Rewind: Impressions of the first Pimcore User Group in Berlin

On 22.03.2023, the time had finally come: Titled "Reality Check AI: Automated Content Production at Enterprise Scale with Pimcore", the official Pimcore User Group | Berlin took place at the Basilicom headquarters in Berlin. More than 70 Pimcore users and other interested parties met to learn more about how to use generative AI in content production with Pimcore.

The Event officially started with a few opening words by Basilicom founder Arndt Kühne and Pimcore CEO Dietmar Rietsch to welcome the guests, followed by two presentations by Prof. Peter Kabel and Basilicom CTO Christoph Lühr.

Creative AI: A Perfect Storm - How to weather the demise of the content industry

Prof. Peter Kabel is a founder, and professor at HAW Hamburg and has been working on creative AI for years. For his latest project, cogniwerk.ai, his team is collecting and categorizing the different creative AI models. Users can discover applications that suit their specific needs from around 200 tools.

Using current examples, he showed that AI models have found their way into our everyday lives a long time ago and how ChatGPT, Midjourney & Co. are turning the content industry upside down, challenging even global corporations and their business models. Whether it's filters on TikTok and Snapchat or automatically editing photos taken with mobile phones, in his view "machine learning" is already a determining factor for our behavior.

The new AI models are going to achieve the same with our way of working. In the near future, there will be a suitable AI model for every task, every topic, and every style. Competitive advantage will be based on the ability to use creative AIs correctly. Prompt crafting, the ability to write ideas and concepts as input for an AI that delivers the desired result, will become one of the most important skills in years to come. A new market for prompting tools is already emerging from the hype around AI. Those who cannot operate AIs sufficiently will have to pay for tools providing specialized interfaces to remain competitive.

The "AI wars" have already begun. Alphabet (Google) has entered the race with Bard, a competitor to ChatGPT. Microsoft has invested billions in Open AI and has launched a challenger to Adobe with Microsoft Designer, who themselves have already sent an AI-based product into the test phase called "Adobe Firefly". The conclusion of the presentation. Anyone who creates content digitally will inevitably come in touch with generative AIs. Experts who master the new tools will be even more efficient and deliver better results. Less complex tasks with high costs will be automated in the future.

Automated Content Production at Enterprise Scale with Pimcore

Christoph Lühr is CTO at Basilicom and a Pimconaut from the very beginning. For the Pimcore User Group, he connected AI tools to Pimcore in order to create content automatically. In his presentation and demo, he did a "reality check" on whether current tools can be used to produce content in bulk with the required quality.

The aim of his tests was to use product data from a PIM to create texts and product images that could be used in a catalog or online shop. Prompts were generated using existing product information and sent to ChatGPT for texts and DALL-E and Midjourney for images.

The results were already impressive, but not suitable to be used for the respective products in a shop. The problem with the texts was that ChatGPT kept adding information in addition to the product information from the prompts. This meant that the generated texts were good, stylistically, but the content was overall not correct because it contained misinformation regarding the product or details were missing.

The results are already equal or better in quality compared to texts of many existing providers but are not usable without being manually edited. Thus ChatGPT can already save a lot of work in connection with Pimcore, but it does not yet offer the necessary scalability to automate content for thousands of products. Just as with content, the automatically generated images did not deliver the required quality. Although appealing images can be generated with existing product information, results clearly deviate from the original product.

Original image (left) vs. DALL-E (right)

A second test, conducted with a trained AI delivered better results. For this purpose, 35 different product images were created using a 360° image in order to train a model from the company Mindverse.

The prompt, therefore, could rely on a known product to be used to set the scene. Unlike the creative AI, the trained model results are almost confusingly realistic, but there are still problems with the logo and some product details compared to the original product.

For companies with sufficient resources and training data, such models could be an interesting alternative in the future, but for most companies, they currently still don’t provide the necessary scalability.

Generative AI needs Guardrails to work effectively

The tests have shown that the AI models can already generate content at a very high quality, but product data as the sole basis for prompts does not deliver the desired results, to enable automated content production at a large scale. As explained by Prof. Kabel, these prompts do not provide enough context to generate the quality of manually created texts and photos. To bridge this gap, Christoph integrated two more tools into Pimcore to present them in a demo. Pebblely automatically generates backgrounds for images with common AI models such as Stable Diffusion and recognizes the size and orientation of a product to properly stage it.

Retresco allows companies to create templates with individually definable rules to generate texts. With the help of structured and unstructured data, web, product, and SEO texts can be generated automatically with the desired quality, variance, and relevance. The team around Retresco CEO Johannes Sommer themselves was present at the user group to join the demo and answer questions.

The images and texts generated in this way were confusingly similar to manually created content. In the case of Pebblely, the tool already manages well to recognize a towel, create suitable backgrounds based on a prompt and put the product in context. Although the product appeared to be rather small in the tests, together with the appropriate text and product information, the generated images could easily have been used in a shop.

The results for the texts were even better. Thanks to the template, based on an existing product text, the texts contained only product information that was stored in the PIM.

The rules defined by Christoph and the dynamic content elements ensured that newly generated texts, even for the same product, were unique.

Dynamic Template and Generated Text

The integration directly in Pimcore made it possible to generate content live within seconds of a click.

The conclusion of the reality check: Content can already be automatically generated en masse in high quality, but not yet with popular AI models like ChatGPT, Stable Diffusion, or Midjourney still. Companies either need the resources to train these models to their needs or use complementary tools and Pebblely and Retresco to achieve the desired quality.