VAEs, GANs, normalizing flows, score-based models, diffusion, flow matching, latent diffusion, ControlNet, consistency models, diffusion-DPO/RLHF, video, 3D, audio. 19 lessons covering the full generative-AI stack — from theory to FLUX/Sora/SDXL Turbo production.
Discriminative vs generative. The model-family taxonomy (autoregressive, VAE, GAN, flow, diffusion) and the metrics that compare them.
VAE, β-VAE, hierarchical VAEs. Latent-space exploration, ELBO, the reparameterization trick.
Codebook latents, vector quantization, EMA. Foundation of DALL-E's image tokenizer and EnCodec audio.
DCGAN, WGAN, StyleGAN. The adversarial game, mode collapse, and what GANs still do best.
RealNVP, Glow, coupling layers. Exact-likelihood generative models — the bridge to flow matching.
Score matching, Langevin dynamics, the SDE/ODE framework. Yang Song's theoretical backbone of diffusion.
Forward + reverse process, L_simple, classifier-free guidance, DDIM and modern samplers (DPM-Solver++, Euler).
Conditional flow matching, rectified flow. The training paradigm behind FLUX, SD3, and Stable Video.
Compress to latent space first, diffuse there. SD 1/2, SDXL, SD3, FLUX — the production stack.
ControlNet, IP-Adapter, LoRA-Dreambooth, prompt2prompt, inpainting. Production controllable T2I.
Consistency Models, LCM, SDXL Turbo. Diffusion distillation from 50 steps to 1-4 steps.
Sora, Veo, Runway Gen-3, Hunyuan, Kling. Diffusion transformers (DiT), latent video VAEs, temporal attention.
NeRF, 3D Gaussian Splatting, DreamFusion (SDS), Shap-E. Generating geometry from text and images.
TTS (VALL-E/ElevenLabs/OpenAI), music gen (Suno/Udio/MusicGen), audio codecs (EnCodec/SoundStream).
FID, KID, CLIP-score, DINO-score, prompt-following benchmarks, human eval — measuring generative quality.
Stable Signature, C2PA, deepfake detection, Glaze/Nightshade, copyright lawsuits, EU AI Act provisions.
DALL-E 3, Midjourney, Imagen, FLUX, SDXL, Sora 2, Veo 3, Suno — model survey and selection rubric.
Capstone — wire prompt → SDXL/FLUX → ControlNet → upscale → safety → watermark → serve via FastAPI.
DALL-E 3, Midjourney v6, FLUX.1 [pro], SDXL Turbo — every frontier image model in 2024-2025 ships a base diffusion checkpoint and then spends weeks of GPU time bending it…
12 interactive labs — hands-on exercises for this track
Two AIs compete: one creates fake art, the other detects fakes — watch them get better together
Start with pure noise and gradually reveal an image — this is how Stable Diffusion works
Step through each noise level of the diffusion process — from crystal-clear image to pure static and back
Drag a point through the latent space and watch the decoder produce different outputs — this is how generative models work.
Token by token, probability by probability — this is exactly how ChatGPT generates text.
Drag through a 2D latent space with four semantic anchors. Compare lerp vs slerp and see how latent arithmetic edits attributes.
Compare linear, cosine, and sigmoid diffusion schedules side-by-side. See β(t), ᾱ(t), and visual signal destruction at any timestep.
Drag the CFG scale from 1 to 15 and watch prompt adherence, sample diversity, and over-saturation change. Find the sweet spot.
Watch 50 particles flow from a Gaussian noise cloud to a two-moons target by integrating a learned velocity field.
Step through reverse diffusion one denoising step at a time. Watch structure emerge from pure noise, see the predicted noise the model outputs.
Target = 8 Gaussians in a ring. Watch generator samples cluster at 1-2 modes (collapse). Stack 4 fixes to see coverage climb toward 8/8.
Drag a latent vector → snaps to nearest of 8 codebook entries (Voronoi cells shown). VAE blurs, VQ-VAE stays sharp.
475 questions across 19 modules — check how well you understood this track.