# Reference Boards & Consistency Grids — Which Model for What

Nodaro's **Generate Image** node ships a family of factory presets for *identity consistency* —
keeping the same person, pet, product, or place recognizable across every generation in a project.
This guide explains the two families, which model each one uses, and the head-to-head experiment
behind those defaults so you can make your own trade-offs.

## The two families

| | **Reference Sheet boards** | **Cast & Consistency grids** |
|---|---|---|
| Presets | Character, Pose, Location, Product, Outfit, Scene, Creature, Vehicle, Food, Mascot, Pet Board | Character Reference Grid, Cast Mega Grid, Cast Scene |
| What they make | A dense, editorial production sheet: hero shot, metadata block, labeled panels (views, expressions, details, lighting), HEX color palette | A sterile, neutral-background grid of canonical angles — no text, no decoration |
| Made for | **Humans** — art direction, briefing, a project's visual bible | **Models** — feeding back into later generations as an identity anchor |
| Default model | `nano-banana-pro` @ 2K, 16:9 | `nano-banana-2` @ 4K, 3:4 |

Both follow the same workflow: **connect one sharp, well-lit, front-facing photo** → generate →
**reuse the result as a reference image** in every later generation (image or video) featuring that
subject. For multi-character work, build one **Cast Mega Grid** (2–4 characters as labeled strips in
a single image), then stage scenes that reference cast members **by the names on the grid** without
re-describing them — less re-description means less identity drift.

## Which model should you use?

| Job | Use | Why |
|---|---|---|
| Any Reference Sheet board | `nano-banana-pro` (default) | Best identity fidelity across panels and best text rendering for the metadata block, panel labels, and HEX swatches |
| Identity grids to feed back as references | `nano-banana-2` (default) | Nearly Pro-level identity at lower cost and higher speed — consistency work is iteration-heavy, so cost-per-attempt matters; 4K keeps panel faces sharp when reused |
| Layout-critical sheets where likeness is secondary | `gpt-image-2` | In our tests it followed multi-panel layout instructions the most completely and produced very uniform panel sizing — but the face drifts (see below) |
| Label/edit workflows (Edit by Name, annotations) | `gpt-image-2` | Strong instruction-following for overlay/labeling tasks |

## The experiment behind the defaults

We ran the same prompts with the same source photos head-to-head (June 2026, one generation per
cell — treat as directional, not statistical):

- **Test 1 — Character Board:** `nano-banana-pro` vs `gpt-image-2`, both 2K / 16:9
- **Test 2 — Clean Reference Grid:** `nano-banana-2` vs `gpt-image-2`, both 4K / 3:4
- **Test 3 — Cast Mega Grid (2 people):** `nano-banana-2` vs `gpt-image-2`, both 4K / 3:4

### What we found

1. **Identity fidelity is the differentiator — the Nano Banana family wins it.** Across all three
   tests, `nano-banana-pro` / `nano-banana-2` reproduced the *same person* from the source photo
   (face shape, stubble pattern, freckles, fabric texture). `gpt-image-2` consistently produced a
   convincing *casting double* — similar, attractive, clearly inspired by the source, but visibly
   not the same face. For a preset whose entire job is identity anchoring, that decides it.

2. **Both render text cleanly now.** No garbled labels, no misspelled headings on either model —
   the old "use GPT for anything with text" rule no longer holds at these tiers. `nano-banana-pro`
   went further: its color palette had *named* swatches with HEX values that matched the actual
   outfit and skin tones.

3. **`gpt-image-2` follows layout instructions most completely.** It rendered every requested
   panel heading and the most uniform panel sizing; `nano-banana-pro` merged one panel heading into
   a neighbor on its run. The factory board prompts now include an explicit *"render ALL panel
   headings, never merging or omitting a panel"* clause to close that gap.

4. **Resolution and speed are not equal at the same setting.** At "2K", `nano-banana-pro` returned
   2752×1536 while `gpt-image-2` returned 2048×1152 — and GPT took ~1.8× longer on the dense board
   (210s vs 118s). The cheaper-per-image model is not cheaper per usable pixel.

5. **No identity blending on multi-character grids.** Both models kept the two cast members
   cleanly separated in their labeled strips with correctly spelled names — the Cast Mega Grid
   technique is robust across providers.

## Practical tips

- **The source photo is the #1 quality lever.** Sharp, evenly lit, front-facing, one subject. A
  soft or filtered photo is the most common cause of distorted or drifting faces.
- **Regenerate, don't settle.** If a face drifts on a board, regenerate the board — don't carry a
  drifted board forward as a reference. Two or three takes usually lands a clean one.
- **Feed the board back.** A board only pays off when you attach it as a reference image to later
  generations — including video (the *Scene Recipes* presets in Generate Video are built to consume
  these boards).
- **Boards for people, grids for machines.** The editorial styling on a board (dark background,
  neon accents, labels) is noise when a model consumes it as a reference; that's exactly why the
  clean grids exist. When in doubt: show the board to your team, feed the grid to the model.

See the [factory preset catalog](./nodes/presets.md) for the full list of boards, grids, and the
Scene Recipes that consume them.