Depict real people and places
When a still shows someone or somewhere real, the viewer knows what it should look like
and the model doesn't. Image models don't reliably know a famous face, and asked for "a
ballpark" they draw the most famous one. So the first question is: would a viewer
recognise this person or place? If yes, it comes from photos, and the prompt names it
exactly. Invented places and characters are guide-design-a-set and
guide-recast-a-person.
Routes
| Subject | Route | Pick it when | Cost | What you get | Load |
|---|---|---|---|---|---|
| A real person | A photo of them at the age shown, as a labelled reference in a nano banana 2.1 edit | Always, a celebrity included | One image per take | Their face and hair in the job's style | nano-banana-prompt-format |
| A real place, seen once or from far off | A photo as a labelled reference, the scene drawn new | A facade, a skyline from a known spot, one or two shots | One image per take | The place's look; geometry roughly right | nano-banana-prompt-format |
| A real place whose layout matters | Restyle a photo per camera direction into a plate, then stage people into the plates | Many shots in one place, a distinctive roof or layout, views in and out | One image per plate, then one per scene | The real geometry in every shot, the same from shot to shot | nano-banana-restyle-photo-as-plate, nano-banana-prompt-format |
| A place as it was | Photos from that era, or the plate route on them | The place was rebuilt, renamed or replaced | As above | The right version for each scene | as above |
What each route did on a tribute video (about 30 poster-style stills of a late broadcaster and his team's ballpark):
- The person. Invented from words, he came out as a generic white-haired man who looked like a different public figure, and every later still copied him. From a free-licensed photo taken at the age the story wanted, he was recognisable on the first pass.
- The place from text. "A classic ballpark … a city skyline behind" drew a more famous ballpark and another city's skyline, with light towers the real park doesn't have. 14 of 22 notes in one review were the light towers.
- Photos as references. With the place named and photos given, 13 of 22 stadium scenes still had its unusual roof wrong: the model drew the common kind of roof.
- Plates. Restyled into plates, the photos kept the real geometry: 6 plates for 24 credits, all approved, and 16 scenes staged into them came back with one note.
Process (each gate a review page)
- List the real subjects from the script or transcript: each person and the age
each scene shows; each place and the version each scene shows (before or after a
rebuild, a rename); each event. Write them in
AGENTS.md. - Get the photos. Ask the person first: they often have the best ones (their own
photos from a visit gave the plates the shots used most). Then free-licensed
sources: Wikimedia Commons search and category pages, recording each file's source
and licence in
refs/<subject>/sources.md. For a person, one clear photo at the age shown. For a place, every direction the shots look, plus the outside and any sign. - Write the facts,
refs/<place>/facts.md: its exact name (and former names), what is there, what isn't (no light towers; no skyline from inside), where things stand and which way they face (where the roof opens from; which side of the river the park is on, and where downtown is from it), and the exact text of each sign. Prompts are written from this file, and stills are checked against it. Gate: the photos and the facts on one page; the person corrects them. They know the place: in this job every geography fact came from their notes, three rounds late. - The style frame carries the look, not the subjects. Draw it on a neutral scene, or on a real subject drawn from its photos. Label it "ONLY for the art style" whenever it goes in as an image.
- Plates for each place whose layout matters, one per camera direction
(
nano-banana-restyle-photo-as-plate). Gate: each plate beside its photo. - One or two scenes first, staged into the plates with the person from their photo
(
nano-banana-prompt-format: camera, facing, every word spelled). Gate. Then the batch, and every still on one page before anything is animated. - Check every still yourself first, against the facts and the photos, one fact at a time. A set can look consistent and still be the wrong place: this job's first self-check called 30 stills "very consistent" when every stadium in them was wrong.
Choosing
- Name and photo, never one alone. A name without a photo gets the model's idea of it; a photo without a name or facts loses what the photo doesn't show.
- The person's model: nano banana 2.1 edit. GPT Image 2.5 edit gave an unnatural head
and hair on both likeness tests; 2.1 got both right. Pro for precise detail edits
(
guide-edit-a-frame-before-animating). - Older photo for "in his prime": the photo sets the age better than words do.
- Plates when the place is seen from three or more directions or has a structure the model gets wrong; photo references when it is seen once.
- Nano banana 2.1 has an
enable_web_searchoption; it is untested here and never a substitute for photos. - Texture: 2.1 comes out smoother than a gritty style frame. A grit pass (a redraw against the style frame) and fine grain in post matched them.
Avoid
- A generic description of a real place or person ("a classic ballpark", "an elderly broadcaster with white hair").
- Inventing a real subject in the style frame: it becomes the reference every still copies.
- Banning all text: spell the real signs (
nano-banana-prompt-format). - Generating a place again for each shot when plates would hold its geometry.
- Mixing eras by accident: decide for each scene which version of the place it shows.
Related
nano-banana-prompt-format: how to write each prompt.nano-banana-restyle-photo-as-plate: plates from photos.guide-design-a-set: places that don't exist.guide-animate-a-shot: the stills once picked; derive a clip's end frame from its start (nano-banana-derive-last-frame) so the place matches.
