Pablo Dawson
Hi, I'm Pablo. I teach computers to make pictures, videos, and reconstruct the real world. ML engineer in Santiago, Chile
Hi, I'm Pablo. I teach computers to make pictures, videos, and reconstruct the real world. ML engineer in Santiago, Chile
Generative AI for fashion retail. I develop the face swap, virtual try-on, video generation and detail-fixer pipelines that run in production for 14 retail clients, including Falabella, Ripley and Paris. Face Swap V2, built with the team, took client approval from 82.7% → 97.3%. I also made the internal Playground where every pipeline gets tested first (~400 of its 415 commits).
Visual engineering studio for spatial imaging and volumetric video (Munich / Eindhoven, remote from Chile). A volumetric viewer for web and VR, an in-house Gaussian Splatting pipeline, and VR productions in Germany and the Netherlands.
Image and video diffusion (SDXL, AnimateLCM, IP-Adapter) for the art exhibition “Ecosistemas Híbridos”, with an API that gave access to the models during the show.
Researched and trained video depth estimation models for fine detail and temporal consistency.
A VR player for volumetric video on Meta Quest and SteamVR, plus converters that turn monocular, VR180 and 360° video into it. Built solo, 10,000 downloads.
Inference server for Dekorai, a room visualizer, and virtual try-on prototypes for digital retail.
One photo in, a seamless 360° camera orbit out, landing back on the first frame. Trained on orbits rendered from Gaussian splats. 38k+ downloads · 550+ likes
Diffusion depth estimation for video, reusing previous frames through optical flow. Less flicker, and 0.6 s vs 2 s per frame.
Depth-regularized 3D Gaussian Splatting from only 5 views. Test PSNR 8.55 → 19.45.
Pruning that cuts the number of Gaussians needed for dynamic scenes, plus an exporter to Unity.
Retrained LDM3D's UNet with mask channels so it inpaints color and depth together.
Two order-independent transparency methods for a Unity Gaussian Splatting renderer.
An LLM-assisted app that helps people with diabetes manage carbs, connected to glucose sensors. 2nd in votes
A fully volumetric music composition for 6DoF VR: a dance captured with 112 DSLRs, rendered as 4D Gaussian Splatting. I built the volumetric pipeline and the VR player, taking it from 10 GB → ~1.5 GB.
For Wunderkammer's viewer and the volumetric scenes we made.
A visual explainer of dynamic Gaussian Splatting, condensing dozens of papers into one read.
Lectures in CENIA's diploma program: GANs (2025); RNNs, tokenizers and seq2seq (2026).
Hi! I'm Pablo, an ML engineer from Santiago, Chile.
These days I build diffusion pipelines that put clothes and faces on people who never wore them: virtual try-on, face swap and video for fashion retailers, at RIAL and CENIA, Chile's national AI center.
Before that I spent a few years going the other way, turning flat video into 3D you can walk around in VR. First with Touchly, an app I built alone, then at Wunderkammer, the volumetric video studio I co-founded.
Long before any of this, from 2014 to 2019, I made music as Kays. It came out on international labels and got a mention on BBC Radio 1.