Runway Gen-3 Alpha Turbo

À propos de Runway

Fondée 2018 · New York City, USA

Runway was founded in 2018 in New York City by Cristobal Valenzuela (CEO), Anastasis Germanidis (CTO) and Alejandro Matamala-Ortiz (CDO), three NYU ITP graduates who set out to build creative tools for the next generation of media production. Runway co-authored the seminal Stable Diffusion paper with CompVis and LMU Munich (Rombach et al., 2022) and shipped one of the first commercial text-to-video products, Gen-1 (2023), followed by Gen-2 (2023), Gen-3 Alpha (June 2024), Gen-3 Alpha Turbo (July 2024) and Gen-4 (2025). The company has raised over $300M from investors including Google, NVIDIA, Salesforce Ventures and General Catalyst, reaching a $3B+ valuation in 2024. Runway also runs the AI Film Festival and partners with Hollywood studios for on-set previs and post-production tooling.

Architecture

Latent video diffusion model (DiT-style transformer denoiser)

Gen-3 Alpha Turbo is a distilled, lower-step variant of Runway's Gen-3 Alpha latent-video-diffusion model, designed to deliver Gen-3-level visual fidelity at roughly 7x faster inference and lower cost. It uses a transformer-based denoiser operating on a compressed spatio-temporal latent representation, conditioned on text embeddings from a multilingual text encoder and optional image embeddings. Turbo is image-to-video focused: a user-supplied first frame is mandatory, and the model generates 5- or 10-second clips at 720p / 24 fps. Runway describes Gen-3 as trained on a 'new infrastructure built for large-scale multimodal training' and on a curated mix of internal video, licensed footage and partner data. The Turbo variant is produced via consistency-/distillation-style training that compresses the diffusion sampler into far fewer denoising steps.

Paramètres

Undisclosed

Capacités

• Image-to-video generation in 5s or 10s clips at 720p / 24 fps
• Approximately 7x faster than Gen-3 Alpha base at similar fidelity
• Strong camera-motion vocabulary (handheld, dolly, crane, orbit)
• Lip-sync workflow when combined with Runway's Act-One and Lip Sync features
• Extend video up to ~40 seconds via successive 10s extensions
• Available via Runway web app, mobile app and Gen-3 API
• Per-clip pricing well below Gen-3 Alpha base
• Tight integration with Runway editor, green-screen, motion-brush
• Best for: production-line iteration, ad creative, social shorts, animatics.

Entraînement et licence

Curated mix of internal video, licensed footage and partner data on Runway's large-scale multimodal training infrastructure; exact corpus size undisclosed.

Licence: Proprietary commercial licence via Runway Terms of Service; generated content usable commercially on paid plans.

Tests de sécurité: Multi-stage content moderation including NSFW, CSAM, celebrity-likeness and trademark filters; public model card and ongoing red-team programme.