# How do you keep characters and scenes consistent across AI video shots?

Select an image Variant that already has the right character, scene, and composition; use that exact Variant as the video keyframe Reference; then compare completed video Variants and select the main clip. Preserve the image and video Recipes, exact References, and Lineage so the shot remains connected to its source. Reference-conditioned generation reduces identity and scene drift, but it cannot guarantee identical pixels, timing, or motion between shots.

## Why should the image keyframe come first?

A cold video prompt asks the model to invent identity, composition, and motion together. A selected image settles the visible starting point first. The public [Night Market recorded session](/#showcase) uses Mara image Variant `4904331f` as the exact Reference for **Mara — idle loop**.

![Selected Mara image keyframe](/showcase/4904331f-1c34-416f-b7db-7f242412fe17-inspect.webp)

## What does the image-to-video handoff look like?

```sh
makefx video generate --refs asset:mara-the-lantern-keeper~0f4af497@4904331f --name "Mara — idle loop" --type animation "Subtle idle animation: steady rain, warm lantern flicker, gentle breathing and blinking; camera locked, motion returns to the starting pose" --model veo-3.1 --tier fast --duration 8 --resolution 1080p --aspect 16:9
```

```text
✓ asset:mara-idle-loop~7d22189e@3828d140 · 8s video · $0.96
```

![Poster from the derived Mara idle-loop video](/showcase/3828d140-b39e-4177-a825-0f6447d25abb-poster.webp)

The completed video keeps its motion Recipe, exact image Reference, and Lineage back to Mara. The compact workflow is: select image → generate video from its exact Variant reference → compare sibling clips → Activate the strongest completed clip.

## What can still drift between shots?

An image Reference constrains the starting evidence, not every frame. Facial detail, props, motion, camera paths, lighting, timing, and audio can vary. Even “return to the starting pose” does not guarantee a perfect loop. Inspect the full clip, not only its poster. Regenerate deliberate changes as siblings; the current main stays selected while work runs, and failure does not replace it.

Build the keyframe with the [image consistency workflow](/docs/image-playbook), inspect current inputs in [Model & Parameter Selection](/docs/model-and-parameter-selection), and review the record in [Core Concepts](/docs/concepts). [Start the MakeFX quickstart](/docs/quickstart) to run the human selection loop.