Hypernatural

TechniqueBy HypernaturalVersion 1, Oct 21, 2026

nano-banana-swap-actor-keep-wardrobe

Swap a new actor into an existing frame from a character sheet with nano banana edit, keeping the original's wardrobe, pose, eyeline, framing, light and grain: three seeds per frame at 2K, then Lanczos to 1080p. Use to recast the stills of a spot shot for shot; on Pro it can restyle the held product in the same pass.

SKILL.md

Swap the actor, keep the wardrobe

You give it a frame from the source and a character sheet of the new actor; you get the same frame with the new face, hair and build, and everything else (wardrobe, pose, eyeline, hands, set, light, grain) kept. Use it for the stills of a shot-for-shot remake. When the new character should wear their own outfit, or the swap changes gender or build, use nano-banana-recast-in-own-outfit instead.

Inputs

  • The source frame (the shot's first, or last, still).
  • A character sheet of the new actor (front, back and close-up views).
  • Optional, Pro only: a pre-hazed product photo when the held product must change too (nano-banana-swap-held-product says how to prepare it).
  • Who is replaced, as seen ("the woman in the green cardigan"), and what to call the new character ("the woman", she / he / they).

Run it

Call describe on the model before its first run in a job. One call per take, three takes per frame, each with its own seed (separate seeds spread more than num_images: 3):

run(model="fal/google/nano-banana-2.1/edit",
    input={"prompt": <the swap prompt>,
           "image_urls": ["<job>/shots/shot01_first.jpg", "<job>/references/sheet_b.jpg"],
           "aspect_ratio": "16:9", "resolution": "2K", "num_images": 1, "seed": 1000,
           "output_format": "png", "safety_tolerance": "5", "thinking_level": "high"},
    into="<job>/swap_b", unit="shot01_first")

Seeds 1000, 1001, 1002 for the three takes. Add the product photo as the third image for a one-pass swap, on fal/fal-ai/nano-banana-pro/edit (no thinking_level there).

  • 2.1 edit for a plain swap. Pro edit when a prop must change in the same pass, and for a one-pass swap of a shot that will get a pinned last frame.
  • 2K, never 4K. 2K 16:9 comes back 2752×1536; conform it to 1920×1080 with python-conform-to-1080p (centre-crop, Lanczos). 4K sharpens lettering a little, but the downscale averages the film grain away and redraws background faces.

Prompt

The full template and a filled example are in references/prompt.md. It names image 1 as the shot to reproduce and image 2 as the sheet, takes only identity from the sheet (face, skin, eyes, hair colour, texture and style, build), drops any hair clip from the original actor, and lists everything kept. Two things in it matter most:

  • The eyeline is spelled out: if they glance sideways in the source, the irises sit just as far into the corners of the eyes. Without it, side-eyes drift to camera.
  • Every change is stated. Anything the shot needs different (eyes closed, cups swapped, a tray empty, keep the source's grain because it is grain and not haze) goes in as its own sentence at the end. Frames that need their own note (a smirk, a side-eye) go in their own call.

When the new character has hair accessories, say where they sit and that they come from the sheet.

Tips and traps

  • A prop changed after the swap from a raw product photo pastes in a studio-lit, too-sharp object. Put the pre-hazed photo into the swap itself (Pro), so the prop is rendered in the scene's light. Hazing the result afterwards turned the product grey.
  • Small print comes out garbled at frame size. Don't fix it with a second whole-frame pass: use nano-banana-rerender-small-print on the picked still.
  • First and last stills swapped independently don't match (hair, cuffs, cup, eyes), and a video pinned to both morphs between them mid-shot. Swap the first frame only and make the last with nano-banana-derive-last-frame.
  • One correction left after picking (a hair strand, a letter): nano-banana-targeted-fix.

Check it

Put the three takes per frame on a review page with the source frame as the reference. Look at identity against the sheet, the eyeline, the hands and grip, and the grain at 100% against the source. Logos and print at playback size. Record the seed of the pick.

Keeping small products and print exact through the whole shot: guide-keep-small-details.

Go make something cool.

Install the app, open your agent, describe the video. That's the whole onboarding.

Sticker HypernaturalSticker BannerSticker TVBoy