What is the difference between a reference board and a consistency grid?

A reference board is an editorial sheet for people, with a hero shot, labels, panels and a color palette. A consistency grid is a clean grid of the subject from standard angles, with no text, made to be fed back to a model as a reference.

Which model makes the most consistent reference boards?

In our tests the Nano Banana family kept the same face most reliably. Nano Banana Pro is the default for boards, and Nano Banana 2 is the default for grids because it is nearly as faithful at a lower price.

What kind of photo should I start from?

One sharp, evenly lit, front-facing photo with a single subject. A soft, filtered or crowded photo is the most common cause of a face that drifts.

How do I use a board in a video?

Connect the board image to the reference input of an image or video node, or pick one of the Scene Recipes presets in Generate Video, which are built to read these boards.

Reference boards — keep a character, product or place consistent in AI images and video

Build a reference board or a consistency grid from one photo, pick the right model, and reuse it so a person, product or place looks the same everywhere.

A reference board is a single image that captures everything about one subject — a person, a pet, a product or a place — so that every later generation can match it. You make it once from a photo, then attach it as a reference to every image and video that features the subject. Nodaro has two kinds: reference boards for people to read, and consistency grids for models to read.

Boards or grids

	Reference boards	Consistency grids
Presets	Character, Pose, Location, Product, Outfit, Scene, Creature, Vehicle, Food, Mascot, Pet Board	Character Reference Grid, Cast Mega Grid, Cast Scene
What you get	An editorial sheet: a hero shot, a details block, labeled panels (views, expressions, details, lighting) and a color palette with HEX codes	A clean grid of standard angles on a neutral background, with no text
Made for	People: art direction, briefs, a project's visual bible	Models: an identity anchor for later generations
Default model	Nano Banana Pro, 2K, 16:9	Nano Banana 2, 4K, 3:4

Show the board to your team. Feed the grid to the model.

Make a board

Start from one good photo

Add an Upload Image node with one sharp, evenly lit, front-facing photo of the subject. This photo matters more than any setting.

Add a Generate Image node and choose a preset

Wire the photo into the References input of a Generate Image node. Open Presets on the node and choose a board, such as Character Board, or a grid, such as Character Reference Grid. The preset fills in the prompt, the model and the frame.

Run it, and run it again if the face drifts

Run the node. If the face on any panel does not match the photo, run it again. Two or three runs usually give a clean result. Do not carry a drifted board forward.

Reuse it everywhere

Wire the result into the References input of every image or video node that shows this subject. In Generate Video, the Scene Recipes presets are built to read these boards.

One grid, made once from a photo, anchors every later image and video of the subject.

The Reference Board node is a dedicated way to make a board in one pass and refine it with a mask, and the Reference Sheet node builds a sheet whose panels you can reuse one by one.

Cast several characters in one grid

For a story with two to four characters, build one Cast Mega Grid: each character gets a labeled strip in the same image. Then, in each scene, refer to the characters by the names on the grid instead of describing them again. The less you describe a face, the less it drifts.

Which model to use

Job	Model	Why
Any reference board	Nano Banana Pro (default)	Keeps the same face across panels best, and renders the labels and HEX palette cleanly
Grids you feed back to models	Nano Banana 2 (default)	Nearly as faithful as Pro, faster and cheaper, which matters because consistency work takes many runs. 4K keeps faces sharp when reused
Sheets where the layout matters more than the likeness	GPT Image 2	Follows multi-panel layout instructions most completely, but the face drifts
Labels and annotations on an image	GPT Image 2	Strong at following editing and labeling instructions

What our tests found

We ran the same prompts from the same photos on different models, one generation per test, in June 2026. Treat the results as a direction, not as statistics.

The Nano Banana family keeps the face. Nano Banana Pro and Nano Banana 2 reproduced the same person: face shape, stubble, freckles, fabric texture. GPT Image 2 produced a convincing look-alike, clearly inspired by the photo, but not the same face. For an identity anchor, that decides it.
Both write text cleanly. Neither model garbled the labels or headings. Nano Banana Pro also named its palette swatches with HEX values that matched the real outfit and skin tones.
GPT Image 2 follows the layout most completely. It rendered every requested panel heading at the most even sizes. The board presets now ask every model to render all panel headings.
The same setting does not mean the same size or speed. At 2K, Nano Banana Pro returned 2752×1536 pixels and GPT Image 2 returned 2048×1152, and GPT Image 2 took about 1.8 times longer on a dense board.
Characters stay separate on a cast grid. Both models kept two cast members in their own labeled strips, with their names spelled correctly.

Tips

Fix the photo first. A soft, filtered or crowded photo is the most common reason a face drifts.
Regenerate instead of settling. A slightly wrong board makes every later image slightly wrong.
Keep boards away from the model when a grid will do. The dark backgrounds, colors and labels of a board are noise to a model that reads it as a reference.
Name what each reference is for. When you attach several references, reference roles tell the model what to take from each one: the face, the outfit, the background or the style.

Reference boards and consistency grids