Neural Monsters

Neural Monsters was the second stage of the research begun in Panelki. Could we automate more of the journey from a generated motif to knitting, while exploring a collection that combined physical garments and digital-only looks? Our model was trained on images of Japanese monsters from manga.

I identified and articulated the production problem, defined the task for our machine-learning specialist and directed development around the needs of the garment-making workflow. He implemented the solution and wrote the code. I led the development from a design and production perspective, working with the ML specialist and knitting programmer.

I also developed the collection concept, designed the knitwear and digital garments, created 3D looks and directed the visual identity and runway film. My research partner trained the model and generated characters. Other 3D designers contributed to the runway environment, animation and rendering, while specialists supported physical production. Full team credits are linked below.

The workflow generated original monster images, pixelated them and limited the palette to the required number of yarn colours. It also partially automated stitch preparation in M1. We tested the sweaters in 3D to review motif scale, placement, colours and the overall look before deciding what to manufacture.

We photographed our team wearing physical sweaters and added digital trousers and accessories to those photographs. Some people wore entirely digital sweaters in the final images, and some styles were never manufactured. We aimed to make the physical and digital garments visually indistinguishable within the campaign.

The collection produced physical knitwear, digital-only looks, a hybrid photoshoot and a Brazilian Digital Fashion Week film. Motivated by the sustainability discussion at the time, we explored whether clothing made for image-making could remain digital rather than requiring another physical product.

We did not achieve full automation. A knitting programmer still needed to inspect the images and programs, check and correct stitch instructions, and prepare the work for the machine. Automated image preparation reduced one part of the manual work, but the technical handover remained unfinished. This unfinished transition is a starting point for my proposed research.