On Lost Garden, my solo AI anime project, I needed one shot: a knight turning toward a doorway while the camera arced slowly around him. I wrote “camera slowly orbits the character, cinematic, dynamic movement” and generated. Then I generated again. By the twentieth version the knight still wasn’t turning, and the camera had tried everything except an orbit: a push, a whip pan, once a move that circled the wrong way entirely.
The fastest fix for inconsistent camera movement in AI video isn’t a better prompt. It’s a short reference clip you film yourself and feed to the model as the motionaround. By mid-2026 it’s a built-in feature in several tools filmmakers already use, and it solves a problem text prompts were never going to solve on their own
Why does camera movement in AI video look wrong even with a detailed prompt?
Because the model isn’t simulating a camera. It’s predicting pixels that resemble footage people have labeled with words like “dolly” or “orbit,” and that association gets less reliable the more specific or compound the move gets. Runway’s own guide to camera prompts recommends naming the exact cinematography term, a “slow dolly push-in” instead of “cinematic,” and attaching a speed to every movement word. That advice is correct, and it still isn’t enough once a shot needs two things to happen at once: the camera arcing while the subject turns, a push that also tilts, a handheld wobble with a specific rhythm. A sentence can describe that. It can’t direct it.
A locked-off shot or a simple pan still prompts fine most of the time. The model has plenty of training examples that look exactly like what you typed. It’s the shots with real blocking, the kind a cinematographer would sketch on a diagram before rolling, that start drifting.
What’s the real difference between describing a camera move and directing one?
A prompt describes an outcome and hopes the model fills in the physics. A reference video is the physics: the exact path, speed, and timing of the camera, stripped of everything else and handed to the model as a signal to copy.
That distinction is a named feature now, not a workaround, and different tools are solving it in different ways.
Kling’s Motion Control, released on Kling Video 2.6 in mid-2026, takes a motion reference video as one of its three required inputs, alongside a character image and a short text prompt for context, and transfers the camera and body motion from that video onto the generated shot, confirmed
Kling AI’s Motion Control feature page for Kling Video 2.6.
Higgsfield takes a different route with Cinema Studio: instead of a reference clip, it locks camera behavior, lighting, and lens into explicit settings before generation, so the same movement wording produces the same physical result on every run, a fix Higgsfield itself documents for prompts that leave speed, path, and end position unstated.
Higgsfield’s Cinema Studio documentation on controlling camera movement, angle, and lens.
Runway, by contrast, still routes camera direction mostly through text and dedicated UI sliders rather than a reference clip, which is consistent with how its own camera-prompting guide teaches cinematography vocabulary instead of a filming step.
Runway’s AI camera prompts guide, dated August 2026.
Not every shot needs this. Save it for the moves that are genuinely hard to put into words:
- Write the prompt for a locked-off shot, a simple push-in, or a pan with rough timing. It’s cheaper, and it usually works on the first try.
- Film a reference for anything handheld, arced, or compound, for a move that has to land on a specific beat, or for a camera move you need to repeat identically across several shots in the same scene.
- Film a reference the moment a prompt-only shot fails twice in a row for the same reason. Two failed rerolls on the same movement is the signal to stop rerolling and start filming.
The twentieth generation is not a better dice roll than the first one. If the same part of the prompt keeps failing, the prompt isn’t the problem.
Prompt versus reference: which camera moves need which approach.
How do you actually film a usable motion reference?
Three things matter, and they come straight from how the tools require the input to work.
- Keep it short. Kling’s own spec caps a usable motion reference between 3 and 30 seconds, and a clip close to your target shot length transfers cleanest.
- Match the opening frame. Line up your phone’s starting angle with the first frame of the shot you’re generating. That alignment, more than anything else about the footage, decides how well the motion transfers.
- Ignore what’s actually in frame. The content of your reference video gets discarded after the motion is extracted; only the path and timing of the camera carry over. You can film the move in your kitchen and apply it to a throne room.
The four-step method for getting an AI video camera move right.
One production outside my own put a number on what this saves: a creator who had burned 50-plus generations, hundreds of credits, and over an hour chasing a single camera move got it in one take once they filmed the motion instead of typing it. On that orbit shot for Lost Garden, re-filming the move on my own phone and feeding it in as the reference got me a usable take on the next generation, not the twenty-first.
Does this replace planning the shot before you generate?
No, and this is where it connects to the step before generation, not after it. A motion reference fixes a camera move you’ve already decided you need. It doesn’t tell you which shots in a scene are worth that move versus a plain static frame, and guessing wrong burns the same hour you were trying to save. That decision belongs in the storyboard, before a single generation runs. It’s part of why ScreenWeaver has you lock camera language into the shot list during pre-production: not to replace filming a reference, but so you already know, before you generate anything, which shots are simple enough to prompt and which ones are worth the extra two minutes with your phone.
I write and teach around this at frankhoubre.com, mostly because I keep running into the same prompt-versus-camera walls my students do, on their projects as much as on mine.
Mini-FAQ
Does the reference footage show up in the final AI video?
No. The motion gets extracted and applied to your generated scene; the original footage is discarded.
How long should a motion reference clip be?
Between 3 and 30 seconds, matched as closely as possible to the length of the shot you’re generating.
Does motion transfer work with more than one character in frame?
Not reliably yet. Kling’s current documentation specifies a single character in the motion reference, so multi-character blocking still leans on prompting and manual correction.
Do I need special gear to film the reference?
No. A phone held in your hand is enough. What the camera is pointed at while you film doesn’t matter, only how it moves.
Keeping camera direction out of the prompt and in your hands instead feels like a small shift. On a shot list with a dozen moving shots, it’s the difference between losing an afternoon and losing a week. If you’re building that shot list before you generate anything, ScreenWeaver’s storyboard stage is free to try against your own script, and Lost Garden is where I keep testing this stuff before I write about it.
