Sampler Acceleration (DDIM/DPM-Solver/LCM)

Original DDPM sampling needs 1000 steps; acceleration falls into three families:

1. DDIM (Denoising Diffusion Implicit Models, deterministic implicit sampling): rewrites the forward chain as non-Markovian so sampling becomes a deterministic trajectory of the probability-flow ODE, while the training objective stays identical to DDPM (no retraining) — 10–50 steps suffice:

$\sigma_t = 0$ yields deterministic DDIM (Euler discretization of the probability-flow ODE), enabling latent-space interpolation and image reconstruction; $\sigma_t > 0$ gives stochastic sampling (DDPM as a special case). 2. High-order ODE solvers: DPM-Solver / DPM++ use 2nd/3rd-order exponential integrators on the probability-flow ODE, reaching near-1000-step quality in 10–25 steps. 3. Distillation: Progressive Distillation halves the step count per training round; Latent Consistency Model (LCM) applies consistency models to the latent space, learning to map any noisy sample to the start of the solution trajectory in 4 steps (even 1–2); LCM-LoRA works as a plug-in without retraining; SDXL-Turbo/Lightning use adversarial distillation for 1-step generation.

Tradeoff: fewer steps → lower latency but slightly worse FID/detail; e.g. 4-step LCM is usually a bit below 20-step DPM++, but roughly 10× faster.

interview question — “how do you cut diffusion from 1000 steps to real-time?”; production deployment, interactive generation, video generation and mobile latency optimization.

1000-step sampling is unusable for real-time interaction, video and mobile; DDIM accelerates an order of magnitude without retraining, high-order solvers reach 10–25 steps with nearly no quality loss, and distillation methods push single-image latency to tens of milliseconds — bringing diffusion into real-time applications.

Write the DDIM sampling formula; why is σ_t=0 a deterministic ODE sampler?

Do DDIM and DDPM share the same training objective; why does DDIM need no retraining?

How do LCM/consistency models achieve 4-step generation; what does the consistency loss constrain?

Difference between DPM-Solver/DPM++ and DDIM; why do high-order solvers need fewer steps for better quality?

Steps vs FID/detail tradeoff; when to use 4-step LCM vs 20-step DPM++?