Course Resources
This page contains curated resources to support your learning throughout the course. Resources are organized into monographs, tutorials, and research papers. This list will be updated throughout the class.
📚 Books
2025
Comprehensive monograph covering diffusion models, flow matching, and transport-based generative modeling from first principles.
🎓 Courses
2023
Stanford course on generative models including VAEs, GANs, EBMs, normalizing flows, diffusion models, and autoregressive models.
2025
CMU course on generative models including LLMs, GANs, and diffusion models.
2025
CMU course on generative models including LLMs, VAEs, and diffusion models.
2025
MIT class on diffusion and flow matching from a flow-based theoretical perspective.
2025
CMU course that focuses on probabilistic modeling (including some deep generative models from a more theoretical perspective).
2024
Stanford course that focuses on probabilistic modeling.
📝 Tutorials
2021
Introduction to score-based generative models and their connection to diffusion models.
2021
Comprehensive introduction to diffusion models with clear explanations and intuitive visualizations.
2022
Unifies VAEs, hierarchical VAEs, and diffusion models under a single framework.
2024
Comprehensive guide to flow matching with code examples and applications.
2024
Written by the first author of SEDD (ICML 2024 Best Paper). Explains the motivation for discrete diffusion, the concrete score, score entropy loss, and sampling from first principles.
2024
Blog post and video walkthrough by the MDLM authors. Connects absorbing-state diffusion to masked language modeling (BERT); the video tutorial is especially helpful.
2022
Detailed walkthrough of D3PM with the author's own clarifications. Great for understanding the transition matrix formulation and why the network predicts x₀ instead of x_{t-1}.
📄 Key Papers
2015
The original paper introducing diffusion probabilistic models for generative modeling.
2020
Landmark paper that made diffusion models practical and effective for high-quality image generation.
2021
Unifies score-based models and diffusion models using SDEs, enabling new sampling methods.
2022
Introduces flow matching as a simulation-free approach to training continuous normalizing flows.
2022
Introduces rectified flow to learn straight trajectories between distributions for fast sampling.
2022
General framework for building normalizing flows via stochastic interpolation between distributions.
2021
OG fast sampling algorithm paper for diffusion model.
🚀 Advanced Papers
2022
Systematic analysis of design choices in diffusion models with improved sampling.
2022
Fast high-order ODE solver for diffusion models.
2022
Improved solver with guided sampling support.
2021
Improved DDPM with learned variance and cosine noise schedule.
2022
Introduces v-prediction parameterization and progressive distillation.
2022
Guidance method that combines conditional and unconditional models.
2021
Introduces latent diffusion models, the foundation of Stable Diffusion.
2021
Introduces classifier guidance for conditional generation.
2021
Training-free image editing by hijacking the diffusion process with user inputs.
2022
Time traveling technique for training-free inpainting.
2022
Using off-the-shelf discriminative models for diffusion guidance.
2022
Discovered that diffusion manifolds are layered bubble shells.
2023
ICCV 2023 - Talks about which range of the time steps matter more for controllable generation.
2023
Guidance method that preserves the data manifold during sampling, making the guided sampling faster and better.
2017
Discovered that we can access the tangent space of the data manifold via the Jacobians of the decoder in an autoencoder.
2017
Introduces vector quantized variational autoencoders for learning discrete latent representations.
2020
Combines VQ-VAE with adversarial training and autoregressive transformers for high-resolution image synthesis.
2021
Learns visual representations from natural language supervision, widely used as the text encoder in text-to-image models.
2019
A text-to-text transformer used as the text encoder in many text-to-image models including Stable Diffusion.
2022
Replaces the U-Net backbone with a transformer, establishing the foundation for modern diffusion transformer architectures.
2024
Combines rectified flow matching with a multimodal DiT architecture (MMDiT).
2024
Rectified flow transformer for text-to-image generation; no official paper, open-weight model released August 2024.
2025
32B flow matching transformer that couples a vision language model with a rectified flow transformer; supports multi-reference generation and up to 4MP output.
2025
Efficient 6B-parameter model built on a single-stream DiT architecture; trained in only 314K H800 GPU hours.
2025
Native multimodal MoE model with 80B+ total parameters (13B activated per token); unifies multimodal understanding and generation within an autoregressive framework.
2024
Trains a single transformer over mixed-modality sequences by combining next-token prediction loss with diffusion loss.
2025
Google's native multimodal image generation model built into Gemini.
2025
GPT-4o's native autoregressive image generation capability.
2023
New family of models that enable single-step generation while maintaining sample quality.
2023
Generalizes consistency models to learn the full trajectory of the probability flow ODE.
2025
Provides a theoretical framework for building consistency models via flow map learning and self-distillation.
2025
Scales continuous-time flow map distillation for high-quality few-step generation.
2025
Joint distillation framework that enables both fast sampling and likelihood evaluation in flow-based models.
2021
The D3PM paper; introduces a family of discrete diffusion models with structured transition matrices including absorbing (mask), uniform, and embedding-based diffusion.
2022
Extends discrete diffusion to continuous time using Continuous Time Markov Chains (CTMCs), enabling more principled training and sampling.
2023
The SEDD paper (ICML 2024 Best Paper); introduces score entropy as a training objective for discrete diffusion by estimating ratios of the data distribution, analogous to score matching in continuous diffusion.
2024
Proposes MDLM, a simple masked diffusion language model with an efficient training objective and absorbing-state noise schedule that matches or outperforms autoregressive models on language benchmarks.
2024
Unifies and simplifies masked diffusion models, showing that a simple masked diffusion objective generalizes prior work and yields strong performance on text generation.
2025
A masked diffusion language model trained from scratch at scale that matches LLaMA3 8B in instruction following, demonstrating the viability of discrete diffusion for LLMs.
2024
Extends flow matching to discrete state-spaces using CTMCs; demonstrates applications in protein co-design.
2024
Proposes a flow matching framework for discrete data that enables flexible noise schedules and competitive language modeling performance.
2025
Introduces edit flows, which define flows on discrete state-spaces via structured edit operations (insert, delete, substitute), enabling more expressive discrete generative models.
2025
Builds on edit flows to enable unified mixed-modal (text and image) generation and editing within a single model.
2025
Introduces block diffusion, which interpolates between autoregressive and masked diffusion LMs by denoising blocks of tokens, enabling flexible context length and KV-cache compatibility.
Compute Resources
Access to GPU compute resources is essential for training your generative models. Below are the available platforms with tutorials and guides to help you get started.
Modal
AvailableServerless cloud computing platform for running GPU workloads
AWS (Amazon Web Services)
Coming SoonCloud computing platform with EC2 GPU instances
SLURM Clusters
AvailableTutorials for navigating slurm-based clusters (like the ones we use at CMU)
Additional Resources
- Discord: Communication server for discussions, questions and collaboration
- Office Hours: In-person and virtual communication directly with the instructor, check the home page and the schedule page for the time and locations.
Note: This list will be updated throughout the course. Additional readings specific to each lecture can be found on the schedule page.