjusts again. Then the client watches and has notes. Then the music changes and everything needs to be retimed. A tool that can’t support fine, local, repeatable adjustments is not a tool for motion design, no matter how beautiful its first output is.

Motion Has a Grammar

My starting belief, which came from years of making animations and games, was that motion, like language, has a grammar.

There are units. A move from one position to another. A hold, where nothing changes. An anticipation, a small movement in the opposite direction before the main action, the way a person crouches before jumping. A follow-through, where parts keep moving after the main action stops, the way hair settles after someone turns their head. A settle, where an object arrives and comes to rest, perhaps with a small bounce.

There are rules for combining these units. Ease in before a big move so it doesn’t start abruptly. Stagger related elements so the eye can follow them one by one instead of being overwhelmed. Never let two things fight for attention at the same moment. Let important elements move last, so they’re the ones the eye ends on. Match the energy of the motion to the energy of the content and the sound.

And there are styles that bend the rules on purpose. A playful brand might use exaggerated bounces that would feel ridiculous in a banking app. A minimal brand might avoid overshoot entirely. A glitch aesthetic deliberately breaks continuity. Good style is knowing the rules well enough to break them meaningfully.

Good motion designers know this grammar deeply, even if they never name it. Animators have written about many of these principles for decades. But in software, most tools treat motion as raw data: keyframes, curves, numbers. The grammar lives only in the designer’s head.

So instead of generating pixels, Deep Motion works with motion itself: keyframes, easing curves, timing, and relationships between elements. The AI’s output is an animation you can open and edit, not a video you can only accept or reject. Every element is still an element. Every timing is still a number you can change. Every curve is still a curve you can grab.

Learning the Grammar

Teaching a model this grammar was the core technical challenge.

Generative models learn from data. For pixel video, there are oceans of data. For structured motion, with real keyframes, curves, and element relationships, there is much less. Most animations in the world exist only as rendered videos; their underlying structure was never saved or shared.

We approached this from several directions. We worked with motion data in structured formats where we could. We built representations that capture motion at the level of grammar rather than raw keyframes, so that “anticipation, then fast move, then overshoot and settle” is something the model can express directly. And we leaned on the fact that much of motion grammar can be described in rules and principles, which gave us a way to guide and evaluate the model even where data was thin.

One lesson that surprised me: relationships matter more than individual motions. A single element moving well is easy. Five elements moving together in a way that feels coordinated and intentional is hard. The model had to learn not just how things move, but how they move in relation to each other: who leads, who follows, what waits for what.

Sketch to Motion

One of the workflows I’m proudest of is going from a rough sketch to animation.

A designer draws a few arrows and scribbles on a frame: this comes in from here, this pops, this follows, this fades as the other one arrives. It takes thirty seconds. Deep Motion turns that into a real timeline with sensible defaults: reasonable durations, appropriate easing, a stagger that makes sense, timing aligned to the music if there is any.

The designer then refines from there. Maybe the pop should be bigger. Maybe the follow should be slower. Maybe the whole thing should be tighter. All of that is direct manipulation of a real timeline, exactly like they’d do in any professional motion tool.

It doesn’t get everything right. It doesn’t need to. It needs to get the structure right so the designer spends their time on feel, not setup. Setup is the tedious part: creating layers, placing keyframes, applying default easing, aligning things to beats. Feel is the creative part: the precise timing that makes something delightful. We want to give designers more time for feel.

Watching designers use this has been one of the most satisfying experiences of building Charsook. There’s a particular expression people have when their scribble turns into motion that already feels close to what they imagined. And then, almost immediately, they start adjusting it. That adjustment is the point.

What Was Hard

Timing is unforgiving. A transition that’s 80 milliseconds too long feels sluggish. One that’s 80 milliseconds too short feels jittery. Human perception of motion is extraordinarily precise, and small errors that would be invisible in a still image are obvious in movement.

A model trained on averages loves the safe, slightly-too-long middle. It produces motion that is never terrible and never great: soft, cautious, a bit sleepy. We had to push it toward decisiveness. Good motion commits. It moves fast when it should move fast and holds when it should hold. Teaching a model to commit, without making it reckless, took many iterations.

Evaluating motion is hard. You can’t judge motion from a single frame. You can’t even judge it reliably from a sequence of frames laid out side by side. You have to watch it, at real speed, and often more than once.

This pushed us to build visual evals that look at motion over time: measuring things like whether elements collide, whether the eye has a clear path, whether energy matches the content, and whether human viewers prefer one version to another in side-by-side comparisons. Later, this work became part of the broader design-judgment layer I’ve written about separately, because the same tools that judge motion can help judge interfaces.

Music changes everything. Much motion design is set to sound. Aligning motion to music is not just matching beats; it’s matching phrasing, energy, and emotion. A crescendo wants motion that builds. A pause wants stillness. Getting the model to listen well was its own long project.

Designers have strong habits. Every motion designer has a set of tools they know deeply, and Deep Motion has to fit into workflows that already exist. We learned to prioritize clean export and compatibility over adding features. A brilliant tool that doesn’t fit into a real pipeline won’t get used.

Pixels and Structure Will Meet

I don’t think the future is either pixel generation or structured motion. I think they’ll meet.

Pixel models are wonderful at texture, atmosphere, lighting, and things that are hard to describe structurally: smoke, water, the way light moves across skin. Structured motion is wonderful at timing, intention, and precise control. Real projects need both.

I imagine tools where a designer defines the structure, which elements, which timing, which relationships, and a pixel model renders the atmosphere around it. Or where a generated video can be decomposed into editable elements after the fact. The research is moving in this direction, and I’m excited to see where it goes.

The tools that win will let designers move fluidly between the two: generating when they want inspiration, structuring when they want control.

Why Control Is a Creative Value

There’s a philosophical point underneath all of this that I care about.

When people talk about generative AI in creative fields, they often frame control as a technical limitation that will eventually be solved, as if the ideal is a model so good that you never need to adjust anything. I disagree with that framing.

Control is not a workaround. Control is where creativity lives. The ability to make small, deliberate adjustments is how a person expresses their specific vision instead of an average one. A tool that removes the need for control also removes the possibility of a personal voice.

This is the same principle behind everything we build at Charsook: humans and AI making work together, with the human’s hands on the work. Deep Motion is that principle applied to time and movement.

Until pixels and structure fully meet, if your work depends on the beat landing exactly where you want it, you need motion you can grab. That’s the bet behind Deep Motion, and so far, every designer who’s tried it has told me the same thing in different words: finally, something that moves the way I think.