See the families as one map
Connect likelihood models, latent variables, flows, energy methods, and diffusion through their probability and sampling interfaces.
Generative Models · from first principlesModel possibility.
Model possibility.
Then sample it.
Start with the familiar act of drawing different outcomes, then compare how model families create, steer, and evaluate them. Formal probability and implementation remain optional depth after the sampling idea is clear.
30 lessons · about 13 hours
THE GENERATIVE LOOP
data + condition→randomness + model
generated samplebounded evidenceillustrative flow · not a measurement
Turn randomness into controlled creation.
Generative models do more than produce striking samples: they offer different ways to represent probability, compress structure, transform noise, and steer what is possible. Learning the families side by side reveals which tool fits which creative or scientific problem.
Control what gets generated
Work with conditions, guidance, inverse problems, multimodal signals, and the trade-offs between fidelity, diversity, and speed.
Move beyond attractive samples
Design matched experiments, preserve seeds and budgets, diagnose missing modes or sampler bias, and make claims the evidence earns.
The finish lineComplete the course with a principled generative toolkit and an original experiment you can defend.
From a probability table to a controlled generative research system.
Six build territories expose a different generative interface, then connect implementation, sampling, evaluation, safety, and research evidence.
NEURAL FIELD GUIDE / COURSE ARC01 · 06
Probability to Generation
Build and evaluate an inspectable distribution workbench.
You unlockTested sampler/evaluator package
Autoregressive & Latent Models
Build and diagnose autoregressive and variational generators.
You unlockVAE/VQ-style latent comparison dossier
Flows & Energy Models
Implement exact-density transforms and energy-based sampling.
You unlockMatched-target flow/EBM comparison
Score & Diffusion Models
Build a diffusion model and trace its complete sampling path.
You unlockSmall image diffusion model with sampling trace
Conditional Generation
Control generation while measuring fidelity, diversity, and failure.
You unlockFidelity/diversity/memorization audit
Generative Research
Run a matched-budget, reproducible model-family study.
You unlockReproduced baseline plus controlled original ablation
Inside the course
One course. 6 connected frontiers.
Lessons 01–30 build selected core families—autoregressive, latent-variable, flow, energy-based, and diffusion systems—then culminate in a matched original study. GANs and several specialized families are outside this course's current scope.
01
Model a distributionProbability to Generation
Build and evaluate an inspectable distribution workbench. Each lesson adds one tested component to the territory build.
The payoffBuild and evaluate an inspectable distribution workbench.
02
Infer hidden causesAutoregressive & Latent Models
Build and diagnose autoregressive and variational generators. Each lesson adds one tested component to the territory build.
The payoffBuild and diagnose autoregressive and variational generators.
03
Transform and sampleFlows & Energy Models
Implement exact-density transforms and energy-based sampling. Each lesson adds one tested component to the territory build.
The payoffImplement exact-density transforms and energy-based sampling.
04
Denoise into samplesScore & Diffusion Models
Build a diffusion model and trace its complete sampling path. Each lesson adds one tested component to the territory build.
The payoffBuild a diffusion model and trace its complete sampling path.
05
Steer the processConditional Generation
Control generation while measuring fidelity, diversity, and failure. Each lesson adds one tested component to the territory build.
The payoffControl generation while measuring fidelity, diversity, and failure.
06
Compare with evidenceGenerative Research
Run a matched-budget, reproducible model-family study. Each lesson adds one tested component to the territory build.
The payoffRun a matched-budget, reproducible model-family study.
30connected lessons
30hands-on labs
30code notebooks
No black boxes. Build intuition, see the mechanism, then make the real engineering trade-offs.
Built from the field, not the hype
Primary work behind the course
Ready when you are
Complete the course with a principled generative toolkit and an original experiment you can defend.
30 connected lessons, hands-on labs, and a complete end-to-end build.