Engineering / research
Execute generation samplers efficiently
Compare sampler discretization, guidance, caching, batching, and quality measurements without treating fewer function evaluations as a universal quality guarantee.
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
Execute generation samplers efficiently
Compare sampler discretization, guidance, caching, batching, and quality measurements without treating fewer function evaluations as a universal quality guarantee.
- 01
Reverse SDE and probability-flow ODE
Separate stochastic reverse dynamics from deterministic probability transport, with explicit time direction, score assumptions, and numerical limits.
30 min - 02
Sampler discretization and local error
Separate one-step numerical defects, accumulated trajectory error, model error, and perceptual quality when comparing sampling budgets.
30 min - 03
Noise schedules and step placement
Distinguish a trained noising family from the inference grid and preserve prediction conversions, coordinates, and endpoint conventions.
30 min - 04
Classifier-free guidance cost and bias
Account for conditional branches and altered sampling fields without equating a fused invocation or stronger guidance with free or exact improvement.
30 min - 05
Multistep samplers and state caching
Preserve trajectory-specific solver history and distinguish exact reusable conditioning from approximate denoiser activation reuse.
30 min - 06
Latent decoding and memory peaks
Budget the full pipeline’s live allocations and decoder behavior rather than inferring serving capacity from compressed latent bytes.
30 min - 07
Generation request batching and cancellation
Schedule compatible active trajectories while preserving per-request solver state, random streams, and GPU buffer lifetimes.
30 min - 08
Resolution bucketing and tiling
Bound shape-padding work and tile-memory tradeoffs while respecting global operations, receptive fields, and image-boundary semantics.
30 min - 09
Quality and speed frontier measurement
Compare complete serving configurations using paired inputs, scoped quality metrics, uncertainty, and measured latency rather than step-count slogans.
30 min - 10
Seeded generation reproducibility boundaries
Control request-owned random streams and complete generation manifests without promising identical pixels across arbitrary numerical environments.
30 min