916Studio field notes / AI animation
How to Lock Character Consistency in 9:16 AI Stick Figure & 2D Animations
Build a reusable 2D character lock, separate identity from shot instructions, and diagnose face, costume, prop, and line-weight drift in vertical AI video.
A stick figure can become a different character without gaining a single new facial feature. Change its head-to-body ratio, outline width, shoe shape, or the hand holding its mug and the viewer notices the discontinuity. Consistency starts by separating what the character always is from what the character does in one shot. This guide builds that separation into a practical brief for a ten-second, 1080×1920 animation, then shows how to review the generated frames without rewriting the entire idea.
The example uses a round-headed host, a blue hoodie, and a red mug. These are deliberately few enough to audit on a phone screen. You can substitute your own design, but keep the same discipline: describe observable attributes, freeze the essential ones, and vary only the instructions needed for the next beat. A written lock reduces ambiguity; it cannot guarantee that an external model will preserve every detail.
1. Write a compact identity contract
Start with a character identifier such as HOST-01 and six concrete attributes: head shape, facial marks, body proportions, outline treatment, costume, and footwear. A useful instruction is “circular white head, two black dot eyes, no mouth, narrow black limbs, cobalt hoodie, white oval sneakers.” “Friendly, expressive, professional, and memorable” describes an impression rather than a design. Keep emotional direction elsewhere so a surprised pose does not accidentally invite a new face or costume.
Specify line weight relative to the composition when exact pixels are not reliable. A generated image may interpret “4 px outline” as a loose visual cue, particularly after scaling. Keep it in the prompt as a target and use a reference sheet to show the intended appearance. State that the head remains the same size relative to the torso. A close-up should enlarge the whole character consistently rather than redesigning the head to fill the frame.
Scroll the table horizontally to compare all columns →
| Permanent field | Example lock | Keep outside this field |
|---|---|---|
| Face | Two dot eyes; no mouth | Looks left in beat two |
| Costume | Cobalt hoodie; white shoes | Sleeve swings during gesture |
| Prop | Red mug; right hand | Raises mug at 6.2 seconds |
| Outline | Uniform dark stroke | Rim highlight on final shot |
2. Establish a reference sheet before motion
Create a front, three-quarter, and side view before asking for a complex action. If your selected generator supports reference images, use the same approved reference for every related clip and preserve its documented settings. If it does not support references, the written contract becomes even more important. Do not invent a reference-image parameter because another product exposes one; check the actual interface or API you will use. Store the approved images beside the production brief so collaborators can inspect the same source.
Include a neutral full-body pose and a close view of the prop. Test the drawing at the approximate final display size: a 1080-pixel-wide source viewed at 360 CSS pixels is effectively one third as wide. Small costume details can disappear, while an oversized mug can dominate the silhouette. Favor two or three recognition cues that remain visible at that scale. Approve the neutral design before spending generations on camera movement, because movement will otherwise conceal an unstable starting identity.
3. Treat props as continuity objects
A prop needs its own short contract. Name its shape, color, material treatment, size relative to the hand, and ownership. For the example, write “one red cylindrical mug, no logo, right-hand grip, width about half the head diameter.” The relative measure is easier to compare across camera distances than an absolute pixel count. If the mug moves from one hand to another, write the transfer as an explicit action with a visible starting and ending position.
Look for three common failures: the prop duplicates during a gesture, switches hands at a cut, or changes size when the camera moves. Fix the smallest ambiguous instruction first. A sentence asking the host to point while also holding the mug may leave hand assignment unclear. Rewrite it as “left hand points at the diagram; right hand keeps the mug beside the waist.” The solution is specific ownership, not a paragraph of repeated “do not change” instructions.
CHARACTER LOCK: HOST-01
Circular white head; two dot eyes; no mouth.
Slim black limbs; uniform dark outline.
Cobalt-blue hoodie; white oval sneakers.
PROP LOCK: One red cylindrical mug in the right hand.
Left hand remains available for pointing.
Preserve the same proportions in every camera distance.4. Build the shot brief around the locked design
For a ten-second test, reserve 0–3 seconds for the interruption, 3–7 seconds for the demonstration, and 7–10 seconds for the result. In beat one, the host notices a warning light. In beat two, the free left hand redirects a diagram. In beat three, the light turns off and the mug remains unchanged. This gives the model one primary physical action per beat and gives you three clear moments to inspect. Camera and lighting instructions should support those actions rather than compete with them.
Use a close-up only when it helps the viewer understand the face or object. A wide full-body view is better for a pose change involving both feet. Return to a medium framing for the payoff if the mug and hand must remain visible. Avoid asking for a wide view, a face close-up, and a full orbit inside the same two-second interval. When the action requires a genuinely new view, render it as a separate controlled shot and check continuity before editing the cut.
5. Make revisions comparable
Save the complete brief, generator name, model version, reference asset, aspect ratio, duration, and any seed or camera controls actually exposed by that generator. A seed can help reproduce some outputs, but the same seed is not a cross-product identity guarantee. Keep a plain revision note such as “v03: clarified left-hand gesture; all identity fields unchanged.” Without that record, a better result may be impossible to distinguish from a lucky variation or a model update.
Change one category per experiment. First test a static shot with the identity contract. Then add the gesture. Then add a camera change. Finally adjust the lighting. If the costume only drifts after the rim-light instruction is introduced, the revision target is now visible. Batch variation is useful after a reliable baseline exists; before that point it multiplies uncertain variables. Compare the same three timestamps in each render instead of choosing a favorite thumbnail and ignoring the rest of the sequence.
6. Review identity at meaningful frames
Inspect the start, the middle of the largest motion, and the held result. Add a frame immediately before and after any cut. At each location, check head proportions, eye spacing, hoodie color, limb count, shoe treatment, prop count, and prop hand. A simple pass/fail checklist is more useful than a vague quality score. If identity is correct at the start but changes during the motion, shortening or simplifying that motion is a more targeted fix than rewriting the entire character description.
Separate continuity errors from intentional perspective. A hand closer to the camera may appear larger, while an unexpected second mug is never a perspective effect. Compare silhouette and color relationships rather than expecting identical pixel coordinates. Also inspect the safe area: a perfectly consistent host can still be unusable if the right-side interface covers the mug. Review the actual video in the Safe Zone Checker, including its largest gesture, rather than only checking the neutral reference image.
7. Hand off a complete clip specification
A clip specification should survive being copied into a blank production session. Include the character contract, prop contract, format, art direction, beat times, camera choices, lighting, actions, and the exact spoken lines. Avoid phrases such as “same as before” when the destination tool cannot see the previous session. The Prompt Builder exports these fields in readable text or JSON. Choose the readable version for a manual handoff and JSON when your own documented integration consumes named fields.
Rehearse the voiceover before rendering the final clip. At the builder’s planning rate of 2.5 words per second, a three-second beat has space for about seven complete words before pauses are considered. The spoken line should reinforce the visible action, not explain every movement. Keep the finished prompt alongside the approved reference and final media. That package makes a later correction, a new episode, or a different language version much easier to reproduce without slowly redesigning the host.
Frequently asked questions
Can a character lock guarantee identical AI animation frames?
No. It makes identity explicit and reviewable, but output still depends on the model, references, supported controls, and generation settings.
Should I put a seed in the prompt?
Use a seed only if your generator exposes a documented seed control. A number typed into ordinary prompt text is not necessarily interpreted as that setting.
How do I stop a prop changing hands?
Assign the prop to a named hand, assign gestures to the other hand, and describe any intentional transfer as an explicit visible action.
Does 916Studio generate or upload my character?
The Prompt Builder creates a local text or JSON production brief. It does not generate video or submit the prompt to an external generator.