Video Providers

OpenAI's video generation model. Strong at cinematic motion, human subjects, and complex scene compositions. Supports both image-to-video (animating a still) and text-to-video generation. Offers the widest duration range with 12-second clips.

Remix support — Sora can create alternate interpretations of a scene via its remix capability, available as a scene recovery option.

Google's latest video generation model. Excellent at natural scenes, landscapes, atmospheric effects, and smooth camera movement. The 6-second mid-length bucket provides flexibility that other providers lack.

Google's Gemini-native video model. Renders any clip length from 3 to 10 seconds in 16:9 or 9:16 at 720p or 1080p, with a generated audio track, so scene windows rarely need capping or slow-down. Supports image-to-video and text-to-video.

Runway's latest generation model. Great for stylized and abstract visuals with consistent artistic aesthetics. Works best with strong reference images and specific art direction.

Strong at dynamic camera movement and 3D-aware scene understanding. Produces smooth, physically plausible motion with good depth perception.

A generative model with a distinctive visual style. Operates primarily as text-to-video, bringing its own interpretation to prompts.

ByteDance's model, hosted on Replicate. Any clip length from 4 to 15 seconds at 480p, 720p or 1080p in 16:9, 9:16, 1:1 and more. On a performance film, every vocal shot is sent with its own window of your audio track as a reference, so the performer mouths the words at generation time instead of needing a correction pass. The scene still is kept as the character and framing reference for those shots.

Kuaishou's model, hosted on Replicate. Any clip length from 3 to 15 seconds; 720p renders in standard mode and 1080p in pro mode. Starts from the scene still when there is one and from the prompt alone otherwise.

Alibaba's model, hosted on Replicate. Image-to-video only, 5 or 10 seconds at 720p or 1080p, and it can follow a 3-30 second audio track, so vocal shots are generated with the scene's audio window. Every Wan clip is asked without on-screen text, captions and watermarks up front.

Alibaba's Wan 3.0, hosted on Higgsfield, in its reference-to-video mode: every clip is generated from your song. Each shot is sent with its storyboard still and its own slice of your track, starting exactly at the shot's start, so a sung shot is mouthed to the real words and a dance shot moves on the real beat. In our tests it reproduced the song in time with the picture more closely than any other provider, so pick it for performance films with on-camera singing. Any clip length from 2 to 30 seconds at 480p, 720p or 1080p.

ByteDance's Seedance 2.0, hosted on Higgsfield, in the same song-driven mode as Wan 3.0: still plus the shot's slice of your track. Its timing to the song is looser than Wan 3.0's, so use it when Wan 3.0 refuses a shot. Any clip length from 4 to 15 seconds, in 16:9, 9:16, 1:1, 4:3, 3:4 or 21:9.