Parallel encoding · Built on NVIDIA CUDA

Transcode once.
Play everywhere.

Upload a video and get crisp 1080p, 720p, 480p and 360p versions in about a minute, with live progress, instant previews and one-click downloads.

  • No credit card
  • Up to 30 MB per video
  • Auto-deleted after 30 days
  • Private by default
vidflow-transcoder.vercel.app

4× parallel

resolutions encoded at once

CUDA

NVDEC → scale → NVENC pipeline

< 1 min

for a short clip, end to end

H.264 + AAC

plays on every device

Features

Everything you need to ship video

Simple on the surface, a proper pipeline underneath.

Every resolution, one upload

1080p, 720p, 480p and 360p are produced together, keeping your video's aspect ratio and audio.

Live progress

Follow each video from queued to ready in real time. Close the tab and come back; it keeps going.

Preview and download

Play any version in the browser with a resolution switcher, grab one file, or download all four as a ZIP.

Private by default

Files go straight to private cloud storage over short-lived signed links. Only you can play or download them.

CUDA-accelerated pipeline

Decode once, scale on the GPU, encode all four resolutions in parallel with NVENC. Each job runs in its own isolated container, orchestrated on AWS.

You stay in control

Delete any video whenever you like. Everything is removed automatically after a set number of days.

Built on NVIDIA CUDA

Faster by design.
Parallel at every step.

VidFlow's pipeline is built around NVIDIA CUDA: decode once, scale on the GPU, and encode all four resolutions in parallel with NVENC, orchestrated end to end on AWS.

Right now: all four versions are already encoded in parallel in a single pass. The GPU path is built and tested; it switches on as GPU capacity is enabled.
  • Decode once, encode four in parallelYour video is decoded a single time and every resolution is produced at the same moment, not one after another.
  • Frames stay on the GPUNVDEC decodes, CUDA scales and NVENC encodes without copying frames back to system memory. That zero-copy path is what makes GPU transcoding fast.
  • Safe fallbackIf a file can't be handled on the GPU (an unusual codec, say), the job transparently runs on CPU instead of failing.
  • Elastic by designGPU workers start on demand, run only while there is work, and shut down when idle, so capacity follows demand.
DecodeNVDEC · once

One pass over your source video

ScaleCUDA · in GPU memory

Four sizes from the same decoded frames

1080pNVENC
720pNVENC
480pNVENC
360pNVENC

Four encoders, side by side

Orchestrated end to end on AWSFrom the upload to the finished files, every step is a managed AWS service, defined as code.
  • Amazon S3storage
  • AWS Lambdaupload trigger
  • Amazon ECSFargate + GPU workers
  • API Gatewaysecure API
  • CloudWatchmonitoring
  • Terraforminfrastructure as code

How it works

From upload to ready in four steps

1

Upload

Drop a video. It goes straight to secure storage.

2

Queue and launch

A scheduler starts a dedicated encoder for it.

3

Transcode

Decoded once, encoded to four resolutions together.

4

Deliver

Preview, stream or download every version.

FAQ

Questions, answered

A short clip is usually ready in under a minute: about 15 seconds to start the encoder, then roughly the length of the video to encode it. A 2-minute video takes a few minutes. You can close the tab; progress is kept and the page picks up where it left off.

Typically under a minute for a short clip

Drop one file.
Get four ready to play.

Sign in with Google, upload a video, and watch it turn into 1080p, 720p, 480p and 360p versions you can preview and download.

Try it with your video Private by default
  • No credit card
  • Google sign-in only
  • Delete anytime
Drop a video here
holiday-trip.mp428.4 MB
1080p4.9 MB
720p3.5 MB
480p1.0 MB
360p0.6 MB