How do you structure the pipeline?

The reliable pattern is reference-first:

1. Start with one clean, high-resolution studio shot as the master reference.
2. Define a fixed style sentence — lighting, color grade, camera feel.
3. For each scene, keep the product description identical and change only the setting.
4. Generate 3-5 candidates per scene, then select by eye.
5. Run every selected image through a similarity check against the master.

Step five is what most teams skip. Comparing the generated product against the original reference, even with a simple perceptual hash, catches the drifts before they ship.

What prompts keep the product identical?

Separate the immutable part from the variable part. “The ceramic mug with the matte navy finish and visible handle” stays byte-for-byte in every prompt. Only the background and lighting change. Adding style adjectives to the product clause is how identity creep starts — resist it.

What is the cost per set?

Twenty final images, generated five candidates deep and filtered, means roughly 100 generations. At typical per-image API pricing in the cents range, a full set lands in the low single-digit to low double-digit dollars. That compares to a physical reshoot costing hundreds plus studio time, which is why the pipeline is worth building even with some manual selection.

What are the failure modes?

• Logo and text on packaging render inconsistently — add them in post if they must be exact.
• Reflective and transparent products (glass, chrome) drift more than matte ones.
• Aggressive scene changes can alter lighting on the product, breaking the set.
• Low-resolution references produce soft, unshippable outputs.

For packaging with mandatory text, compositing the real label over the generated scene is more reliable than prompting for it.

Where does this fit in a content stack?

The image pipeline is one stage. Teams pairing it with an automated editor — the pattern behind tools like droid for content repos, or a scripted CMS uploader — can go from one shot to a published set without manual file handling. Keep the human in the loop at the selection step; that is where quality is decided.

Quick Summary

• Build sets from one master reference and a fixed style sentence.
• Add a similarity check to catch product drift before shipping.
• A 20-image set costs roughly low single to double-digit dollars.
• Glass, chrome, and in-image text are the weakest cases.
• Composite exact logos in post rather than prompting for them.

Last updated September 2026.

What is Nano Banana 2?

It is a multimodal image generation and editing model. You give it an input image plus a prompt, and it returns an edited version — change the background, restyle an object, add a character in a consistent look. It also generates from scratch. The headline feature is identity preservation: the same face or product stays recognizable across edits.

How do you edit an image with it?

The workflow is short:

• Upload the source image as the reference.
• Write an instruction that names what changes and what stays.
• Ask for a specific output aspect ratio and style.

Prompts that work name the subject and the change separately: “keep the woman’s face and jacket; replace the street background with a plain studio wall.” Vague prompts like “make it nicer” produce generic results.

How do you keep a character consistent?

1. Generate one strong reference image and keep it.
2. Reuse that image as input for every subsequent scene.
3. Change only the scene description, never the subject description.
4. Lock the style wording — if you say “cinematic” once, say it every time.

This is the practical trick behind consistent storyboards and product shots. Narrative consistency does not happen automatically; it comes from holding the reference fixed.

What does Nano Banana 2 cost?

Access depends on the surface. Inside Google’s consumer tools it is bundled with a subscription. Through the API, image generation is billed per image, typically in the cents range per output depending on resolution. Reusing an input reference in an edit is metered similarly to a fresh generation.

Where does it fall short?

Fine text inside images is still unreliable — expect occasional misspelled signage. Hands and fine jewelry remain a weak spot. And heavy edits can drift a subject’s likeness, so for brand-critical work, generate several candidates and pick. It is a strong tool for iteration, not a replacement for a careful retoucher on final assets.

Quick Summary

• Nano Banana 2 is Google’s image generation and editing model.
• Edits take an input image plus a text instruction.
• Identity consistency comes from reusing a fixed reference image.
• API pricing is per-image, typically cents per output.
• Weak on in-image text and hands; iterate and pick the best result.

Last updated September 2026.