The Way @ Runway

The Way @ Runway

Nov 3, 2019
I spent most of my time this weekend working on knowSound, which I thought I might be able to finish training and bring into Runway by yesterday, while now I seriously doubt if I can finish it by the end of the semester. (2019.11.04)


As a platform developed for artists and designers, Runway is well designed to be user-friendly with a beautiful interface. It's still true nowadays that technology sounds some kind of restricted area to the majority of creatives, and the idea of “making technology accessible to artists (or everyone)", just like the collapsible paint tube empowering painters with more and easier to obtain colors, is definitely of great demand.

Compared to other softwares we've used through the whole semester, Runway is definitely a winner. Because of its friendly GUI alone, it's already way easier to use than models outside of it that we need to configure and run ourselves with CLI; and better than ml5.js since you don't need to code at all and look up the documentations. And because of it being customizable, that users can add new models as they like or even train new models in the future, it has much more application scenarios than Teachable Machine.

It's not hard to find that the software is still in its fast growing phase and 3 new versions had come out during the past two weeks when we started getting our hands on it. While I do have some suggestions regarding to the design choices made to its interface:

  • The Welcome screen, where users are asked to choose from starting with current workspaces and models, seems unnecessary since the switch tab for these two pages are already at the highest level in the main software window. Adobe Creative Suite treats the problem with another solution: the majority of starting page are thumbnails of rent works (would be workspaces in this case), while on the side users can also choose to start with new canvas or templates.

  • This is more like a concept question: How wrapped up should the software be eventually? I also want to take Adobe as an example. Photoshop and Illustrator are also developed for creatives, yet actually have a quite steep learning curve that people need to spend hundreds of hours to use them as a master. This might be the inevitable trade-off of any tool platform and is related to the targeted user groups - everyone or professionals.

Way to Grasshopper

Not many news about how machine learning helps industrial design or product design has caught people's eyes. It might because of the difficulty of connecting old-fashioned modeling tools to latest ML models, or it's just because of the cooling-down of the whole industry as a result of the raise of the demanding UI/UX design. Runway, however, can at least solve the first problem.

Hence, for this week's small experiment, I decided to use Runway to feed data into Rhino's most famous plug-in - Grasshopper, to knock open the door between industrial design (my original field) and machine learning. This could also be a good kick-off of my undecided final project: Machine Learning for Computational Design.

Since data flowing through Grasshopper are only either numbers or Rhino.Geometry objects, which as far as I'm concerned none of Runway models could return right now, I decided to try with Face-Landmarks, which are supposed to return straight forward coordinates of key points around the face.

Script Overview

Currently there're no nice and clean way to directly visit Runway from Grasshopper, which is also what I'd like to contribute to the runwayml/RunwayML-for-Grasshopper repository. So I need to make an HTTP request - just as in p5.js - but a little harder as Grasshopper wasn't originally designed for streaming data. I found a piece of C# code and tuned into the way that works best with Runway. (Actually, Grasshopper also supports Python scripts, but installing one requests package into Rhino's Python environment took me forever.)

After loading data (sample data), which is in JSON format, into Grasshopper with C# component, I pass it into another Python component, read point locations from it, and draw Rhino.Geometry with it. The script is now on GitHub.

Script Points (Above GIFs might take some time to load.)


From this post I'll divide References that are related to topic and concept from Technical References, which is a composite of how to learning materials I came across during the process of accomplishing the final outcome documented in the posts.

  1. runwayml/RunwayML-for-Grasshopper is waiting for its first PR!
  2. davisking/dlib
  3. Machine Learning En Plein Air: Building accessible tools for artists

Technical References

  1. Crow Tutorial1: Neural Sensing and Control in Grasshopper
  2. Basic HTTP request - Grasshopper Forum
  3. Install Python packages in Rhino - Grasshopper Forum
  4. C# Guide
  5. How to Parse JSON into a C# Object
  6. ghJSON (Grasshopper extra component)
  7. What Every Computer Scientist Should Know About Floating-Point Arithmetic