Swap the on-screen actor, keep the format

Take a talking-head template and replace the person on screen with your own character — while the timing, captions, and cut rhythm stay exactly as they are.

Copies the whole guide — prompts, next steps & all — for your AI agent

Watch this chapter

The actor-replacement demo: build a reusable character sprite card, then regenerate one scene with that character as the reference, keeping the template's timing and captions, and approve.

Key takeaways

Actor replacement regenerates a scene conditioned on a reference image — it's a new AI clip, not a face pasted on the old body.
generate_layer with intent: replace_layer copies the slot's timing and full-canvas geometry automatically.
Consistency is mechanical: feed the same character sprite card as the reference on every generation.
On most video providers only the first reference image conditions the clip — pick your single best view and over-describe in text.

The idea: same scene, new person

You keep everything that makes the template work — its pacing, its captions, its energy — and change only who's on screen. The copilot doesn't composite a face onto the original footage; it regenerates that scene as a fresh AI clip, told to reproduce the same action but with your actor. Because it uses generate_layer with intent: replace_layer, the new clip inherits the old slot's timing and full-canvas framing automatically.

The engine behind consistency is a reference image. When you fork and decompose a source video, VidFarm builds a cast array of the recurring people (each with an isolated reference_url); and any character you save lives at /files/characters/<slug>/ as a sprite card. You pass that reference on the generation so the model knows exactly who to draw. Master this once and it powers every generated shot — see Character consistency across shots .

Illustration

character_maya.json Look mid-20s, warm smile, freckles Wardrobe mustard hoodie, gold hoops Voice upbeat, fast, friendly

Maya's sprite card: one reference sheet — front, side, back, face + expressions — the AI conditions every shot on.

Manual walkthroughDo it yourself (the web UI)

Fork a talking-head template

On the Discover feed , open a character-driven / talking-head format and click into the Trackpad Editor . Forking a template that decomposes cleanly gives you a populated cast — the copilot can then resolve "the guy in scene 2" against it.

Open Discover feed
Set up a reusable character

Make the actor you're swapping IN once, so you can reuse them forever:
"Create a recurring character named Zara — a red fox mascot in a teal bomber jacket, cream chest fluff, warm palette. Save her sprite card to My Files."
The copilot generates the image (image gen is cheap) and saves character_sprite_card.png under /files/characters/zara/, annotated so future sessions find it.
Replace the actor in a scene

Now swap her into the shot, keeping the scene's timing:
"Replace the actor in scene 2 with Zara, same pose and energy, 9:16, keeping the scene's timing."
Crucially, copy the path to Zara's character reference card into the chat — e.g. /files/characters/zara/character_sprite_card.png — so the generation locks onto the exact character instead of leaving the copilot to guess which reference to grab. Point it at the precise card and every shot stays on-model.
The copilot fires generate_layer intent: replace_layer on that scene's key, with that sprite-card view_url as input_references. Because AI video is expensive, it asks your permission on cost before it generates.
Re-voice if you need to

If the original narration no longer fits, re-voice it:
"Keep the script but say it in a new voice for Zara."
That runs /audio/regenerate-speech (an approximation, not a clone), mutes the old span, and adds an add_layer kind=audio track.
Re-create music if you need to

If the swap changes the mood, refresh the soundtrack too — either AI-generate a fresh track or drop in one of your own audio files:
"Replace the background music with an upbeat track that matches Zara's energy." generates a new score (ElevenLabs Music). Or "Use badge-anthem.mp3 from My Files as the music" lays down a track you already own — just copy its path from My Files into the chat.
Either way the copilot adds it as its own add_layer kind=audio track so it mixes under the narration.
Review, render, approve

Wait for the pending generation to auto-swap onto the timeline, scrub it, and say "Looks good, render it." The copilot confirms first, renders, and you approve into a shareable post .

The Trackpad Editor showing scene 2 selected, the chat dock asking to replace the actor with Zara, a cost-permission prompt confirming an AI video generation, and a pending-generation placeholder on that timeline slot. — Step 3 — generate_layer replace_layer regenerates scene 2 with Zara as the reference, after confirming the video-gen cost.

🤖 AI Agent · devcli

Replace the actor with your agent

With the vidfarm skill installed and VIDFARM_API_KEY set, hand it the whole swap:

Paste to your AI agent

Using the vidfarm skill, replace an actor in a template:
1. Fork a talking-head template.
2. Create a recurring character "Zara" (red fox mascot, teal bomber jacket, cream chest fluff) and save her sprite card.
3. Replace the actor in scene 2 with Zara — same pose and energy, 9:16 — keeping the scene's timing. Use her sprite card as the reference. Confirm the video-gen cost with me first.
4. Regenerate the VO in a matching voice.
5. Render, approve with a caption, and give me the share link.

Or run the raw devcli steps yourself

Terminal

# generate video for the scene, conditioned on the character reference,
# and drop it into the replace slot (keeps timing + geometry)
vidfarm generate video --prompt "Zara (red fox, teal bomber jacket) delivering the line, same pose" \
  --ref ./characters/zara/character_sprite_card.png --aspect-ratio 9:16 --duration 4 \
  --place ./work --replace scene-2

# render + approve
vidfarm render <forkId> --wait
vidfarm approve --video <renderUrl> --caption "New face, same energy."

Heads up Even though the reference field is an array, on the two default video providers (OpenAI Sora and Gemini Veo) only the first reference image actually conditions the clip — so pick your single strongest sprite-card view and restate the character's look in the text prompt. OpenRouter/Seedance does pass the whole array. Consistency lives or dies on feeding that same card every time.

What it costs This demo crosses into the expensive band: AI video generation runs roughly $1–$10+, so the copilot always confirms budget before generating (image gen, like the sprite card, is cheap and used freely). And you need a BYOK image/video key — Gemini, OpenAI, or OpenRouter — because the free nvidia "lite" key is text-and-vision only and can't generate media at all.

Note You're regenerating the shot, not editing the original body — so lip-sync and motion won't perfectly match the source, and re-voiced audio is a preset approximation, not a clone. The original scene's audio stays unless you explicitly re-voice it.

Where to go next

You've swapped an actor while keeping a template intact. Lock your character down for every future shot, or bring a still image to life:

Recast a template now

Fork a talking-head template, save a character sprite card, and regenerate one scene with your own actor. AI video gen needs a BYOK key and the copilot confirms cost first.

Open the Discover feed

Key takeaways

The idea: same scene, new person

Manual walkthroughDo it yourself (the web UI)

Fork a talking-head template

Set up a reusable character

Replace the actor in a scene

Re-voice if you need to

Re-create music if you need to

Review, render, approve

Where to go next

Recast a template now