Delivery Hero
Delivery Hero generates missing menu photos with self-hosted Stable Diffusion
Delivery Hero fine-tuned and self-hosted an open-weight Stable Diffusion model to create food images for menu items that had none. Its AI lead says an A/B test showed a 6% to 8% rise in menu-to-cart conversion when a product had an image, at under 0.3 cents per image.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
6%
to 8%
Increase in conversion rate from menu to cart when an image is added to a product, in an A/B test1
85%
Reduction in the time it takes to compute an image after optimization (baseline in seconds not stated for this figure)1
less than 0.3 cents
Cost per generated image on the self-hosted setup (the team cites 2 cents per image with the OpenAI service it used before)1
What it took
The problem
Many menu items on food delivery platforms have no photo. Delivery Hero's data analysis found many products with no image, and the speaker says only 14% of products were bought without images. Taking or buying photos for every dish at every restaurant is slow and costly, and it varies by country.
What they built
Delivery Hero built a system that generates a food image for a menu item from its description. A first version used a hosted image service from OpenAI. The company then moved to a self-hosted Stable Diffusion model, an open-weight image model, meaning its files are published so a company can run and adapt them. One of the best setups in its tests was a community variant called Juggernaut V9 with an IP adapter. Restaurant owners can see suggested images and pick the most relevant one.
How they did it
The team fully fine-tuned the model, taking several days on an A100 graphics processor, because it needed to cover thousands of local dishes. This improved local dishes but made some others, such as pizza, less appealing. For serving, it picked L4 graphics processors for the best cost per image and sped up inference with compile-time optimization. Production runs on AWS in Kubernetes, scaling from 0 to 36 nodes with KEDA.
Automatic checks reject images with people or creatures, text, poor framing or blur. Offline tests used 1,000 products spread across entities and categories.
Results
The speaker reports a 6% to 8% increase in conversion from menu to cart when a product had an image, based on an A/B test. Time per image fell to 11.2 seconds, which the speaker calls an 85% decrease. Cost fell to under 0.3 cents per image, against 2 cents per image for the OpenAI service the team used before. More than 100,000 products now have generated images.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Increase in conversion rate from menu to cart when an image is added to a product, in an A/B test | 6% to 8%1 | |
| Image generation time per image after optimization | 11.2 seconds1 | |
| Reduction in the time it takes to compute an image after optimization (baseline in seconds not stated for this figure) | 85%1 | |
| Cost per generated image on the self-hosted setup (the team cites 2 cents per image with the OpenAI service it used before) | less than 0.3 cents1 | |
| Products that have generated images on Delivery Hero platforms | More than 100,0001 |
Lessons for your project
Test the business case before building at scale. Delivery Hero first analysed its menu data to decide where generative AI would matter most, then measured the lift from images with an A/B test.
Start with a hosted service for a quick prototype, then self-host when volume makes unit cost matter.
Build automatic quality checks early. Delivery Hero screens for text, people, blur and poor framing, and lets restaurants choose among suggested images.
Keep infrastructure in one cloud where you can. The speaker called splitting services across two clouds a mistake that delayed delivery.
Caveats
- All figures come from a conference talk by the Delivery Hero employee who leads its AI solutions development, who speaks for the company. They are not independently confirmed.
- The speaker gives inconsistent counts of generated images (more than 100,000 and then around a million), so only the count of products with generated images is used.
- The 85% reduction is stated without the baseline in seconds. The only baseline in the talk, about three minutes per image for default Stable Diffusion on a T4 GPU, is a different setup.
- The talk gives no go-live date and no team size or overall timeline. The deployed date shown is the recording date printed on the InfoQ page (the announcement date); the actual go-live is earlier and unknown. The talk refers to the Talabat IPO as last year.
- The A/B test result is stated as a range without sample size or duration. Delivery Hero says images are marked as for reference only.
- The model is an image model, not a text model. Juggernaut V9 is a community variant of Stable Diffusion, so the vendor attribution to Stability AI is an approximation.
Sources
- AI for Food Image Generation in Production: How & Why
InfoQ (QCon London 2025, talk by Iaroslav Amerkhanov of Delivery Hero) · 2025-08-27 · Conference talk · infoq.com
Summarized in our own words from the sources above. Spotted an error? See corrections.