Google’s Neural Head Model (GNM) in Houdini — Take 1

[00:00:00]Hi everyone. The last few days have basically reset everything I was working on because Google released GNM, the generative neural morphable model.

[00:00:10]GNM is a neural parametric model of a human head and it can generate different facial identities, expressions, size movements, head poses while preserving a consistent mesh topology.

[00:00:23]So, naturally I dropped everything and started bringing it into Houdini.

[00:00:29]Just to put this in the perspective, SideFX released Houdini 2020 about yesterday and I almost completely ignore it and ignore the whole event until I had a first working version of my GNM running inside Houdini.

[00:00:43]What are you seeing here? It's only the first step. For now, I have recreated the core functionality of the Google two demos, the semantic demo and the head demo, directly inside Houdini.

[00:00:53]Later on, I want to reinterpret the system, make it more procedural and start adding some actual Houdini power to it.

[00:01:00]But for now, the workflow starts by importing the static GNM base mesh together with its UVs, attributes, metadata, and point groups for the most important anatomical regions of the face.

[00:01:13]Next comes the semantic sampler.

[00:01:16]This semantic sampler is basically user-friendly way to explore the model without manually editing hundreds of coefficients.

[00:01:23]It allows me to generate different identity using a broad categories such as gender and ethnicity. Uh and it's also give me access to the predefined facial expression classes.

[00:01:36]Under the hood, it generates a complete set of identities and expression coefficients by combining a random light and latent sample with a selected semantic categories.

[00:01:47]So, the result is partly controlled and partly random, which makes it very useful for quickly exploring uh wide ranges of faces or generates random characters.

[00:02:00]I then combine the semantic sampling with manual identity and expression controls. This means I can start from a generated face, a random one, and continue adjusting individual components by hand.

[00:02:13]I'm still exploring exactly what every component does, but the model contains 253 identity components and 383 expression components. The identity section is divided in 170 components for the head and skin, three for the eyes, and 80 for the teeth.

[00:02:32]>> [sighs] >> To wrap things up, I think GNM could eventually become a very interesting open-source alternative to systems like MetaHuman, especially for artists who wants more direct control and more freedom to customize their underlying model.

[00:02:47]My next step is to explore whether the system can be remapped onto custom scans, so you could create your own character, like MetaHuman does, while preserving the GNM topology and deformation system.

[00:02:59]The model also seems to include something similar to facial landmarks, so it may make this process possible. I also want to investigate how I could connect to other systems, such as text-to-speech or audio-driven facial animation.

[00:03:15]In any case, this is still only the beginning. I think we are going to see some very interesting things built with it.

[00:03:23]So, until next time, download it, explore it. It is really something.

[00:03:28]See you. Bye-bye.