The Generative AI path
How machines make images, sound and text, in the order that works. Start at Level 0 if the field is new to you. Otherwise begin at Level 1, or pick one of the two routes.
You can also open the Generative AI track.
Two routes through the lessons
Practitioner
Fast path
You want to use and build with image generators. The shortest order that gets you there.
- 1.Generative Models Landscape & Evaluation
- 2.Generative Adversarial Networks
- 3.Diffusion Models (DDPM, DDIM, CFG)
- 4.Text Conditioning: How a Prompt Steers the Picture
- 5.Latent Diffusion & Stable Diffusion
- 6.Conditional & Controllable Generation
- 7.Safety, Copyright & Watermarking
- 8.Few-Step Generation: Consistency Models and Distillation
- 9.Modern GenAI Landscape
- 10.Building a Diffusion App
Researcher
Theory path
You want the maths behind the models. The fast path plus every derivation lesson.
- 1.Generative Models Landscape & Evaluation
- 2.Autoencoders
- 3.Variational Autoencoders
- 4.Generative Adversarial Networks
- 5.Diffusion Models (DDPM, DDIM, CFG)
- 6.Latent Diffusion & Stable Diffusion
- 7.Conditional & Controllable Generation
- 8.Safety, Copyright & Watermarking
- 9.Few-Step Generation: Consistency Models and Distillation
- 10.VQ-VAE & Discrete Latents
- 11.Diffusion Transformers (DiT)
- 12.Autoregressive and Discrete Generation
- 13.Modern GenAI Landscape
- 14.Building a Diffusion App
- 15.Score-Based Generative Models
- 16.Normalizing Flows
- 17.Flow Matching & Beyond
- 18.Diffusion Conventions and Samplers
- 19.Diffusion DPO & RLHF: Fine-Tuning Image Models on Preferences
- 20.Generative Evaluation in Depth
- 21.Generative Expert Lab
Level 0
Start here
No calculus and no code. Pictures and sliders that explain what generative AI is, how a chatbot writes, how a picture comes out of static, and why AI makes things up.
- Generative AI: Start HereWhat it means for a computer to make something new instead of sorting something old, the four everyday ways it is done, and which lesson to read next.15 min
- Sampling and Latent SpaceHow a model can make something new by reaching into a bag of examples, and how a short list of numbers can act as a map of everything it knows.15 min
- How a Chatbot WritesA chatbot writes a reply one small piece at a time, each piece a weighted guess, and this lesson shows the guess, the dice and the dials.15 min
- From Noise to PictureHow a computer makes a picture by starting from random static and cleaning it up a little at a time, and what the guidance dial does.15 min
- Why AI Makes Things Up (and What to Do)Why an AI tool can sound sure and still be wrong, where bias comes from, and the simple habits that keep you safe.15 min
Level 1
Foundations
The map of generative model families, then autoencoders, the VAE and GANs.
- Generative Models Landscape & EvaluationDiscriminative vs generative. The model-family taxonomy (autoregressive, VAE, GAN, flow, diffusion) and the metrics that compare them.25 min
- AutoencodersSqueeze a picture through a narrow bottleneck and rebuild it. The denoising autoencoder, the PCA link, and why a plain autoencoder cannot make new pictures.15 min
- Variational AutoencodersA cloud of possibilities instead of one point: the ELBO, the reparameterization trick, smooth walks between latents and beta-VAE.25 min
- Generative Adversarial NetworksDCGAN, WGAN, StyleGAN. The adversarial game, mode collapse, and what GANs still do best.30 min
Level 2
Core
How pictures are made: diffusion, how a prompt steers it, working in latent space, and how to judge the results.
- Diffusion Models (DDPM, DDIM, CFG)Forward + reverse process, L_simple, classifier-free guidance, DDIM and modern samplers (DPM-Solver++, Euler).60 min
- Text Conditioning: How a Prompt Steers the PictureHow a typed prompt becomes numbers, how a text encoder such as CLIP or T5 builds them, how cross-attention lets each part of the picture read them, and where the known weaknesses come from.35 min
- Latent Diffusion & Stable DiffusionCompress to latent space first, diffuse there. SD 1/2, SDXL, SD3, FLUX — the production stack.35 min
- Evaluating Generative Models: The BasicsWhat FID, precision and recall, CLIP score and human judgement each measure, and how to spot a model that makes only a few kinds of output.20 min
Level 3
Advanced
Control, ethics, fast sampling, discrete latents, video, audio, 3D, diffusion transformers, autoregressive image making, today's landscape and a capstone app.
- Conditional & Controllable GenerationControlNet, IP-Adapter, LoRA-Dreambooth, prompt2prompt, inpainting. Production controllable T2I.40 min
- Safety, Copyright & WatermarkingStable Signature, C2PA, deepfake detection, Glaze/Nightshade, copyright lawsuits, EU AI Act provisions.25 min
- Few-Step Generation: Consistency Models and DistillationWhy diffusion is slow, one map of the ways to make it fast, the consistency idea in pictures, a runnable toy, and an honest guide to the quality and diversity you give up.30 min
- VQ-VAE & Discrete LatentsCodebook latents, vector quantization, EMA. Foundation of DALL-E's image tokenizer and EnCodec audio.35 min
- Video GenerationHow image diffusion grows a time axis. Diffusion transformers (DiT), spacetime patches, latent video autoencoders, temporal attention.30 min
- Audio Generative AITTS (VALL-E/ElevenLabs/OpenAI), music gen (Suno/Udio/MusicGen), audio codecs (EnCodec/SoundStream).25 min
- 3D GenerationNeRF, 3D Gaussian Splatting, DreamFusion (SDS), Shap-E. Generating geometry from text and images.25 min
- Diffusion Transformers (DiT)Why image generators swapped the U-Net for a transformer, and how that network works: patches, positions, time and class conditioning, joint attention with text, and what it costs.30 min
- Autoregressive and Discrete GenerationHow a picture or a sound becomes a sentence of codes, and three ways to write that sentence: one code at a time, one scale at a time, or the surest cells first.30 min
- Modern GenAI LandscapeA dated survey of image, video, audio and 3D model families, plus a six-axis rubric and break-even maths for choosing between them.20 min
- Building a Diffusion AppCapstone — wire prompt → SDXL/FLUX → ControlNet → upscale → safety → watermark → serve via FastAPI.20 min
Level 4
Expert
Score-based models, normalizing flows, flow matching, a translation table for the conventions and samplers, preference tuning for diffusion, deeper evaluation and a from-scratch lab.
- Score-Based Generative ModelsScore matching, Langevin dynamics, the SDE/ODE framework. Yang Song's theoretical backbone of diffusion.45 min
- Normalizing FlowsRealNVP, Glow, coupling layers. Exact-likelihood generative models — the bridge to flow matching.35 min
- Flow Matching & BeyondConditional flow matching, rectified flow. The training paradigm behind FLUX.1 and Stable Diffusion 3.30 min
- Diffusion Conventions and SamplersA translation table for the time, noise and prediction conventions of DDPM, EDM, score-based models, v-prediction and flow matching, why schedules, loss weights and samplers are three separate dials, and how Euler, Heun, DDIM and DPM-Solver step along the same path.40 min
- Diffusion DPO & RLHF: Fine-Tuning Image Models on PreferencesFine-tune an image diffusion model on human preferences: reward models, DDPO, Diffusion-DPO, and how to avoid reward hacking.35 min
- Generative Evaluation in DepthFID in depth, precision and recall computed with nearest neighbours, KID, DINOv2-FID and FVD, the 2025-26 metrics, and when to use each.20 min
- Generative Expert LabNine from-scratch numpy exercises with automatic checking: forward noising, the DDPM and DDIM steps, guidance, velocity prediction, score matching, flow matching, token unmasking and FID, plus the traps and a capstone checklist.55 min