Note: Mandatory to fill both RSVP and fill in the Google https://forms.gle/hgBmNo5NHNekVYvm7 no last minute exceptions will be possible without completing both these steps.
Build a Diffusion Model from Scratch on 6 GB VRAM | Image → Video → Audio → Text → Omni
How does a model learn to create an image, generate a video, produce text, or work across multiple modalities—starting from nothing but noise?
In this hands-on workshop, we will build an intuitive understanding of diffusion models from first principles, starting with the mathematics and fundamental idea of adding noise and learning to reverse that process.
We will then trace the evolution of diffusion architectures, from the early diffusion models to modern approaches used across image, video, text, audio/multimodal, and Omni-generation systems.
We’ll explore concepts such as:
- Forward diffusion and the reverse denoising process
- Noise schedules, timesteps, and denoising objectives
- DDPM and DDIM
- U-Net–based diffusion architectures
- Latent Diffusion Models
- Diffusion Transformers (DiT)
- Flow Matching and modern diffusion/flow-based architectures
- How diffusion evolved from image generation into video and multimodal generation
- How modern Omni models combine multiple modalities within generative architectures
The workshop will conclude with a live implementation of a diffusion model from scratch—building the forward-noising process, neural network, training loop, and reverse-generation process ourselves.
The entire hands-on implementation will be designed to run on a consumer GPU with just 6 GB of VRAM, making it possible to understand diffusion without requiring expensive hardware.
From noise → mathematics → architecture → a working generative model.
