
Examples of knit code and renderings of the resulting patterns. Move operations are shown on the left, cross operations on the right.

Preprocessing of a pattern: End-of-line tokens (here shown in orange are added to the right side of the pattern. Start- and end-of-sequence tokens (here shown in blue) are added to the bottom left and top right. The pattern is then split by row and concatenated.

Making a swatch of generated patterns using our design tool
Knitting Equipment Support
Shima Seiki
DeepKnitMachine Learns and Machine Knits
DeepKnit
Collaborators
Summary
In this project, we worked on simplifying the complex process of programming modern knitting machines, which are capable of producing intricate textile products. Typically, programming these machines demands expert knowledge, but we developed a long short-term memory (LSTM) based deep learning model to address this. Our model generates low-level code for novel knitting patterns based on high-level style specifications. Instead of using image-based approaches, we described knitting instructions as one-dimensional sequences of tokens, making them processable by our model.
.
We integrated this into a design tool that allows users to assemble atomic patterns into larger swatches or garments. To evaluate our approach, we formalized requirements for syntactically correct and manufacturable patterns. Although the generated patterns appear more random and less like traditional human designs, our evaluation demonstrated that their knittability is significantly better than patterns generated randomly.