6.2 Image-to-Video: The Workshop

The still is validated. Now it moves.

This chapter is the practical layer: which models do what, how they differ in the features that matter for a persona, how to write a motion prompt, and the preparation step that improves consistency more than any model choice.

[AS OF 2026-10] Everything in this chapter is point-in-time. Model versions, prices, and reference limits change within months, and this field has already lost one major platform mid-book. Verify before committing a production budget.

What to look for in a model

Five features matter for this use case, and they are not the features the marketing leads with.

Image-to-video with a strong start-frame adherence. The model must respect your keyframe. This is the single most important property, and it is what separates a usable model from an impressive demo.

Reference conditioning beyond the first frame. Support for one or two additional reference images, or for a short reference video, lets you keep the subject bound as the clip proceeds — the difference between a face that holds and a face that migrates.

Start and end frame control. Specifying both ends turns generation into interpolation between two validated images. Hugely valuable for consistency, because both endpoints are yours.

Maximum duration per generation with shot control. Useful, and to be treated with suspicion. A model that advertises fifteen seconds and up to six shots in one generation is advertising exactly the capability 6.1 says is the field's open problem. Test it on your own persona before planning around it.

Native audio. Convenient, and mostly irrelevant here: chapter 6.5's rule is that dialogue comes from a dedicated voice model, never from the video model.

The current field, by role

Not a ranking. A description of what each option is for, with the caveats that matter.

The self-hosted path deserves a specific note, because it inverts the economics. Wan's open weights mean the marginal cost of a clip is electricity, and the practical cost is setup time and hardware (3.3's 24 GB floor). For an operator generating hundreds of clips a month, the arithmetic can favour a local setup decisively. For everyone else, the setup cost is real and the hosted tiers are cheaper than they look.

Reference limits, concretely. Current frontier models accept roughly one to two reference images, and some accept one to two short reference clips in addition. That is enough to bind a subject and not enough to build a character that walks through a crowd. Plan shots around the limit rather than against it.

Preparing the start frame

This step improves consistency more than any model selection, and it is the one operators skip.

Upscale the keyframe before generating motion. Feed the video model the highest-quality version of the frame you have — the upscaled output from 5.5, not the raw generation. Two reasons: the model propagates the detail it is given, so a sharp start frame produces a sharper clip; and the identity is carried in those pixels, so a mushy start frame gives the model license to invent a mushier face.

Crop to the clip's aspect ratio before generating. If the final video is vertical, the keyframe handed to the video model should already be vertical. Letting the model generate a wide frame and cropping afterwards discards the composition you validated and often the face's position in frame.

Choose the start frame for motion, not for beauty. The best still for a video is one where the pose implies the movement you want, the body is not occluded, and there is room in the frame in the direction of travel. A gorgeous tight portrait with no room to move produces a clip where the subject can only shift in place.

Keep a library of motion-ready keyframes. Same as 5.3's assets, for the same reason: a validated keyframe is a reusable input, and generating one on demand costs more and performs worse.

Writing the motion prompt

A motion prompt is not a scene description. The scene is already in the keyframe. The prompt's job is to specify movement, and to specify what must not move.

Four components:

1. Camera. Static, slow push in, slight handheld drift, orbit, tilt. Static and subtle movements hold identity best. A fast orbit or a whip pan gives the model sixty opportunities to lose the face.

2. Pace. Slow, natural, unhurried. Fast motion is where faces break. Almost every amateur clip is too fast, and the fix is almost always to halve the speed of the action and the camera.

3. Direction and magnitude of subject motion. What changes: head turns slightly to the left, hair moves with a light breeze, she raises the mug. One action, small, completed within the clip. Two simultaneous actions multiply the failure modes.

4. Explicit negatives — what must not change. This is the component people omit and the one that earns its place:

static camera, no camera movement
consistent facial features throughout, no change to the face
no change to clothing, no change to hair length
no additional people
no change to the background

Naming stability explicitly biases the model toward holding the frame it was given. It is the video equivalent of 5.2's retention phrases: it cannot create identity, but it protects the identity you supplied.

A worked motion prompt, minimal by design:

She turns her head slightly toward the window and blinks slowly.
The mug stays at chest height. Hair moves gently.
Slow static camera, eye level, natural pace.
Consistent facial features throughout. No change to clothing or background.
No additional people. No camera movement.

Note what is absent: no scene description, no lighting, no wardrobe. All of that is in the keyframe, and repeating it gives the model permission to re-decide it.

A practical workflow

1. Select the keyframe from the Part V batch (5.6), choosing for motion-readiness.
2. Upscale it (5.5).
3. Crop to the target aspect.
4. Write the motion prompt with all four components.
5. Generate two or three candidates, not one. Variation is real and cheap relative to the review time.
6. Frame-step the candidates (6.3) and reject any with identity drift, even if the movement looks better.
7. Archive the accepted clip with the keyframe, the prompt, the seed if available, and the model version.

Step five is a cost decision worth making explicitly: at $0.10 per second for a fast tier, three four-second candidates cost about $1.20. That is less than the time you would spend trying to fix a single drifted clip in post, and it should be the default.

Failure modes by shot type

A diagnostic table, because the same symptom has different causes depending on what the shot is doing. [UNVERIFIED] Reported by practitioners.

How to use the table. Diagnose the symptom, apply the fix, and if the fix does not work within the three-attempt budget, change the shot. The table's value is that it tells you which failures are worth attempting to fix and which are structural: crowd scenes and morphing hands are in the second category, and the answer there is to stop generating that shot.

Reference clips and motion transfer

Beyond reference images, some models accept a short reference clip showing the movement you want. The model then transfers that motion onto your subject while your keyframe holds the appearance.

When it is the right tool. When motion itself is the content and describing it in words is unreliable — a specific gesture, a dance step, a hand movement that is part of the persona's character.

What it changes about your asset library. A reference clip is an asset, the same way a keyframe is. A small library of motion references — a turn, a walk, a nod, a gesture — is reusable across every future shot, and it is much cheaper than describing the same movement in text each time.

The constraint. Motion transfer inherits the reference clip's framing and body configuration. A reference of someone walking toward the camera produces a walk toward the camera; a keyframe in a seated pose will degrade if the reference clip is standing. Match the reference clip to the keyframe's configuration, the same rule as pose control in 6.3.

Where to get them. Record them yourself with a phone — a person turning, sitting, reaching. The reference clip does not need to be high production quality; it needs to show the motion clearly. And because it is a recording of a real person, chapters 11.4 and 6.5's consent rules apply: get permission for the recording, and check the model's terms on how reference footage is used.

Cost planning per clip

[AS OF 2026-10] [VENDOR CLAIM] Prices are vendor-published and change frequently.

Three notes on reading that table.

Per-second pricing hides the rejection rate. If one clip in three is usable, the effective cost is three times the figure. This is 5.6's survival-rate logic applied to motion, and it is the reason video budgets surprise people.

The tier choice should follow the shot's role. Standard tier for a hero shot that will be seen at full quality; fast or lite for a background clip in a montage where the face is small.

Time is the larger cost for a solo operator. At $0.10/second, a five-second clip is fifty cents. The review, the frame-stepping, the re-generation, and the assembly are minutes of your attention, and attention is the binding constraint.

Where this goes wrong

Feeding the video model a raw, un-upscaled keyframe. You have thrown away the identity detail.

Generating wide and cropping to vertical. Composition and face position both suffer.

A motion prompt that repeats the scene description. The model re-decides the frame it was given.

No stability negatives. The most commonly omitted component.

Fast camera movement. Identity's worst enemy, and rarely necessary.

Two actions in one clip. Multiply the failure modes.

One candidate at a time. Turning a cheap parallel choice into an expensive serial one.

Choosing the model from a leaderboard. Your persona is the benchmark that matters.

The gate

1. Does your chosen model respect the start frame, and have you verified it on your own persona?
2. Is the keyframe upscaled and pre-cropped before generation?
3. Does every motion prompt have camera, pace, subject motion, and stability negatives?
4. Do you generate two or three candidates per shot?
5. Do you know your cost per accepted clip, not per generated one?
6. Have you tested the self-hosted path against your volume?

What you leave with

Five features that actually matter in an image-to-video model, the current field described by role rather than rank, the start-frame preparation step that improves consistency more than any model choice, a four-component motion prompt with a worked example built on stability negatives, a seven-step workflow that generates candidates in parallel, and per-clip cost planning with the rejection rate applied.