You can now add images, title cards, and logos right in the editor. Come try it now →, plus plenty of other fixes and improvements this week. Send bug reports or feature requests to hello@vidclean.net

Under the hood

How VidClean Works: Infrastructure & AI Models

VidClean is a free suite of 32 video, audio, and image tools. This page documents exactly how it runs, the infrastructure, the AI models, and why it stays free, with specific, verifiable facts.

What AI models does VidClean use?

VidClean uses a hybrid speech-recognition engine for transcription and subtitles, NVIDIA Parakeet-TDT 0.6B v3 (int8) for the 25 supported European languages, and OpenAI Whisper (the "small" model, int8-quantized) for language detection and every other language, plus DeepFilterNet3 for background-noise removal, audio repair, and speech enhancement, and the rembg u2net model for image background removal. Every model is self-hosted and runs quantized on CPU, so there is no per-call API cost on those tools.

How is VidClean's infrastructure set up?

VidClean runs on Railway for about $50/month across six services: a web service (the site and upload API), an FFmpeg worker for video and audio processing, a DeepFilterNet worker (8 GB RAM) for noise removal, a transcription worker (4 GB RAM) running the hybrid Parakeet-TDT and Whisper engine, an image worker (2 GB RAM) running rembg for background removal.

What happens to my files?

Files are processed entirely server-side and automatically deleted within about 90 minutes. Uploads up to 2 GB are supported. Nothing is kept, shared, or used to train models.

Why is VidClean free?

Because the core AI models are self-hosted and quantized to run on CPU, there is no per-use API cost to pass on. The free tools have no per-use cost to the user and no monthly caps. Total infrastructure is about $50/month, paid by the founder.

How many tools and how much has VidClean processed?

VidClean offers 32 free video, audio, and image tools and has processed 2,000+ files to date, all free, with no account required and no watermark on any output.