Generative AI - DDPM Diffusion Model
Diffusion, written from scratch — no black boxes.
Framework
PyTorch
Method
DDPM
Credential
NVIDIA DLI
Metric profile
01 / Signal
1000 T
Diffusion steps
full DDPM schedule
02 / Signal
Stable
Sample quality
cosine beta schedule
03 / Signal
Single GPU
Training footprint
NVIDIA DLI workflow
Overview
A full Denoising Diffusion Probabilistic Model built in PyTorch from first principles: forward noising schedule, U-Net noise predictor, and the reverse sampling loop, all implemented by hand.
Completed as part of NVIDIA Deep Learning Institute training on generative AI with diffusion models.
What makes it work
Forward process
A parameterised beta schedule progressively corrupts images to isotropic noise, with closed-form sampling at any timestep.
U-Net denoiser
A convolutional U-Net with timestep embeddings and residual blocks learns to predict the noise added at each step.
Reverse sampling
Iterative denoising reconstructs images from pure noise, making the generative trajectory fully inspectable.
Architecture
- Beta schedule + closed-form forward diffusion
- U-Net with sinusoidal timestep embeddings
- MSE noise-prediction training objective
- Ancestral sampling loop for generation
Stack
- PyTorch
- DDPM
- CNN
Denoising Diffusion Probabilistic Model implemented from scratch in PyTorch (NVIDIA DLI training).
Interested?