When people think about machine learning, the graves of graves may not come to mind. But in one of the most deep applications for AI Tollyth, we have built a system that uses automated learning to design customized souvenir products – including graves evidence – to personal preferences and cultural sensitivities.
This was not just a ml twisted experience. It was a complete application of obstetric models, the treatment of natural language, and human systems in the episode, all to address human needs sensitive and deep: celebrating life.
Problem: Design with dignity
The memorial design is art and traditions. Families want something personal and respectful, and often symbolic. The challenge is that the design process is slow, and the imposition of emotionally taxes, and is restricted by materials, cemetery regulations and religious or cultural traditions.
We have begun to create something that can help – non -replacement – designers: the birth of the grave who can produce realistic and meaningful design options based on previous data and customer preferences.
You can try it here: Headstonesdesigner.com/generator (all training data comes from the direct site – https://headstoneesgner.com/)
Step 1: Understanding the field
Before we touch Tensorflow or wrote one line of the code, we immersed ourselves in the world of memorial art. We studied:
- Traditional and contemporary methods
- Religious and cultural standards
- Material restrictions (granite, marble, etc.)
- Cemetery regulations, such as Max Monument offer for each piece
This was not optional. The design of artificial intelligence for a sensitive field requires such deep respect and differences. The error was not just an UX error – it was offensive.
Step 2: Building a data collection
We have gathered an amazing variety of data:
- Thousands of disclosed design pictures
- CAD files from the existing graves
- Customer preference date
- Text
- Dimensions Cemetery Standards
All of this needs to be clean, normalized and overlooked. The texts are included using models such as BERT. The images are equipped and increased. This was not only about receiving data in a model – it was about making it Learn can.
Step 3: Form Engineering and Training
We have tested some types of models in parallel:
- Stylegan2: To generate high -quality and elegant pictures of memorial designs
- Vaes (diverse automatic coding devices): To fulfill the design patterns and enable user -controlled differences
- Transformers (GPT): To generate the inscriptions that felt the personality, relevant and respected
A particularly difficult part is to make sure the text and visuals match. The Gothic Witness of the Gothic style should not have a comic inscriptions.
We took this with:
- Multimedia training: Combining vision and language models (like the clip) to assess the alignment
- Gas PoliticalUse the text as an entrance to direct the visual generation
Step 4: The Unknown Administration
We had a lot of moments of “strange intelligence”.
- Some early outputs look like modern sculpting from the memorial.
- Transfer the pattern sometimes via cultural lines in embarrassing ways.
- GPT is born from time to time of the deaf tone.
To alleviate this, we have built humanitarian notes in the episode. Cultural designers and advisors reviewed the outputs and the issues that have been marked. This feedback has returned to setting the form.
We also used techniques such as GANS patterns to impose restrictions and post -generation filters to check the text content.
Step 5: Evaluation and results
Not only did we only make the results. It was a multi -side evaluation:
- Fid grades The realism of the image
- Blood degrees And human evaluation of the accuracy of the text
- User studies and Expert reviews For aesthetic and cultural sincerity
The end result? The system that can generate emotionally, visually, visually accurate resonance designs.
You can interact with the generator here: Headstoneesgner.com/generator
The lessons learned
Some fast food:
- The cultural context is not a state of edge – it is the basic requirements in the areas of sensitive design.
- AI The Molidic is strong, but without restrictions, it easily drifting into a super or inappropriate area.
- Human reactions are not only useful; It is mandatory.
the future
We explore how this technique can extend to other areas: wedding invitation design, personal prizes, souvenir, and more. What place is a personal and high -risk design, there is an opportunity to mix ML obstetrics with human care.
