Schedule

Continuous Diffusion

Instructor: Chieh-Hsin (Jesse) Lai

This section develops the theoretical and empirical foundations of diffusion and flow-map models for fast sampling on continuous data.

Learning Objectives:

1. Train a continuous diffusion model 

2. Distill pretrained diffusion models into fast samplers, including  DMD

3. Train flow-map models, including  Consistency Model, Consistency Trajectory Model, MeanFlow

Pre-requisites: 

Familiarity with classic generative models (VAEs, energy-based models, and normalizing flows), and PyTorch.

Contents:

1. Continuous Diffusion Overview from Three Origins

   1. Variational-based: VAE ➞ DDPM 

   2. Score-based: Energy-based model ➞ Score SDE

   3. Flow-based: Normalizing flows ➞ Flow matching & Rectified flow

2. Distribution-based Distillation from Pretrained Diffusion Models into Fast Generators

Live Demo 1: Quick Play Around Distribution-based Distillation

3. Diffusion-Motivated Flow-Map Models as Fast Generators 

   1. Consistency Model (CM)

   2. Consistency Trajectory Model (CTM)

   3. MeanFlow (MF)

   4. Efficient Flow-Map Training with Consistency Mid-Training

Live Demo 2: Training a Flow Map Model

4. Audio-Visual Generation & Protection of Creation ーー Instructor: Yuki Mitsufuji 

This part briefly covers AI content creation and protection, from diffusion memorization and attribution to audio-visual models such as  MMAudio, highlighting both creativity and the need for safeguards. 

Reference:

• The Principles of Diffusion Models  by Chieh-Hsin Lai, Yang Song, Dongjun Kim, Yuki Mitsufuji, Stefano Ermon.

Discrete Diffusion

Instructor: Subham Sekhar Sahoo 

This section builds the theoretical + empirical foundation for diffusion when the data lives in discrete space.

Learning Objectives:

1. Train a discrete diffusion model.

2. Perform self-distillation for faster generation.

3. Implement samplers for

1. Faster generation 

2. Inference-time scaling

Pre-requisites: 

Familiarity with neural networks and PyTorch. 

Contents:

1. Discrete Diffusion Overview

   1. Uniform-state diffusion: Forward / reverse process.

   2. Training + Implementation.

   3. Samplers.

2. Live Demo 1: Image generation

   1. Training a discrete diffusion model for image generation.

   2. Implement samplers:

      1. Ancestral-sampler.

      2. Psi-sampler [2] for inference-time scaling.

3. Self-Distillation to accelerate sampling

   1. The discrete-space and continuous-space diffusion duality [1].

   2. Use the duality to design Discrete Consistency Distillation [1].

   3. Implementation.

4. Live Demo 2: Discrete Consistency Distillation

   1. Distill a pre-trained discrete diffusion model for faster inference via Discrete Consistency Distillation.

   2. Implement Greedy-Tail sampler [1] for few-step generation.

Reference:

1. Subham Sekhar Sahoo, Justin Deschenaux, Aaron Gokaslan, Guanghan Wang, Justin Chiu, Volodymyr Kuleshov, "The Diffusion Duality", ICML 2025.

2. Justin Deschenaux, Caglar Gulcehre, Subham Sekhar Sahoo, "The Diffusion Duality, Chapter II: Psi-Samplers and Efficient Curriculum",  ICLR 2026.