Publications

Why Multi-Source Audio Fails to Compose: A Geometric Diagnosis and Its Remedy

Anurag Singh, Prajwal Singh, Shanmuganathan Raman

BMVC 2026, Lancaster, UKAlso at the BMVC 2026 Doctoral Consortium

This paper explains why multi-source audio-to-image generation fails. Audio tokens sit in misaligned regions of CLIP space, so objects get ignored or merged. CLAP-based tokens and spatially separated cross-attention fix this and give faithful images of every sound source.

Synthesizing Compositional Videos from Text Description

Prajwal Singh, Kuldeep Kulkarni, Shanmuganathan Raman, Harsh Rangwani

WACV 2026, Tucson, AZ, USA

A training-free method for compositional video generation from a text prompt, built on a pre-trained video diffusion model. Video-ASTAR optimizes the cross-attention maps with a centroid loss and beats baselines in extensive experiments and ablations.

PaperProject page

Incremental Multi-Scene Modeling via Continual Neural Graphics Primitives

Prajwal Singh, Ashish Tiwari, Gautam Vashishtha, Shanmuganathan Raman

BMVC 2025, Sheffield, UK

C-NGP encodes many 3D scenes in a single NeRF, conditioned on pseudo-scene labels. It learns continually with minimal forgetting and renders high quality novel views across scenes without adding parameters.

arXivProject page

BloomCoreset: Fast Coreset Sampling using Bloom Filters for Fine-Grained Self-Supervised Learning

Prajwal Singh, Gautam Vashishtha, Indra Deep Mastan, Shanmuganathan Raman

ICASSP 2025, Hyderabad, India

BloomCoreset stores and retrieves image features with Bloom filters for fine-grained self-supervised learning. It cuts sampling time from large unlabeled datasets by 98.5% at a 0.83% accuracy cost across 11 downstream tasks, and plugs into the SimCore framework.

arXivCode

Learning Robust Deep Visual Representations from EEG Brain Recordings

Prajwal Singh, Dwip Dalal, Gautam Vashishtha, Krishna Miyapuram, Shanmuganathan Raman

WACV 2024, Waikoloa, Hawaii, USAFeatured in WACV DailyBest of WACV 2024

A two-stage method that learns deep visual representations from EEG features. It generalizes across datasets and outperforms GAN-based baselines on EEG-based image reconstruction and classification.

arXivCode

Single Image LDR to HDR Conversion using Conditional Diffusion

Dwip Dalal, Gautam Vashishtha, Prajwal Singh, Shanmuganathan Raman

ICIP 2023, Kuala Lumpur, Malaysia

We cast single-image LDR to HDR conversion as image-to-image translation and solve it with a conditional diffusion model using classifier-free guidance. A CNN autoencoder improves the latent representation of the LDR image used for conditioning.

arXiv

EEG2IMAGE: Image Reconstruction from EEG Brain Signals

Prajwal Singh, Pankaj Pandey, Krishna Miyapuram, Shanmuganathan Raman

ICASSP 2023, Rhodes Island, Greece

Contrastive learning extracts features from EEG signals, and a conditional GAN synthesizes images from them. A modified GAN loss lets the model produce 128x128 images from a small training set.

arXivCode

APEX-Net: Automatic Plot Extractor Network

Aalok Gangopadhyay, Prajwal Singh, Shanmuganathan Raman

NCC 2022, Mumbai, India

A deep learning framework with new loss functions that extracts data from plot images and cuts down manual effort. We also release APEX-1M, a large dataset of plot images paired with their raw data.

PaperarXiv

LS-HDIB: A Large Scale Handwritten Document Image Binarization Dataset

Kaustubh Sadekar, Ashish Tiwari, Prajwal Singh, Shanmuganathan Raman

ICPR 2022, Montreal, Canada

A dataset of one million challenging handwritten document images with accurate segmentation ground truth, built for handwritten document image binarization.

arXivProject page

Search Me Knot, Render Me Knot: Embedding Search and Differentiable Rendering of Knots in 3D

Aalok Gangopadhyay, Paras Gupta, Tarun Sharma, Prajwal Singh, Shanmuganathan Raman

Computer Graphics Forum 2024

A fully differentiable framework for designing 3D tubular knots that resemble a target image from chosen viewpoints. It pairs an invertible neural network with physically constrained optimization, validated by experiments, ablations and a 3D-printed object.

arXivCode

Hand Shadow Art: A Differentiable Rendering Perspective

Aalok Gangopadhyay, Prajwal Singh, Ashish Tiwari, Shanmuganathan Raman

Pacific Graphics 2023, Poster, Daejeon, KoreaBest Poster Award

Differentiable rendering deforms hand models so that the shadows they cast resemble a target image.

Paper

TreeGCN-ED: Encoding Point Cloud using a Tree-Structured Graph Network

Prajwal Singh, Ashish Tiwari, Kaustubh Sadekar, Shanmuganathan Raman

Pacific Graphics 2023, Poster, Daejeon, Korea

A tree-structured autoencoder that builds robust point cloud embeddings from hierarchical information using graph convolution. Experiments and a t-SNE map show it separates object classes well.

PaperarXivCode

DILIE: Deep Internal Learning for Image Enhancement

Indra Deep Mastan, Shanmuganathan Raman, Prajwal Singh

WACV 2022, VAQ Workshop, Waikoloa, Hawaii, USA

Image enhancement with deep internal learning. DILIE improves content and style features while preserving the semantics of the enhanced image.

Paper

Beyond Reconstruction: What EEG-to-Video Decoding Actually Recovers

Prajwal Singh, Anupam Sharma, Pankaj Pandey, Krishna Miyapuram, Shanmuganathan Raman

Preprint 2025

We study the dynamic visual information encoded in EEG and show that it can be reconstructed with temporally conditioned generative models. The work also asks what EEG-to-video decoding actually recovers of continuous visual experience.

alphaXivarXiv