Published backend result
fal's H3 Max launch benchmark reports that a 5-second video at 768p renders in under 3 seconds, with typical inference around 2.5 seconds. This is backend model inference, not a promise that every completed file reaches a ZipReel browser in three seconds.
The API documentation defines its reported timings.inference field as GPU backend denoising time. Prompt expansion, queue time, image upload, webhook delivery, database updates and the viewer's network sit outside that number.
ZipReel end-to-end measurement
For ZipReel, the useful timer begins when the generation request is accepted and ends when the completed video is visible in the user's library. A reproducible report must retain both clocks instead of presenting backend inference as total wait time.
- Run at least 30 requests per configuration, split across different times of day.
- Keep model, duration, resolution and prompt-expansion mode fixed.
- Record request acceptance, provider timing, completion webhook and library-visible timestamp.
- Publish sample size, P50, P90, failure rate and test window.
- Separate text-to-video from image-to-video because upload time changes the path.
Current ZipReel result status
No verified P50 or P90 is published yet. The six original showcase generations predate this measurement plan and do not have preserved end-to-end timestamps. ZipReel will not infer or manufacture those values. The first monthly report will appear here after a complete sample is collected.
- Backend reference
- <3 seconds for 5s 768p, published by fal
- ZipReel P50
- Measurement pending
- ZipReel P90
- Measurement pending
- Failure rate
- Measurement pending