Impact Highlights
Built a full DDPM training and reverse-sampling loop from scratch.
Guided generation with text-conditioned CLIP losses.
Learned disentangled latent factors and autoregressive priors.
Mission Brief
Deep Generative Modeling
A practical generative modeling lab covering DDPM diffusion, CLIP-guided sampling, beta-VAE disentanglement, and PixelCNN.
PyTorchDDPMCLIPVAEGANPixelCNN
Build Notes
Implemented a DDPM with UNet denoiser, cosine scheduler, and reverse diffusion sampling.
Built CLIP-guided diffusion for text-driven MNIST generation.
Implemented beta-VAE for disentangled latent learning and PixelCNN with type A/B masked convolutions.
Image Direction
Recommended Concept
A visual narrative of noise collapsing into structured images through layered denoising waves.
Text-to-Image Prompt
Abstract diffusion process visualization, particles converging into geometric forms, deep black backdrop, cyan and magenta glows, subtle film grain, futuristic and elegant, no text
Fallback asset in use: /illustrations/generative-lab.svg