[x: 1400, y: 0]

// publications

Published Work

[x: 0, y: 0]conf: 0.98
ICLR 2026| first-author

Evolutionary Caching to Accelerate Your Off-the-Shelf Diffusion Model

Anirud Aggarwal, Abhinav Shrivastava, Matthew Gwilliam

International Conference on Learning Representations (ICLR) 2026

We introduce ECAD, an evolutionary algorithm to automatically discover efficient caching schedules for accelerating diffusion-based image generation models. ECAD achieves faster than state-of-the-art speed and higher quality among training-free methods and generalizes across models and resolutions.

PaperArXivCodeProject Page
loading plot data...
Hover to ExploreHover over any blue ECAD point to view generated images
ECAD discovers efficient caching schedules for diffusion models through genetic algorithms
[x: 1024, y: 480]
CVPR 2026

UPLiFT: Efficient Pixel-Dense Feature Upsampling with Local Attenders

Matthew Walmer, Saksham Suri, Anirud Aggarwal, Abhinav Shrivastava

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026

We introduce UPLiFT, a lightweight, iterative feature upsampler that converts coarse ViT and VAE features into pixel-dense representations using a fully local attention operator. It achieves state-of-the-art performance on segmentation and depth tasks while scaling linearly in visual tokens, and extends naturally to generative tasks for efficient image upscaling.

PaperArXivCodeProject Page
conf: 0.96
UPLiFT 4x super-resolution resultBilinear upscale (blurry)
4x super-resolution in latent space
512×512 → 2048×2048
Bilinear
Latency: 1.29s (NVIDIA A100)
UPLiFT
Latency: 1.40s (NVIDIA A100)
PCA visualization: Input Image, DINOv2 Features (32x32), and DINOv2 + UPLiFT (448x448) showing how UPLiFT recovers fine-grained spatial detail from coarse ViT features
confidence:exploratory
CMSC 848R[0.82]

Fast & Faithful: Diffusion Drift

Anirud Aggarwal*, Omkar Pathak*, Nayana Gadde*

Language Model Interpretability (Instructor: Sarah Wiegreffe)

Fast & Faithful: Diffusion Drift visualization

Do accelerated diffusion language models reason faithfully? We introduce a framework for measuring Diffusion Chain-of-Thought (DoT) faithfulness and analyze how train-free acceleration affects reasoning dynamics in LLaDA-8B and dLLM-Cache on GSM8K.

PaperProject Page
CMSC 472[0.78]

Learning to Settle: Reinforcement Learning in Catan

Anirud Aggarwal, Jinhai Yan, Serena Huang, Rohit Kommuru, Monish Napa, Han Lin

Introduction to Deep Learning (Instructor: Abhinav Shrivastava)

Learning to Settle: Reinforcement Learning in Catan visualization

We build a custom PettingZoo environment and training stack for learning to play the board game Catan with reinforcement learning. Starting from a refactored Settlers-RL codebase, we explore both multi-agent methods (via MARLlib) and a single-agent PPO baseline with dense reward shaping. Our experiments show agents that learn to play shorter, higher-scoring games, while highlighting the remaining gap to robust multi-agent performance in non-stationary, multi-player settings.

PaperCodeProject Page