ID AI, built between June 4 and July 4, 2024, is a toolkit of generative AI resources to support industrial designers.
The inspiration came from a conversation with my classmate, Saloni, who was preparing an IDSA conference talk on generative AI tools she used at work. She shared several problems with these products that seemed easily solvable. One tool claimed to create a moodboard from a prompt but instead produced a disjointed image with 'moodboard' in the prompt. While it superficially resembled a moodboard, upon closer inspection, it was nonsense. Another tool failed to accurately apply specific materials to the objects she was generating.
Inspired by these gaps, I created and launched four demos:
1. Image Generator + Moodboard
The initial step in design is to create a moodboard to represent a direction and potential design language. In this demo, the moodboard on the left constrains and guides the generated image results.
2. Seamless Texture Generator
Designers often spend considerable time in Photoshop creating seamless patterns for Keyshot textures. This demo takes an image and generates an infinitely repeating texture..
3. In-Context Images
Another tedious Photoshop task is making renders look realistic within the context of photographs. The challenge here is ensuring every detail remains the same while only the lighting changes, as the lighting affects every pixel.
4. Trained Model (Color, Material, Finish)
This model is fine-tuned on color, material (rubber, acrylic), and finish (gloss, matte). Creating it wasn't scalable, as I manually assembled a dataset of 100 images using resources like material sample catalogs. This ensured that the training images closely reflected the real-life materials users would be designing with.
These demos are designed to support existing workflows, not to build tools that can run autonomously.
Results
2,388 images were generated by 271 users.
These results are just from word of mouth and sharing on Reddit. It averages out to each user generating ~9 images. In reality, that number is skewed towards a handful of returning users.
Process
My work was split between user interviews and building or improving the demos. A cycle would look something like this:
- User interview
- Run quick experiments in Google Colab to show what’s possible
- Show the user interviewee these options
- Build what they said was most useful
- Launch
- Test and repeat
Resources I used to move quickly:
- Gradio
- HuggingFace
- Replicate
- Squarespace
And my favorite part, Saloni included it in her presentation!
Interested in following along?
Site: idai.tools
Newsletter: idai.tools/newsletter
Follow on instagram: instagram.com/idai.tools
