Engineering / research
Train denoising and score models
Connect forward noise, target parameterization, weighting, latent representations, and conditioning to controlled diffusion training experiments.
Your route through engineering
Follow the path or begin with a question. Every lesson is open.
Your place, at your pace. Progress is optional and saved only in this browser when available.
Read / explore / reason
Inside this module
Module 01 research
Train denoising and score models
Connect forward noise, target parameterization, weighting, latent representations, and conditioning to controlled diffusion training experiments.
- 01
Forward noising: state the normalization before quoting an SNR
Derive variance-preserving Gaussian corruption, cumulative schedule coefficients, and the domain of signal-to-noise measurements.
27 min - 02
Noise, score, and velocity: transform the target and its error together
Derive target conversions under a VP convention and distinguish conditional training scores, marginal score estimators, and parameterization-dependent loss scales.
28 min - 03
Denoising weights: transform the objective, not just its target name
Relate clean, noise, and velocity squared errors and derive a stated Min-SNR weighting without claiming universal convergence gains.
28 min - 04
Timestep sampling: preserve the intended objective under a new proposal
Derive importance corrections, support conditions, and variance considerations for nonuniform diffusion-timestep sampling.
28 min - 05
Latent diffusion: measure the representation bottleneck before denoising it
Account for latent tensor size, reconstruction distortion, normalization, and encoder/decoder state without equating dimension reduction with an information-theoretic coding rate.
27 min - 06
Conditioning dropout: train a null branch before choosing a guidance scale
Separate random conditioning removal during training from classifier-free extrapolation during inference, with an explicit guidance convention.
28 min - 07
Conditioning paths: identify where time and content enter the denoiser
Trace timestep embeddings, cross-attention, spatial inputs, and adaptive normalization through architecture-specific diffusion backbones.
27 min - 08
EMA state: define which tensors move and which clock advances
Specify parameter averaging, buffer policy, update cadence, warmup state, and checkpoint behavior for diffusion-model evaluation.
27 min - 09
Instability diagnosis: localize the first broken quantity
Separate data, target, weighting, numerical, optimizer, and sampling failures through fixed-batch probes and controlled interventions.
28 min - 10
Diffusion evaluation: freeze the data and sampling protocol before comparing models
Control source-group splits, preprocessing, held-out denoising, sampling settings, feature metrics, and paired evaluations.
29 min