German Barquero

Excited to spend the week at #CVPR2025 in Nashville! 🇺🇸 If you’re interested in learning more about MixerMDM, come check out our poster at the following sessions: 📍 💃 [@humogen11384] Wednesday 11, 16:00-18:00, ExHall D #430 📹 [AI4CC] Thursday 12, 17:00, ExHall D #412 👑 Saturday 14, 10:30-12:30, ExHall D #165 Feel free to reach out if you want to chat, connect, or just say hi! ☺️

Why do we even need to learn how to mix motions? 🤔In our previous work (DualMDM), we found something very interesting 💡: when mixing motions throughout the denoising process, it actually works better to mix them differently at different steps. Our initial hypothesis was that early in the denoising process, interpersonal dynamics (how people relate to each other) matter more 👫. Later on the denoising, interpersonal dynamics (how a person moves individually) become more important 💃. So, we proposed different mixing schedulers that adapt across steps 🧠📈. This approach worked better than others, but it still had limitations. While some schedulers performed better on average, none of them worked perfectly across all motion combinations 🎭. Different motion pairs needed slightly different schedulers to perform well. 💥 That’s what motivated MixerMDM! We introduce the Mixer — a lightweight transformer module that predicts a mixing weight at every step of the denoising process, conditioned on the motions to be mixed and their textual descriptions 🌀🤖. This allows us to dynamically and consistently mix motions while maintaining strong individual controllability 💃🕺✨

📢 Introducing MixerMDM [#CVPR2025] — the first learnable model composition technique for combining pre-trained human motion diffusion models while preserving their core characteristics. 🟡➕🔵🟰🟢 MixerMDM enables consistent, fine-grained individual controllability in generating human-human interactions by: 🎛️ Mixer – A lightweight module that learns how to dynamically mix motions based on their specific characteristics and input conditions. 👩🏻‍🔬 Mixing – A model composition formula that combines motions dynamically through the denoising process, guided by Mixer predictions. 👮🏻‍♂️ Adversarial Training – A training strategy using adversarial losses, where predictions from the pre-trained models act as pseudo-ground truths. 🏅 Evaluation – A novel metric for measuring mixing quality, evaluating both individual and interaction alignment. ❤️ Huge shoutout to my amazing supervisors and co-authors: @Germs96 @cpalmeroc @SergioEscalera_ and José García-Rodríguez 🌍 Project: pabloruizponce.com/papers/Mi… 📚 Paper: arxiv.org/abs/2504.01019 🧬 Code: github.com/pabloruizponce/Mi…

Web Motion Visualizer 💃🕹️ An interactive visualization component to showcase your human motions on your website. Rendering locally in your browser without frame drops thanks to @threejs and @reactjs 🔥 Stay tuned for the MixerMDM paper release (hopefully) at the end of next week. Soon after its release, I will explain in detail how this component works and its code will be open-sourced.

We have been cooking something for the last few months 🥘🧪 Pre-print and website will be released in the following weeks. Stay tuned 👀