The Research Idea Center.
Hand-curated research ideas across CS, Health and Climate domains. Each idea contains a gap statement, suggested methodology, literature review, and summary brief — organised by Domain → Subdomain → Topic → Research Area.
Test idea
Iot
Test idea
A Diffusion Transformer (DiT) is a generative AI architecture that replaces the traditional, locally-focused U-Net backbone in diffusion models with a scalable Transformer network. By treating visual data as sequences of tokens, DiTs capture global context instantly, allowing for higher-quality, scalable image and video generation.How Diffusion Works (The Basics)Traditional diffusion models operate by progressively adding noise to data (the forward pass) and then training a neural network to predict and remove that noise step-by-step (the reverse pass). This gradually transforms random static into coherent, realistic images or videos.The U-Net to Transformer ShiftThe Old Way (U-Net): Early diffusion models like Stable Diffusion 1.5 used U-Nets, which rely on convolutions. While good at local details, they struggle with global structure and long-range coherence across an entire image or video.The DiT Way (Transformer): DiTs operate by "patchifying" a noisy image or video into discrete tokens (like how a Vision Transformer reads images) and feeding them into a standard Transformer block.Why Diffusion Transformers ExcelGlobal Context: Through self-attention, the model can look at the entire image or video sequence simultaneously, maintaining perfect spatial and temporal continuity.Predictable Scaling: Unlike U-Nets, Transformers follow strict scaling laws. As you add more parameters and computing power, the quality predictably improves. Multimodal Mastery: DiTs can seamlessly integrate text prompts, audio waveforms, and video frames as a unified stream of tokens. Real-World Applications. Because of their efficiency and scalability, Diffusion Transformers have become the engine powering state-of-the-art generative AI systems, ranging from high-definition image generation to long-form, 4K video synthesis (such as systems like OpenAI's Sora)
Low-resource Nepali instruction-tuning corpora curation
Build a reproducible pipeline that converts Nepali web text + glossed parallel data into high-quality instruction-tuning pairs.
Soiling loss models for high-altitude solar plants in South Asia
Quantify monthly soiling losses at 3 Himalayan PV sites and fit a regression model using AOD + local weather covariates.
Stigma-aware chatbot interventions for adolescent depression in Nepal
Mixed-methods evaluation of a Nepali-language conversational agent that adapts disclosure prompts based on caste, gender and family stigma signals.
Battery-less LoRa sensor nodes for Himalayan crop monitoring
Design and field-test an energy-harvesting LoRa node powered solely by indoor PV + thermoelectric scavenging — targeted at year-round crop telemetry in Mustang and Manang.