Video Lab Works On Your Footage — YouTube Rips, Aleph 2 Edits, Magnific 4K, Act-Two

August 9, 2026

Video Lab is no longer generate-only: point it at real footage. Rip a YouTube video straight into your session, ingest just the minutes you care about, then prompt-edit it with Runway Aleph 2, upscale it to 4K with Magnific, drive a character with an acting take via Act-Two, or hand it to Seedance as a continuation reference.

What you can do

• Download YouTube into the session — youtube action download saves the video (480p/720p/1080p ceiling, default 720p) or just the audio track (audio_only, m4a) into the session's in/ folder, ingest-ready for Video Editor and Video Lab. Sources up to 20 minutes by default.
• Ingest only a window — pass time_range: {start, end} to video_editor ingest and only that segment is extracted and analysed. No more paying full Whisper + vision analysis on a 40-minute source to work on 30 seconds of it.
• edit_video (Runway Aleph 2) — prompt-based editing of existing footage (≤30s): "make it golden hour and remove the parked cars". Output follows the input's duration and framing.
• upscale_video (Magnific creative upscaler) — upscale a clip to 720p–4K (upscale_resolution, auto = 2K).
• character_performance (Runway Act-Two) — motion-transfer an acting take onto a character still: your input_image performs the input_video's delivery, with optional body-gesture transfer and expression intensity 1–5.
• Seedance reference_videos — up to 10 session clips as motion/style/continuation references on create: pass a previously generated clip as [Video1] to continue its scene. (Seedance also takes up to 30 reference_images.)
• Remix without pushback — footage you upload or have Alfrada OS fetch is yours to work with. The agent transforms it as directed and keeps the original soundtrack when asked; provider moderation, not a preemptive refusal, is the gate.

Where this shows up

You want a highlight reel from a conference talk. You used to download it yourself, trim it locally, and upload the pieces. Now: "download this YouTube talk, ingest 12:00–14:30, and cut the three strongest moments into a vertical clip."

Your drone clip is beautiful but 720p and shot at noon. edit_video re-lights it to golden hour, then upscale_video delivers it at 4K.

You have one hero portrait of a character and a phone video of yourself acting the lines. Act-Two makes the character deliver your performance.

Try it

• "Download youtube video dQw4w9WgXcQ at 720p, ingest just 0:45–1:30, and give me a storyboard."
• "Take clip.mp4 and make it golden hour, remove the parked cars — Runway Aleph 2."
• "Upscale out/final.mp4 to 4K with Magnific."
• "Animate character.png with the performance in take3.mp4 — expressive, keep body gestures."
• "Continue the chase from [Video1]: the cyclist crests the hill. Seedance, 12 seconds."

Heads up

• YouTube downloads and ToS. Rips are capped at 20 minutes. For summarising, transcript remains faster and cheaper than download.
• The Runway footage operations (edit_video, upscale_video, character_performance) accept inputs up to 30 seconds and ~16 MB encoded — extract a clip first for anything bigger.
• On Seedance, frame anchors and references are mutually exclusive upstream: reference_images/reference_videos apply to text-to-video calls, not create_with_image.
• If a provider's moderation rejects a submission, Alfrada OS reports it rather than working around it.