Pramish Paudel

I am a researcher at INSAIT, where I am fortunate to work with Prof. Luc Van Gool and Prof. Dr. Danda Pani Paudel.

Scientists found an ancient lake on Mars from an orbital image. The discovery was possible because the human mind can reason about cause and effect. My goal is to build world models capable of such causal reasoning, so that AI can make scientific discoveries too.

My UG thesis, iHuman (ECCV 2024), was the first top-tier vision paper from a Nepalese institute. I also co-founded sceneverse.ai (later widelabs.ai), working on 3D and video understanding for clients like Pilatus Aircraft (demos).

E-mail: pramish[dot]paudel[at]gmail[dot]com

Publications

* equal contribution

iSEE: Object Permanence Through Self-Supervision

Pramish Paudel, Ajad Chhatkuli, Luc Van Gool, Danda Pani Paudel

Under review Paper Project Code Thread

A self-supervised method that learns object permanence from video without labels; within 4.1 mAP of a supervised tracker on LA-CATER.
Vision Transformers Need Cross Aggregation for Patch Semantics (CALiT)

Ajad Chhatkuli, Pramish Paudel, Deheng Zhang, Kanzhi Wu, Luc Van Gool, Danda Pani Paudel

Under review Project

A loss on the CLS token sends most ViT patches nearly the same gradient. Cross aggregation fixes this and sharpens patch features, for 0.6% extra FLOPs on ViT-S.
iHuman: Instant Animatable Digital Humans From Monocular Videos

Pramish Paudel, Anubhav Khanal, Danda Pani Paudel, Jyoti Tandukar, Ajad Chhatkuli

European Conference on Computer Vision (ECCV) 2024 Paper Project Code Video

Animatable 3D human reconstruction from monocular video. My undergraduate thesis.
SceneBench: A Hierarchical Benchmark for Vision-Language Understanding of 3D Scenes

Anubhav Khanal, Prabigya Acharya, Roshni Poudel, Sujan Kapali, Bigyan Bhatta, Pramish Paudel, Francois Rameau, Danda Pani Paudel

Under review Paper

A benchmark for vision-language understanding of 3D scenes.
iARCS: Iterative Agentic RL for Controllable 3D Scene Generation

Pramish Paudel*, Saugat Adhikari*, Ashok Prasad Neupane*, Ajad Chhatkuli, Danda Pani Paudel

Under review; presented at the Synthetic Data for Computer Vision Workshop, CVPR 2026 Paper Project Code Data

Agentic RL that steers 3D scene generators toward language-specified constraints.
iADD: Improving Alignment and Diversity in Diffusion Policy Optimization

Pramish Paudel*, Saugat Adhikari*, Ashok Prasad Neupane*, Ajad Chhatkuli, Danda Pani Paudel

Under review Paper Project Code

RL post-training that improves both alignment and diversity of diffusion models.

Demos

Research prototypes from my work at sceneverse.ai (later widelabs.ai), including for companies like Pilatus Aircraft.

Understanding the technician's work

Understands first-person video of a technician working on an aircraft.
Parts and actions from CAD

Recognizes parts and the technician's actions in video, given only the CAD model.
Work instructions from CAD

Turns a CAD assembly into step-by-step work instructions.

Honors & Awards

INSAIT Summer Undergraduate Research Fellowship (SURF), 2024. Fully funded research fellowship for undergraduates worldwide; acceptance rate under 2%.
University Grants Commission (UGC) Grant, 2023. Brought ~US$30K from UGC Nepal to my department for my project; one of 18 recipients across Nepal.
Hult Prize Global Accelerator, Boston, 2022. Won the regional summit with MeasureMe.AI and reached the top 16 of 30,000+ teams competing for the US$1 million prize.
Merit-based Scholarship, 2021. Awarded for one semester, Department of Electronics and Computer Engineering, Pulchowk Campus.

Academic Service

Reviewer: ICLR, IEEE RA-L, ACM TOMM