Wirewalk Instituteof AI
mod-350 Expert Waitlist

Diffusion Models

For ML engineers who want to understand diffusion and flow-matching generators mathematically, train and fine-tune them, control them, and sample from them fast.

The curriculum is being prepared and reviewed. Enrollment is not open yet.

About this course

Diffusion and flow-matching models generate today's leading images and video and are increasingly used for audio, 3D, proteins and materials. This course builds the theory you need to work on them rather than only prompt them. You derive the DDPM forward and reverse processes, connect noise prediction to score matching and stochastic differential equations, and then see how flow matching and rectified flow, the formulation used by Stable Diffusion 3 and the FLUX family, recast generation as learning a velocity field.

The practical half uses Hugging Face Diffusers. You train a small model from scratch, examine the latent-diffusion stack (VAE, text encoders, U-Net or diffusion transformer), and fine-tune large open-weight generators with LoRA and DreamBooth-style personalization. You then control generation with classifier-free guidance, ControlNet-style conditioning, image-prompt adapters and editing, and make it fast with modern samplers, distillation to few-step models, caching and TensorRT. The course ends with evaluation (FID, CLIP score and human preference) and with provenance, licensing and safety for generated media. Model names and licenses are current as of October 2026; each open checkpoint carries its own license.

What you will be able to do

  • Derive the DDPM forward and reverse processes and explain the link between noise prediction and the score
  • Explain flow matching and rectified flow and why they enable straighter, faster sampling paths
  • Describe the latent diffusion stack, including VAE, text encoders and U-Net versus diffusion-transformer backbones
  • Train a small diffusion model and fine-tune a large open-weight generator with LoRA using Diffusers
  • Apply classifier-free guidance, spatial control adapters and image editing techniques to steer outputs
  • Reduce sampling cost with ODE solvers, few-step distillation and inference optimization
  • Evaluate generators with FID, CLIP-based and human-preference metrics and apply provenance and licensing controls

Syllabus

6 modules · 20 lessons · about 20 hours

Module 1: The mathematics of diffusion

Noising, denoising and the score function.
3 lessons
  1. The forward process and noise schedules
  2. DDPM: training a network to predict noise
  3. Score matching, SDEs and the probability-flow ODE

Module 2: Flow matching and rectified flow

The formulation behind current state-of-the-art generators.
3 lessons
  1. Flow matching: learning a velocity field
  2. Rectified flow and straight paths
  3. Prediction targets compared: epsilon, v and velocity

Module 3: Architectures for generation

How production text-to-image and video models are assembled.
4 lessons
  1. Latent diffusion: autoencoders and the latent space
  2. U-Nets versus diffusion transformers and MMDiT blocks
  3. Text conditioning: CLIP, T5 and language-model encoders
  4. Extending to video: temporal attention and spatiotemporal latents

Module 4: Training and fine-tuning with Diffusers

Training small models and adapting large ones.
3 lessons
  1. Training a small diffusion model from scratch with Diffusers
  2. LoRA fine-tuning of open-weight generators such as FLUX and Stable Diffusion
  3. DreamBooth-style personalization and dataset captioning

Module 5: Controlling generation

Steering outputs with guidance, structure and reference images.
3 lessons
  1. Classifier-free guidance and guidance scale trade-offs
  2. ControlNet-style spatial control and image-prompt adapters
  3. Inpainting, image editing and instruction-based editing models

Module 6: Fast sampling, evaluation and responsible release

Making generation cheap and measuring and governing what comes out.
4 lessons
  1. Samplers: DDIM, DPM-Solver and Euler for flow models
  2. Few-step generation: consistency models and adversarial distillation
  3. Evaluating generators: FID, CLIP score and human preference
  4. Provenance, C2PA content credentials, licensing and safety filters

Exam and certificate

The exam unlocks once you have completed every lesson. It draws 25 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 60 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.

Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.

The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by Hugging Face (Diffusers, PEFT), Black Forest Labs (FLUX), Stability AI (Stable Diffusion), PyTorch, NVIDIA TensorRT, or ComfyUI. Product and company names are trademarks of their respective owners.

More in Model Development

All 16 courses in this track
mod-201 Intermediate Available

Deep Learning with PyTorch

A hands-on course in modern PyTorch: tensors and autograd, nn.Module design, data loading, a training loop you understand line by line, GPU and mixed-precision training, torch.compile, and exporting a model for deployment.

14 h · 19 lessons See price at checkout
mod-202 Intermediate Available

Transformers from the Ground Up

Build a decoder-only transformer piece by piece: tokenization, embeddings, scaled dot-product attention, multi-head and grouped-query attention, positional schemes such as RoPE, and the training and sampling loop, then map it onto real open-weight model architectures.

16 h · 19 lessons See price at checkout
mod-203 Intermediate Waitlist

Classical Machine Learning in Production

Gradient-boosted trees and linear models still win most tabular problems. This course covers leakage-proof pipelines, honest validation, tuning, calibration and explainability, then packaging, serving and monitoring for drift.

14 h · 19 lessons Join the waitlist