Few-step image generation and the real cost of video
How distillation turns 50-step diffusion into 1 to 8 steps, what guidance distillation removes, why attention cost explodes with resolution and clip length (with a cost model you can run), what sequence parallelism buys, and how to serve video as a job queue. Numbers come from an illustrative transformer denoiser, not a measured product.
The diffusion lesson explained the cost model for image generation: time is steps times evaluations per step times the time for one evaluation. This lesson goes one level deeper into the two things that change that product the most: how few steps you can get away with, and why video is so much more expensive than images.
All numbers come from a small calculation for an illustrative transformer denoiser (40 layers, width 3072, about 4.5B parameters in the blocks, running at an effective 400 TFLOP/s). Real models differ in architecture and speed, so use these as a way to see the trends and then measure your own system.
Pro content
Keep reading with Inference Academy Pro
Unlock the rest of this lesson, including 1 animated diagram, plus every advanced lesson, hands-on lab, the larger playground models and completion certificates.