Research update · 19 September 2026
A 5090 FastH3 run can take 55 seconds for 15 seconds of video.
Platform benchmark says: Sogni reports that one RTX 5090 (32 GB) produced a 15.08-second FastH3 Preview v1 image-to-video MP4 in 55 seconds at 480×640 and 126 seconds at 768×1024. The smaller result is useful for a faster clip-review or prequeue loop. It is still 3.65 times the output duration, so it is not a live-response result.
One test, two useful ratios
The source fixes the output length at 362 frames, or 15.08 seconds at 24 fps. Dividing each reported wall time by that duration gives a more honest planning number.
| Source-stated condition | Complete-MP4 time | Output duration | Wall seconds per output second | What it supports |
|---|---|---|---|---|
| FastH3 Preview v1 I2V · RTX 5090 32 GB · 480×640 · 4 steps | 55 s | 15.08 s | About 3.65 | Faster review or a preloaded scene library |
| Same named route · 768×1024 | 126 s | 15.08 s | About 8.36 | Offline image-anchored clip generation |
The source times from worker job start to a finished MP4. It excludes queue and model-load time. This is not time to first frame, input-to-visible response, or a count of ready clips ahead of a viewer.
The warm machine changes the story
Sogni says its first 480p FastH3 canary after a worker restart took 160 seconds, while the same configuration in its hot matrix took 54 seconds. That does not prove a general cold-start penalty, but it is enough to make warm residency part of the route’s contract. A benchmark without its start state can sell a wait that a creator never sees.
Do not transfer this result to the new Comfy package
The benchmark uses the older four-step FastH3 Preview v1 condition. FastH3 8-Step V2 is a separate checkpoint and its current ComfyUI first/last-frame scope remains disputed between FastVideo and ComfyUI documentation. A 5090 number for Preview v1 is not evidence for V2 speed, quality, or task coverage.
What the source does and does not settle
Platform benchmark says: the source names the worker, task, source still, prompt, seed, steps, audio setting, output shape, and timing boundary. That makes it more useful than an unqualified “5× faster” post.
Unknown: pool reproduction, dollar cost, worker queue time, retries, sustained uptime, time to first frame, delivery, viewer control, quality parity, and lip sync. Sogni operates the tested service, so its own timing and visual assessment are not independent validation.
Smallest useful next move
If you need rapid local-style clip selection, use this as a comparison lead and repeat the same clip cold and warm on your actual route. Record model and task, GPU/VRAM and system RAM, resolution and frames, model residency, queue delay, complete-media time, retries, and the quality you will accept. If you need a viewer to cause the next visible change, this benchmark does not clear that bar.