Shooting your own motion reference on a phone: a practical guide for motion transfer

Mon, 26 Oct 2026 23:59:04 +0000

Motion transfer takes the movement from one video and applies it to the character in an image. That means the reference video is not decoration; it is the performance. A shaky, backlit clip of someone half out of frame gives the model very little to follow, and no prompt can fix that afterwards. The good news is that a clean reference needs nothing more than a phone, a wall and ten minutes of preparation.

This guide covers framing, light, background and clothing, how long to record, and a six-point check against the input rules before you upload.

Head to feet with room to spare is the safe default. Diagram made for this post.

What the workflow needs from the clip

The Kling Motion Control page on Kling 4 describes the input simply: one character image and one reference video. The video guides the movement of the subject in your image. The page's advice for both inputs is to keep the subject clearly visible so the model can follow the movement and facial expression, and to use clear lighting and an unobstructed subject.

The motion control page: a character image plus a reference video. Screenshot captured 2026-10-26.

Everything below is a practical expansion of that one sentence: make the subject easy to see.

Framing: the whole performance, with room

Decide the widest move first. If the performance includes a step to the side or arms raised overhead, frame for that moment, not for the resting pose. Then check three things:

Head to feet in frame if the legs move at all. A clip cut at the knees leaves the model guessing what the feet do.
Space around the widest gesture. Hands that leave the frame edge disappear from the reference.
Not too small. A figure that fills a fifth of the frame carries little detail about hands and face.

Match the shape of the frame to the character image where you can. A vertical character portrait pairs naturally with a vertically shot performance.

Light, background and clothes

Recording basics for a clean reference. Diagram made for this post.

Keep the phone still. Put it on a tripod, a shelf or a stack of books. If the camera moves, the model has to separate camera movement from body movement, and that is one more source of error.

Light from the front or side. A window behind the performer turns them into a silhouette and hides the face, which matters if the expression should transfer too.

A plain background with nothing moving. A wall is ideal. People walking past or a television in the background add motion that is not part of the performance.

Clothes that show the limbs. Fitted clothing in a colour that contrasts with the wall makes arms and legs easy to read. A long loose coat hides the knees and elbows the model is trying to follow.

Know the input rules before you record

The page lists concrete requirements, and it is easier to meet them while recording than to fix them afterwards.

Orientation, quality and input requirements as listed on the page. Screenshot captured 2026-10-26.

The character image is one JPEG or PNG up to 10 MB; the reference video is one MP4 or MOV up to 100 MB. Both inputs must be larger than 340 pixels and have an aspect ratio between 2:5 and 5:2. Standard mode gives 720p and professional mode 1080p. You also choose whether the background comes from the image or the video, so decide in advance whether the wall behind your performer should appear at all.

How long to record

The duration you can use depends on the orientation setting. Video orientation accepts reference clips from 3 to 30 seconds; image orientation accepts 3 to 10 seconds.

Reference-clip duration ranges listed on the page. Diagram made for this post.

Record a few seconds of stillness before and after the move, then trim to the part you need. A clean start and end are easier to cut than a performance that begins the instant you press record. Shorter clips also keep the file well under the size limit, and the page notes that the cost quote uses the inspected reference-video duration, so trimming dead time is worth doing anyway.

Phones often record at high bitrates, so a 30-second clip at the phone's highest resolution setting can exceed 100 MB. Recording at 1080p avoids the problem and is plenty for a reference.

Do several takes

Record three or four takes of the same performance and pick the cleanest: no stumble, hands visible throughout, nothing crossing the background. Choosing a reference from several takes takes a minute and saves a generation spent on a flawed one.

The six-point check

A pre-upload checklist built from the page's input requirements. Made for this post.

Before uploading, run through the list: file type and size for both inputs, minimum size and aspect ratio, length against the orientation, a clearly visible subject, and the background source. The page also notes that the workspace checks the uploaded file before quoting, but catching a problem on your phone is faster than discovering it in the upload dialog.

A note on the character image

The same rules apply to the still. The character should be fully visible, facing roughly the same way as the performer at the start of the clip, and not hidden behind objects. If the performer starts facing the camera, a side-on portrait forces the model to rotate the character before it can follow anything.

A good motion reference is mostly about removing distractions. Steady camera, clear light, plain wall, visible limbs. Get those right and the movement in your clip is what the model sees.

Coder Legion - Recent posts tagged motion-control