Shreshth Saini

Email | Google Scholar | LinkedIn | X | GitHub | CV (2026)

I recently completed my PhD at Laboratory of Image and Video Engineering at The University of Texas Austin, advised by Prof. Alan C. Bovik. My research focuses on the Theoretical Foundations of Generative Models (e.g. Flows, Diffusion, and MLLMs) and their applications in Image/Video Generation, efficient sampling, image/video quality assessment, image/video editing, and inverse problems (eg: ITM). In my PhD, I collaborate with the YouTube/Google Media Algorithms team. I was a Student Researcher in the LUMA team at Google Research. Lately, I am working on video post-training and RL for video.

In a previous life, I was a Research Engineer (AI) at Arkray, Inc. and BioMind AI, Singapore, building scalable AI for medical image analysis. I also co-founded Short-X, a startup that cut long-form podcasts into bite-sized shorts; it didn't make it. During my undergrad, I was fortunate to be advised by Prof. Mengling Feng (NUS, Singapore), Prof. Aditya Nigam (IIT Mandi), and Prof. Anil K. Tiwari throughout my research.

If you want to work with me, please feel free to reach out. I'm happy to advise and collaborate. 📮 saini[dot]2[at]utexas[dot]edu

profile photo
The University of Texas at AustinLIVEGoogle ResearchYouTubeAmazonAlibaba
Scratch Pad (Blogs): Technical Notes & Insights. Explore →
Updates
  • 10/2026: JLD (paper, code) and MEND (paper, code) are released.
  • 07/2026: Serving as Program Committee member for AAAI 2027.
  • 07/2026: CachedSearch paper is out on arXiv, with the code and a blog post released.
  • 07/2026: Fleetcraft, GPU infrastructure for video generation, is open-sourced (code, blog).
  • 05/2026: LumaGuide, training-free HDR generation, is open-sourced with a project page and the code.
  • 05/2026: GameScope (Gaming VQA Dataset) accepted in IEEE ICIP 2026.
  • 04/2026: Successfully defended my PhD thesis.
  • 03/2026: BrightRate got Best (student) Paper Award at WACV 2026.
  • 02/2026: HDR-Q (Dataset+RL Method) accepted in CVPR 2026.
  • 01/2026: Serving as a reviewer for ICML 2026 and ECCV 2026.
  • 12/2025: BrightRate accepted for oral presentation at WACV 2026.
  • 12/2025: Completed Progress review.
  • 09/2025: Rectified CFG++ accepted to NeurIPS 2025.
  • 06/2025: Excited to join the LUMA team as Student Researcher at Google Research.
  • 05/2025: Paper accepted in IEEE International Conference on Image Processing, IEEE ICIP 2025.
  • 05/2025: Paper accepted in 42nd International Conference on Machine Learning, ICML 2025.
  • 05/2025: Paper accepted in 3rd Workshop on Generative Models for Computer Vision, CVPR 2025.
  • 03/2025: Honored to be appointed as Assistant Director of LIVE at UT Austin.
  • 06/2024: Joined the Perception team as Applied Scientist-II Intern at Amazon.
  • 01/2024: Joined as Research Intern to work on diffusion models at Alibaba US.
  • 11/2023: Paper accepted in 3rd Workshop on Image/Video/Audio Quality in CV and Gen AI, WACV 2024.
  • 06/2023: Completed first-ever large-scale HDR subjective perceptual quality study on Amazon Mechanical Turk.
  • 08/2022: Joined as PhD student under Prof. Alan C. Bovik at LIVE at UT Austin.
  • 08/2022: Awarded prestigious Engineering Graduate Fellowship until 2027.
  • 02/2022: Joined as Research Engineer (AI) at BioMind AI, Singapore.
  • 08/2020: Started as Research Engineer (AI) at Arkray, Inc.
  • 05/2019: Research Assistant under Prof. Mengling "Mornin" Feng at National University of Singapore (NUS).
  • 08/2018: Undergraduate Researcher under Prof. Anil Kumar Tiwari at Image Processing and Computer Vision Lab, IIT Jodhpur.
  • 05/2018: Research Intern under Prof. Aditya Nigam at The Multimedia Analytics, Networks and Systems Lab, IIT Mandi.
Research & Development Experience
Research
Selected papers. The full list is on Google Scholar.
CachedSearch: nine cached candidates explored at once, then two winners delivered at full compute
CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion
S Saini, N Birkbeck, Y Wang, B Adsumilli, AC Bovik.
Preprint, 2026
ArXiv / Page / Code / Blog / Explainer video / Diffusers pipeline / ComfyUI node
DiffusionVideo generation

CachedSearch makes best-of-N test-time search in video generation affordable. Candidate rankings survive aggressive training-free caching, so every candidate is explored cheaply and only the winner is re-generated at full compute: 94.7% of best-of-8's quality gain at 63% of its cost, with any cache engine and any verifier.

LumaGuide: Distribution Shaping for Training-Free HDR Generation in Diffusion Models
B Chen, S Saini, B Adsumilli, AC Bovik.
Preprint, 2026
PDF / ArXiv / Page / Code / Weights
DiffusionHDR

LumaGuide is a training-free framework for distribution shaping in diffusion models. Instead of modifying model parameters, it steers the sampling process to match target feature distributions via differentiable energy-based guidance, instantiated for HDR generation by controlling luminance distributions in perceptually uniform PQ space. The framework supports data-driven presets, reference-image targets, and text-driven predictors, and extends naturally to video with temporal consistency constraints.

Fleetcraft: GPU Infrastructure for Video Generation Research
S Saini
Open-source release, 2026
Page / Code / Blog
Infrastructure

Fleetcraft is an open playbook and toolkit for running video generation research on GPU fleets. It packages the skill files, agent guides, Python utilities, spool and fleet scripts, and Slurm templates distilled from real diffusion training and RL projects, with telemetry and monitoring baked in.

BrightRate-LM: Representation-Aware Reasoning Quality Assessment for User-Generated HDR Video
S Saini, Y Wang, N Birkbeck, B Adsumilli, AC Bovik.
Machine Vision and Applications (MVA), WACV 2026 Special Issue, in submission
Page / Code / Models
MLLMQuality assessmentHDR

Journal extension of BrightRate (WACV 2026 Oral) that adapts multimodal LLMs into an HDR-UGC quality model producing a calibrated score with generated reasoning. A controlled study across Qwen2.5-VL, Qwen3-VL, and Gemma-4 covers tone-mapped, multi-exposure, and native PQ inputs; the multi-exposure 7B model surpasses BrightRate.

LumaFlux: Lifting 8-Bit Worlds to HDR Reality with Physically-Guided Diffusion Transformers
Shreshth Saini, Hakan Gedik, Neil Birkbeck, Yilin Wang, Balu Adsumilli, Alan C. Bovik.
Preprint, 2026
Paper / Code / ComfyUI node / Weights / Project page / Blog
DiffusionHDR

LumaFlux is an 8-step, prompt-free SDR-to-HDR converter that is best on every Luma-Eval metric. Its training-free video stabilization cuts signed excess PU21 flicker by 53.3%.

Seeing Beyond 8bits: Subjective and Objective Quality Assessment of HDR-UGC Videos
S Saini, B Chen, N Birkbeck, Y Wang, B Adsumilli, AC Bovik.
The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026
PDF / ArXiv / Page / Code / Dataset
MLLMRLQuality assessmentHDR

We present the first large-scale crowdsourced subjective study and a novel RL trained MLLM-based objective QA baseline for HDR-UGC videos, enabling reliable reasoning-based assessment of real-world HDR content. We also propose a novel RL training strategy "HAPO" to learn HDR aware features.

BrightRate: Quality Assessment for User-Generated HDR Videos
S Saini, B Chen, Y Wang, N Birkbeck, B Adsumilli, AC Bovik
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026 (Oral, Best Student Paper Award)
PDF / Page / Dataset / Code
Quality assessmentHDR

BrightRate is designed for quality assessment in user-generated HDR videos, focusing on capture variability and perceptual fidelity.

Rectified CFG++ for Flow Based Models
S Saini, S Gupta, AC Bovik.
Thirty-Ninth Conference on Neural Information Processing Systems (NeurIPS 2025)
PDF / ArXiv / Page / Code / Diffusers pipeline
Diffusion

Rectified CFG++ enhances conditional image generation with Rectified Flow models by adaptively correcting the latent trajectory. This method improves visual coherence and alignment with text prompts, outperforming existing samplers in generation quality and efficiency.

LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations
S Saini, R Liao, Y Ye, AC Bovik.
International Conference on Machine Learning (ICML) 2025
PDF / ArXiv
DiffusionQuality assessment

This study investigates the exploitation of diffusion model priors to achieve perceptual consistency in image quality assessment. By leveraging the inherent priors learned by diffusion models, the assessment of image quality is made more aligned with human perception, leading to more accurate and reliable evaluations.

HIDRO-VQA: Deep Contrastive Representation Learning for HDR Video Quality Assessment
S Saini, A Saha, AC Bovik.
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2024
PDF / ArXiv / Page / Code
Quality assessmentHDR

Contrastive HDR-VQA introduces a deep contrastive representation learning approach for high dynamic range video quality assessment. By learning robust representations through contrastive learning, the method achieves state-of-the-art performance in predicting the quality of HDR videos.

(M)SLAe-Net: Multi-Scale Multi-Level Attention Embedded Network for Retinal Vessel Segmentation
S Saini, G Agrawal.
IEEE ISBI 2021 & IEEE ICHI 2021 (Oral)
IEEE/PDF / ArXiv
Medical imaging
B-SegNet: Branched SegMentor Network for Skin Lesion Segmentation
S Saini, YS Jeon, M Feng.
ACM CHIL 2021 (Oral)
PDF
Medical imaging
Detector-SegMentor Network for Skin Lesion Localization and Segmentation
S Saini, D Gupta, AK Tiwari.
NCVPRIPG 2019 (Oral)
PDF / ArXiv
Medical imaging
PixISegNet: pixel-level iris segmentation network using convolutional encoder–decoder with stacked hourglass bottleneck
RR Jha, G Jaswal, S Saini, et al.
IET Biometrics, 2019
PDF
Biometrics
Datasets and benchmarks
Sample clips from Beyond8Bits
S Saini, B Chen, N Birkbeck, Y Wang, B Adsumilli, AC Bovik.
The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026
Datasets & benchmarksQuality assessmentHDR

The largest crowdsourced HDR user-generated video quality dataset. Every source is transcoded along a bitrate ladder and rated about 35 times on Amazon Mechanical Turk; MOS comes from SUREAL aggregation. Released with HDR-Q, an RL-trained multimodal LLM baseline.

41,419 clips · 1.46M human ratings · 5,917 HDR sources
CVPR 2026YouTubeGoogleLIVE, UT AustinThe University of Texas at Austin
Sample clips from BrightVQ
S Saini, B Chen, Y Wang, N Birkbeck, B Adsumilli, AC Bovik.
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026 (Oral, Best Student Paper Award)
Datasets & benchmarksQuality assessmentHDR

The subjective study behind BrightRate, the WACV 2026 Best Student Paper: HDR-UGC sources transcoded along a bitrate ladder and rated by crowd workers. The benchmark for no-reference HDR-UGC video quality models.

2,100 clips · 73,794 human ratings · 300 HDR sources
WACV 2026 Best Student PaperYouTubeGoogleLIVE, UT AustinThe University of Texas at Austin
Sample clips from CHUG
S Saini, N Birkbeck, Y Wang, B Adsumilli, AC Bovik
IEEE International Conference on Image Processing (ICIP) 2025
Datasets & benchmarksQuality assessmentHDR

The first crowdsourced HDR-UGC video quality dataset: phone captures in portrait and landscape with authentic capture and compression distortions.

5,992 clips · 211,848 human ratings · 856 HDR sources
ICIP 2025YouTubeGoogleLIVE, UT AustinThe University of Texas at Austin
Sample clips from GameScope
R Sureddi, S Saini, A Saha, AC Bovik.
IEEE International Conference on Image Processing (ICIP) 2026
Datasets & benchmarksQuality assessment

The largest gaming video quality dataset to date, spanning user-generated and professionally created content across H.264, H.265, and AV1, with a vision-language baseline that outperforms prior VQA methods.

4,048 videos · ~37 ratings per video · 3 codecs
ICIP 2026LIVE, UT AustinThe University of Texas at Austin
Prime-EditBench: A Real-World Benchmark for Image and Video Editing Tasks using Diffusion Models
S Saini, P Korus, S Jin.
Amazon Prime Air / Perception (Internal)
PDF
Datasets & benchmarksDiffusion

Prime-EditBench is a real-world benchmark designed to evaluate image and video editing with diffusion models, enabling standardized assessment of editing performance.

Book / Book Chapters
Iris Segmentation in the Wild using Encoder-Decoder based Deep Learning Techniques
S Saini, et al.
AI and Deep Learning in Biometric Security, CRC Press, 2020
PDF

This book chapter explores the use of encoder-decoder based deep learning techniques for iris segmentation in unconstrained environments. The proposed methods effectively handle challenges such as variations in lighting, occlusion, and off-angle images, making them suitable for real-world biometric applications.

Professional Service
Reviewer and Program Committee Member
AAAI2027
ICLR2025, 2026, 2027
ICML2025, 2026
NeurIPS2025, 2026
CVPR2025, 2026, 2027
ICCV/ECCV2025, 2026
WACV2024, 2025, 2026, 2027
IEEE ICIP2025, 2026
IEEE Transactions on Image Processing (TIP)2025, 2026
IEEE Transactions on Multimedia (TMM)2024, 2025, 2026, 2027
TMLR2025, 2026
Other Service
Assistant Director, LIVE at UT Austin2025-2026
Volunteer, Internal Workshop on Deep Learning (IWDL), India2018
Established and ran LAMBDA Lab at IITJ2018-2020
Overall Head, Entrepreneurship and Innovation Cell at IITJ2018-2019
Assistant Head, Counselling Services at IITJ2018-2019