AI-Powered Image Generation for Customizable Grocery Products

instacart · E-commerce · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Grocery retail product imagery

Instacart's FoodStorm needed a practical way for grocery retailers to create images for customizable prepared-food products where professional photography for every filling, topping, side, banner, and thumbnail was impractical.

Constraints:

Architecture

Instacart integrated generative image creation into FoodStorm through its internal Pixel text-to-image service. Pixel abstracts access to multiple image models, including Google Imagen, and includes acceptable-use controls for prompts. FoodStorm added a Create AI Image dialog on top of the existing image upload component so retailers can enter and refine prompts, preview multiple generated variations, inspect a larger selected image for errors, accept terms of use, and publish optimized images across Instacart customer and in-store surfaces.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  foodstorm-oms["foodstorm-oms"] --> image-upload-component["image-upload-component"]
  image-upload-component --> create-ai-image-dialog["create-ai-image-dialog"]
  create-ai-image-dialog --> prompt-refinement-ui["prompt-refinement-ui"]
  create-ai-image-dialog --> terms-of-use-gate["terms-of-use-gate"]
  create-ai-image-dialog --> pixel-service["pixel-service"]
  pixel-service --> acceptable-use-controls["acceptable-use-controls"]
  pixel-service --> google-imagen["google-imagen"]
  create-ai-image-dialog --> variation-preview["variation-preview"]
  variation-preview --> large-detail-preview["large-detail-preview"]
  large-detail-preview --> post-generation-processing["post-generation-processing"]
  post-generation-processing --> content-delivery-network["content-delivery-network"]
  content-delivery-network --> platform-publishing["platform-publishing"]
  platform-publishing --> foodstorm-oms
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
foodstorm-oms Order Management System where grocery retailers manage customizable prepared-food products and imagery. image-upload-component, create-ai-image-dialog, platform-publishing
image-upload-component Existing FoodStorm image workflow extended so upload locations can also create AI images. foodstorm-oms, create-ai-image-dialog
create-ai-image-dialog Retailer-facing dialog for entering prompts, generating variations, previewing selected images, and accepting terms. image-upload-component, prompt-refinement-ui, pixel-service, variation-preview, terms-of-use-gate
prompt-refinement-ui Prompt entry/editing surface that lets users specify product details such as ingredient type, cut, and background. create-ai-image-dialog
pixel-service Instacart internal text-to-image service with model access and acceptable-use prompt controls. create-ai-image-dialog, google-imagen, acceptable-use-controls
google-imagen Image-generation model noted for food components and white-background product imagery. pixel-service
acceptable-use-controls Prompt governance layer inside Pixel for acceptable image-generation use. pixel-service
variation-preview Best-of-n style preview surface where users compare multiple generated options. create-ai-image-dialog, large-detail-preview
large-detail-preview Larger inspection view for spotting artifacts and image errors before publication. variation-preview, post-generation-processing
terms-of-use-gate Legal/compliance step requiring users to agree to terms before using AI image generation. create-ai-image-dialog
post-generation-processing Compresses and scales generated images for web performance. large-detail-preview, content-delivery-network
content-delivery-network Serves generated images efficiently to Instacart surfaces. post-generation-processing, platform-publishing
platform-publishing Publishes accepted images to Instacart App, Storefront, Storefront Pro, and FoodStorm Kiosk. foodstorm-oms, content-delivery-network

Data flow

A FoodStorm retailer opens the existing image upload flow and chooses Create AI Image. The retailer enters a prompt, ideally with product form and background details such as sliced monterey jack cheese on a white background. FoodStorm calls Pixel, which routes to image models such as Google Imagen under acceptable-use prompt controls. The UI returns multiple variations, the retailer selects one, expands the preview to inspect details and artifacts, and only then publishes. After generation, FoodStorm compresses and scales the image, serves it through the CDN, and makes it available across Instacart App, Storefront, Storefront Pro, and FoodStorm Kiosk.

4 supporting quotes
“The new Create AI Image dialog box builds on our existing image upload component — so anywhere that users could previously upload an image, they could now instantly create an AI image.”article, stated
“Pixel provides easy access to a range of different AI models, including Google Imagen, and has controls around acceptable use of prompts to generate images.”article, implied
“Users can also preview multiple image variations to choose the most appropriate result”article, implied
“compressing and scaling the images post-generation to optimize them for web use, and serving them efficiently through our content delivery network.”article, stated

Why this architecture

Instacart used an internal model abstraction and embedded UI because the hard problem was not just model access; it was making image generation controllable and usable by retailers inside their existing product-image workflow. Pixel supplied governed access to image models, while FoodStorm supplied prompt refinement, multiple variations, larger previews, legal gating, and CDN-ready publishing.

4 supporting quotes
“We were lucky to have a dedicated text-to-image service, called Pixel, already operational.”article, stated
“the main challenge was to create the user interface design to allow retailers to create AI images themselves.”article, implied
“builds on our existing image upload component — so anywhere that users could previously upload an image, they could now instantly create an AI image.”article, implied
“AI can be a significant time saver and dynamic creative tool. We’re eager to see how our grocery retailers embrace FoodStorm’s AI capabilities.”article, implied

Models used: Google Imagen · Pixel text-to-image models

Scale

Alternatives considered (roads not taken)

Professional photography for every configuration

Rejected because: Professional photography remains best for real products, but sourcing images for every customizable filling, topping, side, and configuration is impractical.

“nothing can replace the authenticity of professional photographs showing the real product.”article, stated
“for visualizing various fillings, toppings and sides, Generative AI can be a huge help.”article, stated

One-click generation from ingredient names

Rejected because: Simple prompts like cheese were too vague about product type, form, and presentation.

“Simply using the filling name as the prompt, like “cheese,” was too vague — what type of cheese, and sliced or grated?”article, stated

Get Anthony Super’s stories in your inbox

Join Medium for free to get updates from this writer."

— article, stated

First generated image as the final answer

Rejected because: AI image models can miss details or produce artifacts, so users need multiple variations and inspection before publishing.

“Users can also preview multiple image variations to choose the most appropriate result”article, stated
“occasionally you’ll get artifacts that may be slightly off. As we all know, AI doesn’t always nail it on the first try.”article, stated

Patterns

model-routing (model access)

FoodStorm does not integrate directly with one image model; it calls Pixel, an internal service that provides access to multiple text-to-image models including Google Imagen.

“Pixel provides easy access to a range of different AI models, including Google Imagen”article, stated

safety-policy-enforcement (safety)

Pixel includes acceptable-use controls for prompts, and FoodStorm requires users to agree to terms before using the image-generation feature.

“has controls around acceptable use of prompts to generate images.”article, stated
“users must also agree to relevant terms of use before using this feature.”article, stated

human-in-the-loop-gate (quality)

Retailers choose among multiple variations and inspect a larger preview before publishing generated images.

“Users can also preview multiple image variations to choose the most appropriate result”article, stated
“This allows users to take a close look at the details of the image, as occasionally you’ll get artifacts that may be slightly off.”article, stated

prompt-versioning (prompting)

The case does not describe versioned prompt artifacts, but it does show prompt refinement as a first-class UI behavior: users move from vague ingredient names to detailed prompts.

“fine-tune their prompts, specifying details like, “sliced monterey jack cheese on a white background.””article, stated

latency-optimization (delivery)

Post-generation compression, scaling, and CDN delivery optimize generated images for web use after a human accepts them.

“compressing and scaling the images post-generation to optimize them for web use, and serving them efficiently through our content delivery network.”article, stated

Tradeoffs

Use AI for component and promotional imagery

Chose generated images for fillings, toppings, sides, hero images, promotions, category thumbnails, and banners over photographing every customizable configuration.

Rationale: Generative imagery covers long-tail visual gaps where full photography would be expensive or impractical, while real product photos remain preferred for authenticity.

3 supporting quotes
“sourcing images for every possible product configuration.”article, stated
“for visualizing various fillings, toppings and sides, Generative AI can be a huge help.”article, stated
“FoodStorm uses Generative AI to create hero images, promotional content, category thumbnails and more.”article, stated

Embed generation in the existing upload flow

Chose Create AI Image inside the image upload component over a separate standalone image-generation tool.

Rationale: Embedding the capability where users already manage product images lowers workflow friction and makes generation available anywhere upload was already supported.

“builds on our existing image upload component — so anywhere that users could previously upload an image, they could now instantly create an AI image.”article, stated

Expose prompt details to retailers

Chose user-controlled prompt refinement over one-click generation with hidden prompt assumptions.

Rationale: Retailers need to specify product form and visual context to get images that match their actual offerings.

“Our initial design allowed for one-click AI image generation for any sandwich fillings”article, stated
“Simply using the filling name as the prompt, like “cheese,” was too vague”article, stated

Gotchas (what bit them)

Domain shorthand makes bad prompts.

Root cause: Retailers know what cheese means in context, but the model needs explicit details about type, preparation, and visual presentation.

Resolution: The UI lets users fine-tune prompts with specific product and background details.

“Simply using the filling name as the prompt, like “cheese,” was too vague — what type of cheese, and sliced or grated?”article, stated
“fine-tune their prompts, specifying details like, “sliced monterey jack cheese on a white background.””article, stated

Generated food images can contain plausible but wrong artifacts.

Root cause: Image models are nondeterministic and may misinterpret the subject or add unexpected objects.

Resolution: FoodStorm provides multiple variations and a larger inspection preview so users can reject flawed generations.

“AI doesn’t always nail it on the first try.”article, stated
“An example of the AI model getting it wrong — is that an Australian Bilby nibbling those raspberries?!”article, stated

AI images should not be positioned as a replacement for real product photography.

Root cause: Generated images can help visualize components, but real products need authentic photography for trust and accuracy.

Resolution: Instacart frames the feature as a complement for fillings, toppings, sides, hero images, banners, and thumbnails with quality controls.

“nothing can replace the authenticity of professional photographs showing the real product.”article, stated
“it’s essential to balance its use with genuine photography and other creative channels.”article, stated

Evaluation

Evaluation is primarily human-in-the-loop: internal tests validated Google Imagen's ability to generate food components and white-background images, then users preview multiple variations and inspect selected images for artifacts before publishing. No quantitative model metrics are provided.

Metrics: human selection of generated variations · manual artifact inspection · white-background suitability · terms acceptance before use

3 supporting quotes
“Our tests show that AI excels at creating these kinds of images — think perfectly diced onions, finely shredded lettuce, or just the right amount of strawberry jam.”article, stated
“Imagen also does a great job of creating images “on a white background”, which is a challenge sometimes for other AI tools.”article, implied
“Users can also preview multiple image variations to choose the most appropriate result”article, implied

Outcomes

The source gives no adoption counts, latency, cost, or conversion metrics. It does state the feature went through three UI iterations and publishes across four named Instacart surfaces.

3 supporting quotes
“This feature alone has enabled us to produce stunning promotional visuals effortlessly.”article, stated
“AI-generated images can be published across the Instacart Platform, including the Instacart App, Storefront & Storefront Pro and FoodStorm Kiosk for in-store ordering.”article, stated
“AI can be a significant time saver and dynamic creative tool.”article, stated