# ANOMALY Knowledge · Nano Banana Pro reference

**The workhorse.** Every frame after the master character comes from here: hook frames, build shots, end-result shots, and recreations of a competitor's frame. Its real strength is *editing an existing frame* while holding a reference character across many shots.

**Model id on Higgsfield: `nano_banana_2`.** Do not use `nano_banana_flash` — that is the weaker model and it will quietly cost you quality on every shot.

---

## Settings that decide whether it works

| Setting | Value | Why |
|---|---|---|
| Model | `nano_banana_2` | `nano_banana_flash` is worse |
| Aspect ratio | `9:16` | State it in the parameter **and** in the prompt text |
| Resolution | `2k` for anything you post, `1k` for drafts | |
| Prompt format | **Plain text only** | JSON-structured prompts fail on this endpoint |
| Prompt length | **Under ~1100 characters** | Longer prompts get rejected outright |
| References | Upload once, reuse the returned id | Re-uploading the same character every time wastes time and ids expire |
| Runs per prompt | **2** | Outputs vary; best-of-two is free quality |

---

## KEEP / SWAP — the framework that makes recreations work

When rebuilding a competitor's frame, do **not** describe the scene from scratch. Tell the model precisely what to keep and what to change. Being surgical is the whole trick: the model only has to solve the small thing you asked for. Ask it to redo everything and it drifts.

```
Edit reference image 1 ([one-line scene description]).

KEEP EXACTLY: [room, architecture, floor, furniture, background people and decor,
composition, camera angle, and the lighting stated precisely — e.g. "cold fluorescent
overhead, roughly 5500K"]

SWAP ONLY THESE TWO THINGS:
(1) Person, replace with the character from reference image 2.
[Character DNA, verbatim]. [Outfit appropriate to the scene].
Same position, same pose, same body scale.

(2) [Original items], replace with [your product + short description,
referencing your product image].

LIGHTING: preserve the exact lighting from reference 1.
CAMERA: 9:16, same framing as reference 1.
AVOID: wrong face, original items still visible, background changed,
warped hands, watermark, on-screen text.
```

---

## The fresh-scene prompt

When there is no frame to recreate:

```
[Tight macro of the action / the product being worked on / the finished product].
[Character DNA, verbatim], in [setting with specific real details].
[Exact action being performed].
Product: [Product DNA, verbatim].
[Outfit]. Bright, saturated, high contrast.
Leave clear negative space upper-centre for a caption.
Shot on iPhone, handheld, natural light, organic content, real person. Natural [light source] light.
9:16 vertical, 2K.
No studio lighting, no perfect skin, no CGI, no ring light, no posed model,
no watermark, no on-screen text.
```

---

## Three rules from the winner corpus that change what you prompt

**1. Never ask the model to render text.** Garbled text is the single most-cited AI giveaway in the corpus — text appears in the AI-evidence notes for 42 of 77 AI videos, ahead of morphing, physics, hands and faces. Videos with named garbled text sit at roughly half the plays of those without. So: no box art, no signage, no posters, no price labels, no shirt graphics, no embroidered logos. **Every word gets burned in at edit time.**

**2. In AI shots, make a live animal the hero where the product allows it.** Inside AI-generated videos, those featuring a real cat, dog, puppy or kitten ran ~1.4M median plays against ~456k without — a large gap that holds at account level. Note this is AI-specific: on real filmed footage animal videos are slightly *worse*. The reading is not "animals win", it is "animals are what the generator fakes convincingly". A costume or an object alone does not substitute.

**3. Do not burn rounds chasing a perfect frame.** 77 of 160 winners are AI-generated or partly AI, and within an account AI and real-filmed perform the same. The top AI video in the corpus did 5.4M plays with four visible artifacts including a hand merging into a coffee cup. Fix the concept first, then the artifact. Only rendered text, hands and morphing carry a measured cost — fix those three, then post.

---

## Composition that actually drives views

1. **Leave negative space for the caption.** Decide where the text sits (usually upper-centre) and compose the character and product *outside* that zone. Compose around the caption, not the other way around.
2. **The first frame is a tight macro of the action**, not a wide establishing shot. Close on hands, the tool, the material.
3. **Something is always happening.** Cutting, pouring, rolling, stitching. A dead static frame is the tell.
4. **Bright, saturated, high contrast** beats tasteful and muted on a feed.
5. **Show volume in the background** — finished pieces stacked behind the character read as a real operation.

---

## Hard-learned failure modes

| Symptom | Fix |
|---|---|
| Dark scene comes back brightened | The model pulls dim scenes toward the well-lit character reference. Write "do NOT brighten this scene, preserve the dim moody lighting exactly" |
| Face drifted between clips | The DNA block was reworded. Paste it byte-identical every time |
| Scene fell apart on a hard shot | Too many changes at once. Do the person swap first, then the product swap as a second pass on that output |
| Wrong scene lock | Too many references. Do not pass product references into a scene where the product does not appear |
| Warped hands | Reframe so hands are partially out of frame, or simplify the action |
| Prompt rejected | Over ~1100 characters, or you sent JSON. Shorten and use plain text |

**When it keeps getting it wrong:** re-read your own original prompt and identify the specific wording pulling the bad output. The fault is almost always ambiguous language in the prompt, not bad luck in the generation.