Miguel Angel Bautista

I'm a research scientist at Apple MLR in San Francisco, where I lead a small team that focuses on generative modeling research. At Apple I've worked on research on generative modeling methodologies for images, video, 3D, graphs and scientific problems. I did my PhD at University of Barcelona, where I was advised by Sergio Escalera. I spent a big part of my PhD at CMU working with Fernando De la Torre on matrix factorization methods. I did my postdoc training with Björn Ommer working on unsupervised deep learning and generative models.

Email  /  CV  /  Bio  /  Scholar  /  Twitter  /  Bluesky

Research Interests

I'm interested in scalable and efficient generative modeling approaches that make as little assumptions about data as possible. My long term goal is to unify training recipes and architectures across different data domains (image, text, 3D, graphs, video, etc.). My research style is to focus on simplifying overly complex pipelines and design decisions, and finding what are the key designs that makes things actually work.

Prospective Applicants

Im always on the look out for strong full time RS and PhD interns to join my team and work on generative models. If you think you can be a good fit please reach out via email and describe your biggest accomplishment.

Publications

Scales latent normalizing flows so they happily pump out crisp, megapixel images without breaking a sweat.

Teaches implicit neural representations to flow-match in ambient space, making continuous signals play nice.

Dispels the myth that flows lag behind by proving they can hit diffusion-level quality.

Knits diffusion with a global 3-D scene so every video frame keeps its bearings in the same universe.

Shows how to pre-train skyscraper-sized AR image models at you-won’t-believe-it scale and keep them stable.

Simplifies conformer sampling so chemists can spin up billions of candidate molecules before lunch.

Turns complex shapes into bite-sized tokens so transformers can snack on 3-D data.

Drops a tiny LoRA-style adapter that lets you steer or remix any text-to-image model on the fly.

Generalises diffusion to live directly on manifolds, so everything stays on-surface where it belongs.

Extends diffusion to continuous fields, letting you sample functions instead of pixels.

Uses diffusion models to guess value functions that keep robots from face-planting.

Kicks the tyres on dynamic NeRFs and asks if one-camera view synthesis is truly ready for prime time.

Stacks diffusion stages so each cleans up the mess left by the previous one.

Builds adaptive modules that reshuffle themselves to tackle tasks from trivial to torturous.

Lets a neural maestro draft entire 3-D worlds in one forward pass.

Trades a bit of fancy math for a big speed boost in NeRF-style renderers.

Squeezes the most geometry out of a handful of photos by optimising pose and shape together.

Generates scenes everywhere at once by letting every point consult its local pals.

Drops a massive synthetic indoor dataset so models stop over-fitting IKEA.

Bakes symmetry into NeRF so rotating the scene doesn’t scramble its predictions.

Nails down why some recon models wobble on unseen objects—and how to toughen them up.

Learns to sample whole image sets, not just singles, so galleries come pre-curated.

Aligns your training loss with evaluation metrics so your model stops chasing the wrong goal.

Figures out what looks alike without ever seeing a label.

Compresses targets so classifiers stop wasting neurons on zeros.

Learns visual hierarchies by arranging images into gentle ranking chains.

Clusters video frames by matching body poses, turning chaos into choreography.

Lets every image teach itself by forming little friendship cliques.

Breaks matrices into factors that double as error-correcting codes.

Uses body gestures to spot ADHD behaviour without sticking sensors on kids.

Mixes time warping with 3-D bag-of-words to keep gesture detection in sync.

Serves up 8 k limb annotations and a graph-cut method to slice them cleanly.

Evolves error-correcting codes that keep classifiers honest.

Speeds up DTW by sprinkling probability, making gesture matching less twitchy.

Combines 2-D and depth words so gestures pop out clearly.

Shows you don’t need bloated codes to get robust multiclass classifiers.