🤖 🗞️ Free AI Generated Digital Humans Newsletter -> Like everyone else, I've been struggling to keep up with the sheer volume of papers and news in the Digital Human space. Over the past few weeks, I've developed an agentic (ish) AI pipeline to find and

SnapMoGen: Human Motion Generation from Expressive Texts. TLDR: This is a method for animating digital characters from text. The model uses multi-scale tokenisation with a masked generative transformer. It is trained on a novel dataset (which should be released) consisting of 44

Want to get up to three papers like this directly to your inbox and summarised by AI each day? You can sign up for free here:

ScaffoldAvatar: High-Fidelity Gaussian Avatars with Patch Expressions. TLDR: A method for Gaussian Avatars which improves the ability to capture fine details by using a number of patches and a small MLP per patch to do independent expression-conditioning, rather than a single

Want to get up to three papers like this directly to your inbox and summarised by AI each day? You can sign up for free here:

Digital Salon: An AI and Physics-Driven Tool for 3D Hair Grooming and Simulation. TLDR: A tool for modelling hair, with a text interface, physics simulation and high-quality rendering system. With code released!. 📽️ Project Page: 📜 Paper:

Want to get up to three papers like this directly to your inbox and summarised by AI each day? You can sign up for free here:

VisualSpeaker: Visually-Guided 3D Avatar Lip Synthesis. TLDR: Improves speech-driven facial mesh animation by rendering with 3DGS Avatars and running the results through a lip-reading network to get lip sync as a loss. 📜 Paper:

Want to get up to three papers like this directly to your inbox and summarised by AI each day? You can sign up for free here:

HyperGaussians: High-Dimensional Gaussian Splatting for High-Fidelity Animatable Face Avatars. TLDR: Lifts Gaussians to higher dimensions, which can then be conditioned for the problem of Gaussian Avatars. This enables much better reconstruction of high-frequency details and

🤖📰Want to get up to three papers like this directly to your inbox and summarised by AI each day? You can sign up for free here:

ARIG: Autoregressive Interactive Head Generation for Real-time Conversations. TLDR: A method for dyadic conversations (speaking and listening) in real-time. This work uses an autoregressive diffusion MLP with added signals from context learning (by fusing LivePortrait parameters

Want to get up to three papers like this directly to your inbox and summarised by AI each day? You can sign up for free here:

SynMotion: Semantic-Visual Adaptation for Motion Customized Video Generation. TLDR: A DiT-based video generation model designed for motion retargeting. It uses a dual encoder architecture to separate subject and motion, and can be applied to non-human animals. 📽️ Project Page:

3DGH: 3D Head Generation with Composable Hair and Face. TLDR: A Generative model of (static) Gaussian Avatars using a GAN. Here the head and hair are modelled separately with templates, and data is generated synthetically. 📽️ Project Page: 📜 Paper:

💫 Animate any rig using video diffusion models. Not a human-specific method, but really interesting as an idea. AnimaX: Animating the Inanimate in 3D with Joint Video-Pose Diffusion Models. TLDR: Animate any rig with an arbitrary skeleton using multi-view video diffusion

Controllable and Expressive One-Shot Video Head Swapping. TLDR: A reference-net-based approach for one-shot face swapping. This model works with landmark and background conditioning, with the former disentangled using a 3DMM. 📽️ Project Page: 📜 Paper:

One-shot Face Sketch Synthesis in the Wild via Generative Diffusion Prior and Instruction Tuning. TLDR: This work converts photorealstic portraits into sketches. Despite limited data, it is able to do this by optimising text instructions for a diffusion image-to-image model. 📜

SyncTalk++: High-Fidelity and Efficient Synchronized Talking Heads Synthesis Using Gaussian Splatting. TLDR: A Gaussian-Splatting based person-specific talking head model which reduces the effect of OOD audio by training a discrete VQ-VAE over blendshapes. 📽️ Project Page: