Kikomono field notes / AI video workflows
Kling 3.0 Motion Control Guide
Start with an action video you have permission to use and a character image with compatible framing. Match the orientation, use facial-element binding when available, and inspect a simple result before attempting a complex scene.
Prepare the two references
The image establishes your intended character; the action video supplies movement. Pick an easy first task: a slow turn, a small hand gesture, or a short step within the frame. Keep a copy of the unedited references so you can tell which change helped.
Choose the image for the movement
A close face portrait cannot show the body arrangement needed for a full-body action. Match the visible body area between the image and video, leave room around the subject, and check whether hands or feet are already cut off. For a beginner's diagnostic run, a single clearly visible subject reduces ambiguity.
Trim the video without changing the action
- Open your own video in an editor. Choose the continuous section that contains the complete movement you want.
- Set the start before the action begins and the end after it finishes. Remove introductions, cuts, and unrelated actions.
- Preview the trimmed clip at normal speed. Make sure the subject remains visible throughout, especially the head and hands at the extremes of motion.
- If you crop, check every part of the movement rather than just the first frame. A tighter first frame may remove a hand later.
- Export a copy and check the upload panel's current format, size, and duration requirements. Do not silently reuse a limit from a different model or a third-party service.
The official combined guide lists a 3–30 second upload duration and image-edge restrictions in its 2.6 section. Those are not independently verified here as 3.0 upload limits. The selected 3.0 interface is the authority for whether your file is accepted.
From your first idea to a reviewable result
- Open the official service. Use kling.ai. Sign in yourself and confirm the model shown in the Motion Control panel. A guide title cannot guarantee that your account exposes a particular model.
- Upload your action reference and character image. Check each preview after upload. Confirm that you have selected the intended files, rather than relying on similar filenames.
- Choose orientation deliberately. If you want the character's orientation to follow the action video, use the corresponding video-matching setting. Do not switch to image-matching simply because the wording sounds more consistent.
- Check facial binding. The official 3.0 guide describes binding a facial element beneath the character image; it says this is supported when orientation matches the video. If available, use references of the intended face, including relevant views or expressions. Facial binding does not supply clothing, hair, makeup, or props.
- Keep the first request narrow. Describe only scene details you need to specify. A long prompt that requests several new actions makes it difficult to distinguish transferred motion from additional instructions.
- Read the cost before generating. Check the actual panel's credit charge and your available balance. We do not publish a fixed credit cost because mode, provider, and model versions differ.
- Submit once and retain the job information. If the job is queued, do not repeatedly submit the same request. Record its visible status and any error text.
- Review the whole output. Watch the beginning, the largest movement, the turn away from camera, and the final pose. Compare motion, identity, framing, and background separately.
- Change one input for the next diagnostic run. Keep the model and other settings unchanged. This is an editorial testing method, not a promise that regeneration will repair the result.
How should you set motion_score?
We could not verify motion_score as a user-adjustable setting in the official Motion Control guide inspected for this article. There is no defensible universal value or slider range to recommend. Do not paste an invented parameter into an API request.
If a provider exposes that name, first read the documentation for that exact endpoint and model version. It could be a provider-specific control or a returned metric. Record whether the field is an input or output, its allowed values, and its default. Only then compare the documented default against a permitted alternative while keeping your references unchanged.
If your interface has no such field, work with the real controls it offers: reference selection, character orientation, facial binding, and scene description. Adding the text “motion_score” to a prompt is not evidence that a supported parameter changed.
Diagnose the result before spending more credits
| Observed problem | First check | Next diagnostic change |
|---|---|---|
| Face changes during a turn | Compare the character preview and bound facial element. Check whether the needed face angle is represented. | Try suitable facial references or a simpler turn. Keep other inputs fixed. |
| Head or limbs leave the frame | Watch both references through the full movement; compare body scale and available space. | Reframe the source without cutting off the action, or select a movement with less displacement. |
| Wrong person drives the result | Check whether another person dominates the source frame. | Use an unambiguous single-subject reference rather than trying to correct selection with more adjectives. |
| Action ends early | Look for cuts, disappearance, blur, or a rapid change of direction in the reference. | Try a cleaner continuous segment. Do not assume all discarded duration will be refunded. |
| Background changes unexpectedly | Separate motion transfer from full-scene video editing. Check your selected orientation and scene instructions. | Reduce conflicting scene requests and compare a simple fixed-framing example. |
| Upload or generation fails | Save the exact error, file properties, selected model, and job status. | Correct a documented input violation. For an accepted job that fails, use official support with its job identifier rather than repeatedly resubmitting. |
Reddit includes both positive demonstrations and reports of unwanted framing or identity changes. They are self-reports, not controlled comparisons. Our practical response is to isolate the failure category and use a smaller diagnostic action, rather than claim that a single prompt fixes every video.
A practical review sheet
Keep a local note with these fields: provider; selected model; date; orientation; whether facial binding was used; reference filenames; cost displayed; job status; and the exact point where the output departs from the reference. Write “not shown” for a missing setting. This gives support a reproducible problem without inventing internal parameters.
Example description, not a tested recipe
“A single character in a plain room, steady lighting, with the appearance shown in the character image.” Treat this as a simple scene-description example. The action comes from the reference, and the output still needs inspection.
What remains to be supplied
Your own licensed reference image and action video are required to run a generation in the official service. We have not generated or published a before-and-after demonstration. We will only add one when its source rights, settings, and actual result can be documented.
Sources and verification scope
Primary reference: Kling Motion Control user guide, inspected in Chrome on October 3, 2026. The first-run steps and facial-binding statements follow the official guide; our trimming and diagnostic workflow is editorial advice.
- Motion transfer discussion: distinguishes character motion transfer from scene modification; linked as a community question, not a source for prices or performance.
- Creator demonstration discussion: a favorable self-report, not independent validation.
- Community problem discussion: illustrates why framing should be inspected rather than assumed.
These are three inspected discussions, not a verified list of the latest three posts. Automated bot replies were not used as technical authority. No search-volume, difficulty, output-success, or speed estimate is made.
About this guide
OuKiki publishes independent, source-linked workflows for creators. Kling and its product names belong to their respective owners; this guide is not affiliated with or endorsed by Kling. Corrections are welcome through the contact link.