AI VIDEO MODEL
中文
The model directory

Compare AI video models.

Use documented inputs, output requirements and access to build your shortlist. This directory does not rank visual quality or claim a hands-on benchmark.

Showing 5 models

Published capabilities · checked October 25, 2026
ModelSingle passResolution & outputDirectionAccess caveat
Seedance 2.5 ↗
ByteDance Seed
Up to 30sVerify in your chosen platformUp to 30 images, 10 videos and 10 audio referencesPlatform rollout; announcement says API coming soon
Official source ↗
Kling 4.0 ↗
Kuaishou
Up to 30s4K and 1080p described as upcoming in launch releaseUp to 15 combined assets and 10 keyframesLaunched October 15; verify API separately
Official source ↗
Gemini Omni Flash ↗
Google
Up to 10s360p / 720p; 1080p and 4K upscaledText, image and video; conversational editsStable API model: gemini-omni-1.1-flash
Official source ↗
Runway Gen-4.5 ↗
Runway
Up to 10s720p base outputText-to-video and image-to-videoWeb; Standard plan and higher
Official source ↗
Luma Ray3.2 ↗
Luma
Unconfirmed1080p; HDR and EXR workflow documentedUp to 16 keyframes; Modify Video V2Luma App and API
Official source ↗

A specification comparison, not a measured quality ranking. Native generation, upscaling, footage editing and extensions are separate operations.

Download sourced specifications (CSV) ↗ · Read model data (JSON) ↗

Go deeper on each model.

ByteDance Seed

Seedance 2.5

Reference-led storytelling

A candidate for longer scenes and multimodal reference workflows. The launch announcement describes timestamp-level audio and video editing.

  • 30s maximum single-pass generation
  • Verify in your chosen platform
Kuaishou

Kling 4.0

Keyframe direction

A candidate when a shot needs several planned visual beats. Kuaishou describes a 30-second single pass and reference-driven camera and subject control.

  • 30s maximum single-pass generation
  • 4K and 1080p described as upcoming in launch release
Google

Gemini Omni Flash

Conversational editing

A candidate for applications that refine a short video through follow-up instructions. The stable model card lists 3–10-second output.

  • 10s maximum single-pass generation
  • 360p / 720p; 1080p and 4K upscaled
Runway

Runway Gen-4.5

Shot-by-shot creation

A candidate for short directed shots with text or a starting image. The help page specifies 2–10-second generations and a web credit rate.

  • 10s maximum single-pass generation
  • 720p base output
Luma

Luma Ray3.2

Footage transformation

A candidate for editing existing footage and workflows that need keyframes, HDR or compositing exports. Modify Video V2 accepts footage up to 20 seconds.

  • Generation duration not confirmed
  • 1080p; HDR and EXR workflow documented
Start here

The right model starts with your shot.

Compare controls, access and output requirements. Then run a small pilot with an acceptance rubric you can actually use.

Build your test plan ↗

Two useful workflow comparisons.

How to read this comparison

The duration column means a single generation, not an assembled sequence or a footage-editing limit. The resolution column preserves whether a source calls output upscaled or planned. API access is listed only where the cited documentation establishes it; a web launch is not treated as an API launch.

No field is a quality score. Choose a representative task and compare accepted output under the requirements that matter to your project. A short product reveal, a character scene and an existing-footage transformation may produce very different shortlists. Keep costs in their published units until you can establish a valid conversion for your own account.

Sources and verification

Reviewed October 25, 2026. Provider documentation supports specifications; it is not evidence of an independent performance test.