Google Retired Imagen 4: What Gemini 3.1 Flash Image Changes for Product Sellers

Google shut down the Imagen 4 models in the Gemini API on 17 August 2026. Its current deprecation documentation now points developers to Gemini 3.1 Flash Image, also known as Nano Banana 2, as the recommended replacement.

For product sellers, the interesting part is not that one model name disappeared. The workflow is changing. Imagen 4 was mainly a text-to-image generator. Gemini 3.1 Flash Image is built for generation and editing with image inputs, multi-reference composition and fast iterative changes. That makes the replacement much more relevant to ecommerce teams that already have photographs of the real SKU.

The biggest change is that the product can stay inside the workflow

A text-to-image model is useful for concepts, backgrounds and campaign ideas, but it starts from a description. If you ask it for a silver curb-chain bracelet, it still has to design a bracelet that matches those words.

Gemini 3.1 Flash Image can instead take images as part of the prompt and edit or combine them. For ecommerce, that means the genuine product photograph can become the starting evidence. You can ask for a new scene, crop, composition or controlled presentation change without describing the SKU from scratch.

That does not make the model a perfect product lock. It does make the workflow more sensible. Use text generation when you are exploring what an advert could look like. Use image-based generation when the exact item being sold needs to survive into the result.

Up to 10 high-fidelity object references changes difficult product jobs

Google's current image-generation documentation says Gemini 3.1 Flash Image can use up to 10 object reference images with high fidelity in one workflow. You rarely need ten photos for one SKU, but the capacity matters because one product image often cannot show everything a model needs to know.

A jewellery seller could provide a hero view, clasp close-up, side profile and engraving detail. A cosmetics seller might provide the front label, side shape, cap and packaging. An electronics seller could include the front, rear and sides where ports or controls are visible.

The useful rule is not to upload more references simply because the limit is higher. Add a reference when it supplies information the primary image does not contain. Every image should show the exact same SKU unless the prompt clearly assigns it a different role, such as a composition or style reference.

Editing is more valuable than regenerating every product scene

Google introduced Nano Banana 2 in February 2026 as a faster image generation and editing model designed for high-fidelity visual creation at scale. That editing capability is where the migration becomes most useful for sellers.

Suppose you already have an approved product photograph but need three campaign versions. One needs a warm stone surface, one needs extra empty space for copy, and one needs a vertical social crop. Rebuilding the product from a fresh text prompt three times gives the model three opportunities to invent a different object.

A better workflow keeps the approved product image as the anchor and changes the presentation around it. When a result is already accurate, make the next edit as narrow as possible rather than regenerating everything. This is especially important for jewellery, packaging and other products where small geometry or text differences can define the SKU.

The model is designed for volume, but production still needs checks

Google positions Gemini 3.1 Flash Image as the high-volume, high-efficiency image model in the Gemini 3 family. The current developer guide lists image output pricing below the higher-quality Gemini 3 Pro Image tier, with pricing varying by resolution.

That matters when one product needs many listing and campaign assets. Faster, lower-cost generation makes it practical to create several candidates, test different layouts and repeat a workflow across a catalogue.

But lower cost can encourage a bad habit: generating more images than anyone properly reviews. A ten-image batch is not ten usable product images. It is ten candidates that still need to be checked against the source SKU.

A practical migration test for ecommerce teams

If an existing workflow used Imagen 4, do not judge the replacement with one attractive prompt. Pick one difficult real product and compare the workflow around it.

Start with a genuine source photograph and record the details that must survive. For a bracelet, that might include clasp shape, chain pattern, engraving, dimensions and metal finish. For a bottle, include label text, cap geometry, colour and quantity.

Then test three jobs separately: create a new lifestyle scene from the real reference, make one local presentation edit to an approved image, and create a second composition using additional real reference views. Keep the brief simple enough that you can see where an error came from.

Review product fidelity before scene quality. If the clasp changes or the label is wrong, reject the generation even if the lighting is better. Then generate several fresh candidates from the same setup to see whether the workflow is dependable rather than merely capable of one good result.

There is one important migration caveat

Google's current Gemini 3 developer guide still lists Gemini 3.1 Flash Image as a preview model. That is worth noticing because the same deprecation documentation recommends it as the successor to Imagen 4.

For developers, preview status means model behaviour, limits or identifiers can still change. Record the exact model used for approved catalogue work, keep original product references, and retest important image workflows after meaningful model updates. A saved prompt alone is not a complete production recipe.

What this means for Lustra Studio workflows

For a product-focused system such as Lustra Studio, the shift is useful because the underlying model is moving closer to the way sellers actually work. The seller already has real product images. The job is to preserve those references while creating new poses, scenes, colour variations and consistent asset sets.

Multi-reference image models make that structure more valuable. A workflow can keep the real jewellery photos as product evidence and give the model only the creative freedom needed for the requested output. The model still needs verification, but less of the product has to come from guesswork.

The takeaway

Imagen 4's retirement matters because Google's recommended replacement is not simply a newer blank-canvas generator. Gemini 3.1 Flash Image combines image generation with editing, accepts several real object references and is positioned for faster, higher-volume visual production.

For ecommerce sellers, that is a better direction. Start from the real SKU, supply extra views when they add missing evidence, edit approved images instead of rebuilding them unnecessarily, and treat every generated result as a candidate until the product itself passes inspection.

Sources: Google AI for Developers, Gemini deprecations documentation, updated August 2026; Google AI for Developers, Gemini image generation documentation, updated August 2026; Google Developers Blog, “Build with Nano Banana 2, our best image generation and editing model,” 26 February 2026; Google AI for Developers, Gemini 3 developer guide, updated August 2026.