Auto generating AI images

I have been exploring the auto generation of images to see if it could be an option to help a client add offers onto their website more efficiently.

The starting point was a catalogue of offers from suppliers. The information comes with no images. I wanted to explore AI image generation to see if it was a viable option for sourcing the images needed. So far I have tried 2 different image generators and wanted to share my learnings to date.

Tools trialled:

• Google sheets - holding the information of the items to promote

• Open AI: Dall-E - AI image generation (Experiment 1)

• Google - Vertex AI: Imagen-3 - AI image generation (Experiment 2)

• Cloudinary (Free version) - to store the images produced

• Make (Free version) - to automate the process and store the complete catalogue item

• ChatGPT (Premium) - tech support :)

Experience level: Basic

Outcome:

I was able to successfully implement an automated solution that:

• takes entries from the catalogue

• creates a dynamic prompt based on the information

• links to the AI Image generators that generates an image using the prompt

• uploads the image into Cloudinary

• stores a record of the catalogue item and the link to the associated image in a datastore ready to be used to load into the website

Challenge Level:

Considering I had little to no previous experience with most of the tools, with the assistance from ChatGPT, and my own engineering background I was quite surprised at how easy it was. Yes there was a bit of head scratching here and there whilst I figured out how to set-up each of the tools and one very frustrating typo in Make that lost me a chunk of time, but overall I was pleased with how it went.

Costs:

Support:

• ChatGPT - Monthly subscription (£18.99)

Implementation:

• Make - (Free)

• Cloudinary - (Free)

• Open AI - Dall-E-3 ($0.04 = £0.03 per image )

• Goggle - Imagen-3 - (Free - initial credits, $0.04= £0.03 per image after)

Review of output:

• Images (Dall-E-3 and Imagen-3): The images produced by both AI image generators where of a high quality, but I struggled to get the prompts right despite using ChatGPT to help me to rewrite them. The generators took too much creative liberty with the items I wanted pictures of, I needed specific makes, models of items, and so they could not be used for the purpose I intended.

Key takeaways:

• Frustratingly the prompts I used provided useable pictures when tried in ChatGPT and Google's Vertex AI console. My next experiment will be to add an additional AI call to my automation flow to try and refine my image prompt. This will add a small additional cost though probably around $0.005 = £0.004. I'll let you know how it goes.

• AI generated images are great from a copyright perspective and if producing smaller numbers and/or wanting specific tailored pictures. If stock images would be perfectly acceptable, these could potentially be sourced for free and I believe it is possible to automate finding them from sites like Unsplash, Pexels and Pixabay. I will be looking to do this so again I'll let you know how it goes.

• I used the free versions of most of the tools and for a lot of small businesses this would be more than enough to get started. If the limitations of the free version starts to become a problem the next packages up aren't too expensive especially if the automation is paying for itself back in time and quality.