Pouyan Navard

After years of asking “is this publishable?”, I’m increasingly interested in asking “does this actually work?”

Building some new set of skills

Email  /  CV  /  Scholar  /  LinkedIn  /  Github  /  HuggingFace

Publications

Research Interest: Generative AI.

Portrait animation diffusion models stack trained modules for expression, pose, and identity to compensate for conditioning on RGB, where these axes are inseparable. Loki replaces that with an identity-orthogonal parametric face model on the driver and lightweight key–value injection for identity — using ~43% fewer inference parameters and 1,496× less training video, with cross-ID reenactment reducing to a coefficient substitution at inference.

LLaVA-LE captioning and QA on lunar imagery.

We introduce LLaVA-LE, a vision-language model for lunar surface and subsurface characterization. We curate LUCID, a new dataset of 96k panchromatic images with scientific captions and 81k QA pairs from NASA missions. Fine-tuned with a two-stage curriculum, LLaVA-LE achieves a 3.3x gain over base LLaVA, with reasoning scores exceeding the judge's own reference.

Right-click and open the image in a new tab for better resolution

SeamCam-based camouflage image generation vs. SOTA.

We introduce SeamCam, a camouflage evaluation metric that quantifies how detectable an animal is from visual evidence. SeamCam achieves 78.82% agreement with human judgments, outperforming state-of-the-art by ~25%. We further use SeamCam as a preference signal for DPO fine-tuning of diffusion-based inpainting models for camouflage generation, and introduce CamFG-1.5k, a curated benchmark of 1,521 high-resolution images for unbiased evaluation.

Sketch-to-image generation with adjustable detail.

We introduce KnobGen, a dual-pathway framework that bridges the gap between novice sketches and expert-level image generation. Our system dynamically balances fine-grained detail and high-level control using adjustable modules, producing high-quality results from any sketch.

3D medical image segmentation results.

SegFormer3D redefines 3D medical image segmentation with a lightweight hierarchical Transformer that rivals state-of-the-art models. By blending multi-scale volumetric attention with an all-MLP decoder, we achieve competitive accuracy with 33x fewer parameters and 13x lower compute.

Design and source code from Jon Barron's website.