video-gen

Distilled video model: drafts vs finals

2026-10-271× RTX 5090, power-capped 403 W

wall-clock (draft pass)

38 s (median)

Compare — full 20-step final: 320 s (median)

Setup

I use a distilled video model (internal name stays private) in two modes on the same RTX 5090 at 403 W:

  • Draft: 8-step, 107 frames at 768×448
  • Final: full 20-step, 124 frames at 1344×768

What I measured

Median wall-clock from production logs for each mode separately.

Numbers

Mode Median wall-clock Conditions
Draft 38 s 8-step · 107 frames · 768×448
Final 320 s 20-step · 124 frames · 1344×768

Draft is how I iterate; final is what I ship when the motion looks right. The ~8× gap is not a clean speedup number — resolution, steps, and frame count all move at once.

What’s next

Hold frame count fixed for one A/B so the draft savings are honest to compare.

Method
Median over logged draft jobs (8-step) vs separate final renders (20-step); single GPU.
Caveats
Draft and final differ in steps, resolution, and frame count — not an apples-to-apples speedup multiplier.

← All benchmarks