SIREN & Diffusion Experiments

Overview

Two parallel experiments: reproducing SIREN implicit neural representations at high resolution and building diffusion-model sketches (DCGAN → Progressive GAN → diffusion) to study latent trajectories and editability. These exercises were part of coursework and self-directed research to gain intuition about continuous signal representations and modern generative priors. :contentReference[oaicite:11]11

SIREN (Implicit Representations)

• Implemented sinusoidal representation networks (SIREN) to encode images at 1024×1024 resolution.
• Explored positional encodings and activation scaling to stabilize training.
• Observed excellent reconstruction fidelity for textures and high-frequency detail compared to baseline MLPs.

Diffusion & GAN Experiments

• Re-implemented course assignments: DCGAN, Progressive GAN, and a basic diffusion pipeline on curated datasets.
• Performed latent interpolations and denoising trajectories to understand mode coverage and sample diversity.
• Benchmarked training behavior on an A100 and profiled memory/perf trade-offs to inform future model decisions.

Outcomes

• The experiments improved practical know-how around model instabilities, training curricula for generative models, and how implicit representations can be integrated into larger graphics/vision pipelines (e.g., for texture compression or controllable editing in Happenstance / To Wilt).