Case Study: Guapa Juice
AI Workflows for Powerful Food Photography
| Julien Vandanjon-Rancoule | Case Studies

-
Client : Guapa Juice
-
Date : Spring 2023
-
Mission :
-
Packshot Photography
-
Food Photography
-
AI Automation
-
AI Upscaling
-
Catalog Art Direction
-
Deliver Dynamic yet Consistent Recipe designs for Kiosk Self-Service Screens
For a food photographer, creating a pop effect via motion is highly technical. It involves both a perfect knowledge of lighting techniques at high speed to freeze droplets and liquids movements in a splash, and perfect timing to catch the point of contact of fresh food chunks with the cup and its content.
The Challenge
Considering the number of recipes available in the Guapa Juice Catalog, the amount of shots required to get a perfectly consistent series of splashes—while shooting all kinds of fruit and vegetable chunks into accurately textured and colored juices—requires a HUGE amount of trials and errors. Hence, the cleaning of the studio and overall waste of food due to many attempts needed to catch the perfect composition for each recipe... and quite some luck.
Methodology
Considering the amount of defects and uncertainty of the production process of a full catalog of "liquid explosions" based on a traditional workflow, we decided to bring in experimental Artificial Intelligence Workflows based on available literature on this topic by March 2023.

The client actually shortened up the moodboarding process and instead came with a specific—and already sketched—idea of the final expected rendering. This helped the research and development phase

Still Shooting Products to Nurture the AI Learning Curve
From Midjourney to Firefly via Topaz AI
Three tools have been shortlisted to handle the process: first of all, Midjourney AI (Discord Version on private memory aware channel), as it appeared to be the most consistent and reliable creative iteration tool of its time. However, it had limitations (obsolete by now) of delivering only low-resolution assets (less than 2 megapixels) and not yet allowing local adjustments on each iteration.


To properly nurture the learning curve of the AI, a pro license was required, allowing to reset and sharpen
- A first set destined to teach the model what type of assets we expect while describing objects such as "pineapple slice", "banana chunk", or "little carrot with leaves".
- A second set picturing the juice cup with the actual recipe combinations around it.

The model must not only learn the color tiny variations between really close tones (less than 3% tone-to-tone error margin) and also include a texture index (base on bubble ratio per squar inch). Then It must learn the actual "catalog name" the brand attributed to each specific blend.

The Half-Cut Cup issue
Finally, a third bunch of pictures were made specifically to explain to the model how the container (paper cup) would be cut to reveal the inner mix and showcase the client logo in a clean and consistent framing.


Guapa AI Learning Process 04 Mid Journey Iterations

Guapa AI Learning Process 04 Mid Journey Iterations 02

Guapa AI Learning Process 04 Mid Journey Iterations 03
Iterating & rewarding the model when the liquid action is consistent with the guidelines
Then the iterations began (600 on average) to create a consistent art direction within the model for each recipe. The validated outputs were then upscaled to 12 megapixels using Topaz Gigapixel AI, allowing high-quality postprocessing using Adobe Neural Engine and Adobe Photoshop AI Assistants for local corrections (now called Firefly).
At delivery, all assets are downscaled to 2Mpx again and compressed to optimize loading time and ensure a smooth experience using the touchscreen devices.
The Result
From brief to Production Delivery, the actual full process took place in a matter of weeks instead of months. It avoided a significant amount of studio cleaning and potato-tube launcher reloading (even if it sounds funny, it can become a real assistant killer when reproduced hundreds, if not thousands of times).
Along the full process, this included more "traditional" operations for

In fine, even if the full process seemed to take the same amount of energy as if it had been produced in a more conventional way, its effectiveness and the ability to generate new assets almost immediately made it worth the effort.
For the future, this kind of workflow may be even further improved by using cycles of iterations assisted at other steps, such as prompting (involving GPT-4) and the opportunity to animate the results (such as custom workflows available in Runway AI).






