Krishna Srikar Durbha

About Me

I am a fourth-year Ph.D. student at the Laboratory of Image and Video Engineering, The University of Texas at Austin, advised by Prof. Alan Bovik. I collaborate with the Meta Video Infrastructure Team on Video Engineering and Perceptual Quality Optimization.

My current research advances generative modeling for vision and Multimodal Large Language Models (MLLMs) — investigating few-step generation frameworks, MeanFlows, and visual grounding. Building on my prior work in representation learning and perceptual quality assessment (IQA/VQA), I design and train generative and multimodal models that are semantically faithful and perceptually high-fidelity.

Research Interests

  • Generative Modeling
  • Image and Video Representation Learning
  • Perceptual Quality Assessment
  • Multimodal Large Language Models
  • Computer Vision
  • Image and Video Processing
  • Video Streaming Optimization

Education

  • Ph.D. in Electrical and Computer Engineering
    2022 - Present
    The University of Texas at Austin
    Advisor: Dr. Alan C. Bovik
  • Bachelor of Technology in Electrical Engineering
    2018 - 2022
    Indian Institute of Technology Hyderabad

News

Ongoing ResearchActive

  • MeanFlows for Few-Step Generation In Progress

    Investigating MeanFlows as a framework for efficient few-step generation, training generative models that synthesize high-fidelity samples in a handful of sampling steps.

  • Visual Grounding of Multimodal LLMs In Progress

    Studying visual grounding techniques for Multimodal Large Language Models, improving how they localize and reason over visual content to produce semantically faithful, perceptually high-fidelity outputs.

Publications

Preprints & Under ReviewSubmitted

Published PapersPeer-Reviewed

First Author Equal Contribution

Experience

  • MPI Research Intern Summer 2025
    Samsung Research America

    Built a generative framework for high-resolution 4K image synthesis, scaling state-of-the-art pretrained diffusion models to 4096×4096 with structural refinement methods that suppress synthesis artifacts and maximize perceptual quality and spatial consistency.

  • Wireless Systems Intern Summer 2024
    Skydio

    Optimized drone-to-controller video streaming using adaptive streaming protocols and Video Quality Assessment (VQA), curating an empirical dataset and designing multi-tier frame-rate switching driven by visual features and real-time drone telemetry.

  • Application Development Intern Summer 2021
    Oracle

    Engineered an AI-powered conversational interface using Natural Language Understanding (NLU) for the Oracle Journeys application, integrating it with Oracle HCM Cloud via REST APIs and optimizing intent classification and entity recognition.

Projects

Time-Constrained Representation Alignment (TC-REPA)

Tested restricting Representation Alignment (REPA) loss to high-noise timesteps on SiT-XL/2 flow-matching models, establishing that the effective axis for relaxing alignment is the training iteration, not the diffusion timestep.

Diffusion Models Representation Alignment Flow Matching

Per-Prompt Evaluation of Guidance Strategies in T2I Models

Benchmarked seven CFG-family guidance strategies across SDXL, PixArt-Σ, SD3, and FLUX on T2I-CompBench++, revealing that aggregate FID/IS scores conceal compositional hallucinations exposed only by per-prompt distributional evaluation.

Text-to-Image Classifier-Free Guidance Compositional Evaluation

High-Resolution 4K Synthesis using Few-Step Models

Benchmarked diffusion, flow-matching, consistency, and flow-map-matching models for 4K image synthesis, quantifying the perceptual-quality and inference-efficiency trade-offs between knowledge-distilled and principled few-step techniques.

Few-Step Generation High-Resolution Synthesis Diffusion / Flow Matching