6.3 Identity in Motion: Advanced Techniques

Chapter 6.1 gave the rule: five seconds, one shot, no rapid change. This chapter is what to do when the shot you need does not fit the rule.

Four techniques, in ascending order of effort: masking, pose control, subject binding, and giving up gracefully in favour of stills.

Masking: constrain what the model may change

The core insight of masked image-to-video is that most identity drift comes from the model being allowed to re-decide things it does not need to decide. If you mask the regions where change is acceptable and freeze the rest, drift is limited to the masked area.

How it works. You supply the keyframe plus a mask. Outside the mask, the model receives the original pixels as a constraint; inside the mask, it generates motion freely. On a Wan-family image-to-video workflow there is a dedicated control node for exactly this, and the concept transfers to any model that accepts a conditioning mask.

A practical configuration for a persona:

The general shape: freeze the identity-critical regions absolutely, and allow motion in the peripheral ones. A clip where only the hair and the background move is still a clip, and it retains the face perfectly. This is a legitimate aesthetic, not a compromise — much real portrait video is exactly this.

The mask is drawn on the start frame and applies across the clip. It does not track motion, which is the technique's primary limitation: if the subject moves substantially, the mask no longer corresponds to the anatomy, and the model produces a seam or a warped limb.

Honest assessment. [UNTESTED] The published accounts of this technique in operator communities describe results as better than unmasked and not perfect. Failure modes reported: visible transition at mask boundaries when the mask edge crosses texture, a frozen region that looks frozen (a face that holds expressionlessly while the hair moves reads as uncanny rather than stable), and difficulty with any pose change. Treat masking as a technique that raises the floor and does not remove the need for QA.

The conditioning chain

Masking is one input into a larger conditioning setup. In a node-based workflow the chain looks roughly like this:

CLIP Vision Loader ─→ CLIP Vision Encode ─┐
[upscaled keyframe] ───────────────────────┼─→ Wan Image-to-Video ─→ frames
[mask] ────────────────────────────────────┤
[motion prompt] ───────────────────────────┘

Three configuration notes that materially affect output.

Higher start resolution improves identity retention. This is the 6.2 preparation step in architectural terms: more pixels of the reference mean more identity for the model to condition on. If your workflow has a resolution setting for the conditioning input, raising it is usually the single biggest quality lever available.

The vision encoder and the sampling path should agree on the reference. Feeding one version of the keyframe to the vision encoder and a differently processed version to the video path produces inconsistent conditioning, which shows up as a face that fights itself.

Keep the motion prompt short when conditioning is heavy. With mask plus reference plus vision encoding, an elaborate prompt adds conflicting instructions. State the movement and the negatives; nothing else.

Pose control

When a specific movement is required — a turn, a gesture, a walk — specifying it in text is unreliable. Pose control specifies it as data.

How it works. An openpose-style skeleton extracted from a reference video or drawn manually is fed as an additional condition. The model generates motion matching the skeleton while the image conditioning holds identity.

When it earns its cost. Demonstrations, gestures the persona repeats as part of their character, and any movement where the precise motion is the content rather than a decoration.

The failure mode. Pose and identity are separate conditions, and the model must satisfy both. In practice it satisfies the pose more reliably than the face, because the pose condition is spatially explicit. Poses far from the keyframe's pose — a standing keyframe driven toward a seated skeleton — degrade identity badly, and the degradation is not obvious until you watch the clip.

The practical rule: keep the driving pose within roughly the same body configuration as the keyframe. If you need a drastically different pose, generate a keyframe in that pose first (Part V) and drive the motion from there.

Subject binding, and models that offer it

Some video models offer an explicit subject-reference feature — one or two reference images of the subject supplied alongside the keyframe, with the model instructed to keep them coherent through the clip.

When it helps. Multi-shot clips and any sequence where the subject must survive a camera change. It is the closest thing available to the capability 6.1 identifies as unsolved.

What it does not solve. It binds appearance, not identity in the deep sense. A model can keep the face consistent in a shot while changing the person's apparent age, expression range, or body proportions.

How to evaluate it. Not from demos. Take one of your own validated keyframes plus your reference sheet, generate a shot with a camera move, and frame-step it. If the face holds through the move, the feature works for your persona. Two hours of testing tells you more than any amount of model comparison content.

QA for motion: frame-stepping

Video QA is a different procedure from still QA, and watching is not it.

Step through the clip a frame at a time, with attention on three moments: the first frame (should match the keyframe exactly), the midpoint (where drift usually begins), and the last frame (where it is most visible).

Watch the transition, not the endpoint. Drift is a process. The tell is a face that is subtly different at second two and clearly different at second four, which reads as a snap at speed.

Check hands through the whole clip, not only the first frame. Hand errors in motion are worse than in stills, because they morph.

Check the boundaries. If the clip will be joined to another, compare its last frame with the next clip's first frame. Joins are where continuity breaks, and 6.4 is about managing them.

Score identity on the worst frame, not the mean. Chapter 4.6's rule, applied to motion.

Giving up gracefully

The most valuable technique in this chapter, and the one that saves the most money.

If a movement cannot be generated consistently after three attempts, change the shot rather than the technique. Options, in order of cost:

1. Reduce the movement. A smaller action holds better. A slight turn instead of a full turn; a head lift instead of standing.

2. Cut around the problem. Two short clips with a cut between them, instead of one continuous movement. A cut hides drift, because the audience reads it as a new shot rather than a continuity error.

3. Replace motion with stills. A sequence of three validated stills with a push-in and a sound design is a legitimate video, and it costs a fraction of a generated clip. Chapter 6.1 lists this as a substitute; it is also the correct answer when a movement resists every technique above.

4. Change the content plan. Some shots are not worth generating with a synthetic persona at this stage of the field's capability. A walk through a crowded street with a consistent face is one of them. Design the account around what is currently possible rather than around what would be ideal, which is the same discipline as 2.1's feasibility axis applied to motion.