Research Insights
Visual explorations and animations from my research — interactive visualizations, model behaviors, and experimental results that bring the work to life.
Multi-Scale Steering Field Navigation
This animation visualizes a multi-scale steering field over a greenhouse crop canopy sequence. The colored arrows represent directional flow vectors at multiple spatial scales — capturing both fine-grained local patterns and broader structural navigation cues across the plant canopy. We construct a per-pixel entropy map from a single forward pass of the pepper-LoRA model on tomato 🍅 images, smooth it at multiple spatial scales (σ ∈ {20, 60, 150} px) via Gaussian filtering, and sum the resulting spatial gradients — normalised per scale — to produce a dense vector field that points toward regions of high model uncertainty.
Diffusion Model for Synthetic Tomato Data Generation
This animation shows a diffusion model generating synthetic tomato 🍅 crop images from pure noise. Starting from random Gaussian noise, the model progressively denoises across multiple timesteps to produce realistic greenhouse tomato imagery. The 2×2 grid displays four samples being generated simultaneously, illustrating how the learned reverse diffusion process gradually recovers fine-grained plant structures, fruit shapes, and canopy textures — enabling data augmentation for agricultural computer vision tasks where labeled data is scarce.