Recast a person

Replace someone in existing footage with a new actor or character. Ask first: is the original take's performance right, so only the person changes, or does the shot need a new performance (or new words)?

Routes

Route	Pick it when	Cost and time	What it keeps	Load
Swap the still, keep the wardrobe, re-animate	Shot-for-shot remakes; the role's costume stays	Image edits (3 seeds per frame), then a video render per take	Pose, eyeline, framing, light, grain and the original outfit. The performance is new, from the video model	`nano-banana-swap-actor-keep-wardrobe`, then `guide-animate-a-shot`
Swap the still in the new character's own outfit, re-animate	A new character with their own look, any gender or build	Image edits (2 takes per shot), then video	Pose, framing, set, other people; outfit from the references (badge included)	`nano-banana-recast-in-own-outfit`, then `guide-animate-a-shot` or `guide-lipsync-video`
v2v recast	The whole original take is usable and only the face and body change	One H3 Max render per take; source ≥ 5 s (shorter is freeze-padded, then trimmed)	The original performance exactly, and the source's sound	`h3-max-recast`
Recast plus re-voice	The character also gets a new voice or new lines	Voice casting and TTS, then stills, then lip-synced renders	Whatever the still route keeps; new speech	`guide-cast-a-voice`, `guide-voice-a-script`, a still route, `guide-lipsync-video`

H3 Max recast takes 1 to 4 reference photos, one per new person, and replaces the main people left to right by default. Its prompt says who is recast and who stays.

Choosing

Performance right, person wrong: v2v recast. It is the only route that keeps the acting frame for frame.
Costume belongs to the role: keep-wardrobe swap. Costume belongs to the new character (a different gender or build, a uniform with a badge): own-outfit recast.
Image model for the still: nano banana 2.1 edit by default. It is a little faster, more dynamic, and more natural in how it places and forms the new person and composes the scene. Use nano banana Pro edit for a precise detail edit where every other pixel must stay exactly the same; it goes deeper and sharper. The full trade-off is in guide-edit-a-frame-before-animating.
A prop must change in the same pass (the client's product in their hand): the swap on nano banana Pro, with the product reference (guide-keep-small-details).
New words: the audio comes first, then a still route, then guide-lipsync-video.

References

Split the character sheet into single views with python-split-character-sheet (front close-up, profile, full front, full back) and give only the views a shot needs. Single views make better references than the whole sheet.
A shot from behind gets the back view, and the prompt says the shot is from behind.
State every change in the edit (eyes closed, a tray now empty) and what stays; "keep the source's film grain, which is grain, not haze". Where the eyes look and how far the irises turn must be kept explicitly. A frame that needs its own note gets its own call.
To remove an accessory the new character's sheet has (a pin, a badge), edit the picked stills as a patch and edit the reference views too, or the video model puts it back (guide-keep-small-details).

Process

Each gate is a review page (review_open, review_wait): the delivered still or original frame as the reference, the takes as candidates. Never move on unasked.

Map stills to shots by content, never by file name: delivered folders get renumbered and chat uploads lose names. Write the table in AGENTS.md (shot, source still, whether the character appears). For a shot with no still, propose a similar one to reuse. Gate: the person approves the table.
Scrub the footage for the old actor in every shot, including backs of heads, shoulders, reflections and screens. When a still and the footage disagree, recast from a frame of the footage (a delivered still with nobody in it hid the old actor's back of head until after export).
Recast the stills, 2 to 3 takes per shot. Gate: pick per shot; the rest go to rejected/.
Animate the picked stills (guide-animate-a-shot, or guide-lipsync-video for speaking shots), or render the v2v recast. Gate: the takes against the original.
Export each pick with its own audio when the clips were re-voiced.

A woman-to-man recast at a wide conference table worked first time at 2 takes a shot.

Avoid

Attaching every character to every shot: the model puts the wrong people in the room. Cast only who is physically in the frame.
The whole character sheet as one reference image.
Recasting from a delivered still that disagrees with the footage.
The source's last frame as a reference for any edit: its cast leaks back in.

guide-edit-a-frame-before-animating (image models and one-change passes), guide-animate-a-shot, guide-lipsync-video, guide-cast-a-voice, guide-voice-a-script.

Hypernatural