Stage the cast in a set

Give it an empty set plate, a line saying who is where doing what, and single-view references of each person in the shot (plus a picture for any screen); get a shot still with the people in the set at real scale, lit by its light, the set unchanged. Make the plates first with nano-banana-set-from-new-angle. For another still of a framing you've already staged (a new pose, a new expression), edit the picked still with nano-banana-targeted-fix instead of staging again.

Inputs

• The plate at full size (the 2K render in raw/). When the shot is tighter than the plate, crop the plate to the shot's framing and scale it back to 2K first: python-conform-to-1080p with --box and --size 2752 1536. Pro ignores "push in to a medium shot": the presenter stayed small in the wide on 5 takes in a row, and the cropped plate gave a proper medium shot on the first.
• The shot, written down: who stands or sits where, doing what, looking where, with what expression.
• Single views of each person in the shot, cut from their sheets with python-split-character-sheet: usually the front close-up and the full body front; the back view if they face away. Only the people in this shot.
• For a screen in the shot: a frame of the footage that will play on it, and which screen it is.

Run it

Call describe("fal/fal-ai/nano-banana-pro/edit") before the first run in a job. Order the images: plate first, then each person's views grouped by person, screen pictures last. The prompt's numbering follows this order.

run("fal/fal-ai/nano-banana-pro/edit", {
  "prompt": "<the prompt below>",
  "image_urls": ["sets/<set>/angles/raw/host_podium_f.png",
                 "refs/Dana_1.jpg", "refs/Dana_2.jpg",
                 "frames/episode1_call.png"],
  "aspect_ratio": "16:9", "resolution": "2K", "num_images": 1,
  "output_format": "png", "safety_tolerance": "5"
}, into="stills/raw", unit="s00")

Make 2 takes (two calls). Save each prompt beside its take, then conform with python-conform-to-1080p into stills/.

Prompt

The proven template. Fill <the set's light> with this set's own light in words (its key light, any coloured glow or rim light from its screens or practicals, haze), from the set's description and palette:

Image 1 is a photograph of an empty <kind> from the camera position of the shot we need. Stage the shot in it:

<the shot>

Image 2 is Presenter Dana, front close-up
Image 3 is Presenter Dana, full body, front
Image 4 is what the big LED wall shows

The people: take each person's identity from their reference images (face, features, skin tone, hair color, texture and style, build) so they are clearly recognizable, and their outfit exactly as shown (garments, colors, collar). Add nothing to the outfits. They stand, sit and move naturally at real-world scale for where they are in the set, lit by the set's own light (<the set's light>), with real contact shadows and reflections in <the floor>.
Whatever a screen shows comes only from the picture given for it, shown flat on the screen's surface in perspective, inside its border, at the screen's own brightness; it is a screen, not a window.
Keep the set exactly as in Image 1: its design, furniture, screens, lettering, colors, light and grade. Keep the camera position, lens, framing and aspect ratio of Image 1. Photorealistic, like a real frame from a high-end <broadcast shoot of this show>.

Drop the screen sentence when no screen picture is given. To re-angle within the plate (an over-the-shoulder single, a close-up of one person, an insert), replace "Keep the camera position, lens, framing and aspect ratio of Image 1." with the new framing: "Reframe as a chest-up close-up of Lena from the same direction." That worked on about half the takes; for a plain push-in, crop the plate instead.

What you get

A 2K still per call in stills/raw/, conformed to 1920x1080 in stills/, named by shot and letter (s00_a.png).

Tips and traps

• A screen picture from the footage works: a full-screen video-call frame from the Globex training drama, Episode 1, staged on the LED wall came out clean on 2 of 2 takes.
• An empty audience plate fills well with "a mixed studio audience … reacting together" (2 of 2).
• More stills of one framing: edit the picked still rather than stage again. Close-ups staged again from a cropped plate drifted (a generic grey desk, a new background, a different-looking woman, an invented glitch screen); editing the picked close-up with Pro, the character's front view as reference 2 and "change only her pose and expression: …" kept the desk, light and background on 5 of 6 takes.
• A furniture fix on a picked still (a podium too tall) works as a targeted edit with a still that shows the same piece next to the same person as reference 2.
• Drawn characters in a drawn set: replace "Photorealistic" with the look's finish line (guide-design-a-set, Drawn sets) and state every size in every shot ("A about three quarters of B's height"); 8 shots × 2 takes held their sizes. A prop the characters must work around needs its place spelled out ("ALL FOUR cones stay in their row, every one fully visible; they stand on the open lawn BEHIND the whole row"): without it a cone was dropped and the characters stood among the cones.

Check it

Put the takes on a review page beside the plate. Reject a take when the set was redesigned (different wall art, a green screen behind), a person was drawn as someone else (a different hair colour, another face), the show title appears twice, or a screen shows anything but its picture. Check scale and contact shadows where people meet the floor or furniture.