Edit a frame before animating
Change a still before a video model sees it: a new actor, one fix, a logo, a removal, a new aspect ratio. The video model copies whatever the frame holds, grain and flaws included, so get the frame right here. Ask first: does the change touch the whole frame (a new person, a restage) or one small region (a logo, a pin, a strand of hair)?
The two image models
Which one depends on the use case. 2.1 is a little faster, more dynamic, and more natural in how it places and forms characters and composes a scene, with more interesting composition. Pro is better for a specific detail edit where every other pixel must stay exactly the same: it goes deeper and sharper.
| Model | What it gives | Use it for | Notes |
|---|---|---|---|
| nano banana 2.1 edit | Natural placement and form of people, livelier and more interesting composition; a little faster | The default for recasts and actor swaps; targeted corrections, removals in a crop, restage variants | At 2K, then conform to 1080p |
| nano banana Pro edit | Deeper, sharper detail; the better pick when everything else must stay exactly as it was | Precise detail edits where every other pixel must stay the same: small print in an enlarged crop, deriving a last frame from a first, one-pass swaps that must also restyle a prop; restaging, and anything 2.1 gets wrong | At 2K. Follows long, multi-part instructions better |
Both are priced per image, far below a video render; describe(model) has the price.
Render at 2K: 16:9 comes back 2752×1536, which python-conform-to-1080p centre-crops to
the aspect and resizes with Lanczos to 1920×1080. Never 4K then downscale.
Routes
| Route | Pick it when | Cost and time | What you get | Load |
|---|---|---|---|---|
| Whole frame, one pass | A new person, or every change at once from the source frame | One edit per seed (about 3 seeds) | The shot rebuilt with the changes; everything else held as well as the model can | nano-banana-swap-actor-keep-wardrobe (and guide-recast-a-person) |
| One targeted fix | A picked frame with one thing wrong (hair, a prop, a pose, eyes) | One edit per fix | The fix; the reference is the original frame, or a crop of it or of the matching frame | nano-banana-targeted-fix |
| Crop and patch | Small print, a logo, a badge or pin to remove | One edit on an enlarged crop | Only the changed pixels go back; the rest of the frame stays bit-identical | nano-banana-rerender-small-print, nano-banana-remove-small-object, python-crop-and-paste-back |
| Restage for another aspect | A 16:9 frame whose 9:16 or 3:4 crop loses a story beat | Two kinds of variant on 2.1 and Pro | A 16:9 edit judged by its centred crop, or a native portrait redraw | nano-banana-restage-for-new-aspect |
Choosing
- One change per pass. Each pass is one instruction set; a second full-frame pass to fix what the first missed smooths the grain and background faces. Fix what's left as a targeted fix or a patch.
- Small region, small edit. A full-frame edit for a small removal comes back 0.5–1% zoomed with every pixel nudged. Crop the region, enlarge it (about 6× for print; 256 px at 2K enlarged 4× for a removal), edit, paste back feathered. A badge removal changed ~9k pixels; the face and grain stayed bit-identical. If a full-frame edit already happened, align it to the original and keep only the box.
- Logos and products that must hold through the shot:
guide-keep-small-details. - Restaging: a 16:9 edit (subjects moved into the middle third) keeps the look of the neighbouring cropped shots and suits close-ups; a native redraw gives more resolution and suits wide shots, but recomposes more freely. Check every person in a redraw: it can duplicate a background person or bunch two on one side. Small held props lose their logo in a redraw; plan a prop fix.
- A still pinned to many renders (a presenter lip-synced in chunks) re-renders all of them when it changes. Settle it before animating, and make late fixes as patches.
Writing the edit
- State every change, and what stays: pose, eyeline, hands, framing, lens, light, the set, background people.
- Keep the grain: "keep the source's film grain, which is grain, not haze". Never ask for sharpening.
- "Do not reframe, zoom, or change the aspect ratio."
- A reference already in the shot's light (a crop of the original frame) beats a
product photo, which pulls in studio light and sharpness. If a product photo is
needed, pre-haze it first (
guide-keep-small-details).
Check it
Each edit goes on a review page (review_open, review_wait): the original frame as
the reference, the candidates, and a flip between them. Judge a logo both at 100% and at
playback size. To check that an edit kept the source's grain, compare
python-measure-detail on the same flat patch or face in the source and the edit; an
edit that smoothed it scores well below the source.
Avoid
- A second full-frame pass to fix a detail.
- 4K then downscale: averages the grain away and redraws background faces.
- A raw product photo as the reference for a prop in a busy shot; hazing it after the edit instead of before (the prop turned grey).
- The source's last frame as a composition reference: its cast and props leak in.
Related
guide-keep-small-details, guide-recast-a-person, guide-animate-a-shot.
