RTX 5090 Tested: Cutting H3 from 20 Steps to 4 — Where the Time Actually Goes
Updated 2026-09 · Data from local ComfyUI run logs and public benchmarks
This test machine: RTX 5090 32GB + 64GB DDR5-6000, running the official minimum
setup (pruned INT8 main model 19.5GB + NVFP4 text encoder 14.6GB + two VAEs), launch
flags --fast --use-sage-attention, and the LightX2V 8-step acceleration LoRA.
Every number below comes from real logs, not estimates.
Start with the time budget
For the same 1344×768, 73-frame (~3-second) clip, here is the breakdown at 8 sampling steps:
Stage
Time
Share
Text encoder load + encode
~11 s
14%
Sampling (8 steps × 5.45 s/step)
43.6 s
54%
VAE decode + save
~17.1 s
21%
Other (scheduling, caching, etc.)
~8.6 s
11%
End-to-end
80.3 s
100%
Two takeaways: in short clips, sampling is only half the story — loading and decoding
are another big chunk; the longer the clip, the higher the sampling share (token count grows
with the attention quadratic term), and that is when acceleration really pays off in sampling.
Step compression: the real payoff of 20 → 8 → 4
Public benchmarks on the same machine, 1280×704, ~5-second clips (sampling segment only,
excluding loading):
Config
Steps
Sampling time
vs. 20 steps
Official baseline
20
~270 s (extrapolated)
1.0×
+ LightX2V Turbo 8-step
8
125 s
2.2×
+ built-in INT8 attention (Kitchen)
8
79 s
3.4×
+ FastH3 4-step distillation
4
~60–70 s (inferred from 1.87×)
~4×
One counterintuitive thing to note: per-step time is not constant. The 20-step
profile costs about 13.5 s per step, but the 8-step profile actually rises to about
15.6 s — fewer steps does not reduce per-step overhead (LoRA patching and scheduler
switching are still there), yet total time still drops substantially. So "steps ÷ 4 =
time ÷ 4" only holds as a rough estimate.
How to stack the three acceleration layers correctly
Step distillation (pick only one): LightX2V 8-step (smallest quality loss),
4-step (~3.4×), larryvrh Turbo 8-step, FastH3 4-step (~4×). Stacking two distillation
LoRAs will break each other's effects;
Attention backend (stackable, big gains): SageAttention or ComfyUI 0.32+'s
built-in Kitchen attention, speeding up the sampling segment another 1.4–2×. On a
512-step run, this layer is essentially free;
Caching (use with caution): EasyCache / Spectrum saves another 15–40%, but it
works by skipping computation, so spot-check the frames; when batch-generating seeds
from the same prompt, the lossless CLIPCached conditioning cache pays off more.
Quality trade-offs
4 steps for drafts: get a rough sense of composition and motion in the shortest time;
6–8 steps for final renders: 4 steps can occasionally produce over-sharpened
faces and plasticky details;
Sage's limits: completely safe for short clips; 50-series cards risk silent
noise on very long sequences (roughly 160k+ tokens, i.e., beyond ~26 s at 1344×768 or
~10 s at 1920×1088) — switch to the Kitchen backend for long videos.
How much time will your GPU need?
The numbers above only apply to the 5090. Sampling time on other GPUs depends mainly on
memory bandwidth and architecture — open the WhichH3 calculator, pick
your GPU, RAM, and page file settings, and it will estimate time and VRAM usage for every
"resolution × duration × steps" combination, compare the 20-step / 8-step / 4-step
acceleration options, and highlight the best setting for your machine.
Conclusion
In end-to-end time for short clips, half is not sampling — squeeze loading and
decoding overhead first (reuse models, conditioning cache);
Step distillation plus a good attention backend gives a 3.4–4× stacked gain on
the 5090;
4 steps for drafts and 8 steps for final renders is the most stable mix of quality and
efficiency;
For long videos, remember to check Sage's sequence-length limits.