Video generation models

Duration and resolution, side by side

Only the models whose vendor publishes a figure appear here. "Not established" means no page we could retrieve stated one.

All 62

• Adobe Firefly Video Model Adobe

  The Adobe Firefly Video Model generates video from a text prompt, optionally guided by an image used as the first or final frame. Adobe's Firefly Services API exposes it as model version vid

• Allegro Rhymes AI

  A small text-to-video model comprising a 2.8B VideoDiT and a 175M VideoVAE, released under Apache-2.0 in October 2024. It generates 88-frame clips at 720x1280 and 15 fps, roughly 6 seconds,

• Amazon Nova Reel 1.1 Amazon Web Services (Amazon Nova)

  Amazon Nova Reel generates video from text prompts or from text plus an image, and is available only through Amazon Bedrock. AWS documents output at 1280x720 and 24 frames per second, produc

• ByteDance Seedance 2.0 ByteDance

  Seedance 2.0 is ByteDance Seed's unified multimodal audio-video model: a single generation pass accepts text plus mixed references (up to 9 images, 3 video clips and 3 audio files, addressed

• CogVideoX1.5-5B (and 5B / 2B / I2V variants) Zhipu AI / THUDM (Z.ai)

  CogVideoX1.5-5B generates 1360x768 video at 16 fps in frame counts of 16N+1 up to 161 frames (about 10 seconds), with a matching I2V checkpoint. Earlier CogVideoX-5B and CogVideoX-2B produce

• FramePack lllyasviel (Lvmin Zhang)

  A next-frame-prediction network structure that compresses input context to a constant length, so generation cost does not grow with video length. This lets 13B-class video models produce rou

• Haiper 2.x (discontinued) Haiper AI

  Haiper's 2.x text-to-video and image-to-video models were a European alternative to Runway and Pika, built by a London team of Google DeepMind alumni. The product no longer operates independ

• Higgsfield Cinematic Studio 3.0 / Video 3.5 Higgsfield AI

  Cinematic Studio is Higgsfield's current first-party video pipeline, exposed through the product at higgsfield.ai/generate and through the CLI and API as cinematic_studio_3_0 and cinematic_s

• Higgsfield DoP (Lite / Standard / Turbo) Higgsfield AI

  DoP ('Director of Photography') is Higgsfield's own image-to-video model family, exposed on the API as higgsfield-ai/dop/standard and in the product as Lite, Standard and Turbo tiers. It ani

• HunyuanVideo (13B) Tencent Hunyuan

  Tencent's original open video foundation model, with over 13 billion parameters - the largest open-weights video model at its December 2024 release. Generates 129-frame clips at 540p to 720p

• HunyuanVideo 1.5 Tencent Hunyuan

  An 8.3B-parameter video model covering both text-to-video and image-to-video, released with 480p and 720p checkpoints plus a super-resolution stage that upscales to 1080p. Native generation

• HunyuanVideo-Avatar Tencent Hunyuan

  Audio-driven human animation built on HunyuanVideo. Given a character image and an audio track it generates emotion-controllable dialogue video, and supports multi-character scenes and non-p

• HunyuanVideo-I2V Tencent Hunyuan

  An image-to-video adaptation of HunyuanVideo that uses a token-replace technique to condition on a reference image, emphasizing first-frame consistency. Supports up to 720p and up to 129 fra

• Kling 2.5 Turbo Kling AI (Kuaishou Technology)

  Kling 2.5 Turbo (kling-v2-5-turbo) is the previous-generation value model, restricted to text and image input with fixed 5 or 10 second output at 720P or 1080P. It has no native audio and no

• Kling 2.6 Kling AI (Kuaishou Technology)

  Kling 2.6 (kling-v2-6) generates 3 to 10 second clips at 720P or 1080P from text, image or video input, with optional native audio and voice control at 1080P. It is the default model for Kli

• Kling 3.0 Kling AI (Kuaishou Technology)

  Kling 3.0 (API name kling-v3) is Kuaishou's flagship general video model, taking text, image or video input and producing 3 to 15 second clips at 720P, 1080P or native 4K. It can generate na

• Kling 3.0 Omni Kling AI (Kuaishou Technology)

  Kling 3.0 Omni (kling-v3-omni) is the multimodal member of the 3.0 family: it accepts text, images and reference video in one request and outputs 3 to 15 second clips at 720P, 1080P or 4K wi

• Kling 3.0 Turbo Kling AI (Kuaishou Technology)

  Kling 3.0 Turbo (kling-3.0-turbo) is the cheaper, faster member of the 3.0 family, covering text-to-video and image-to-video at 720P or 1080P for 3 to 15 seconds with native audio. It drops

• Kling O1 Kling AI (Kuaishou Technology)

  Kling O1 (kling-video-o1) is a unified video model that takes all-in-one multimodal instructions including reference video, and outputs 3 to 10 second clips at 720P or 1080P. Kuaishou positi

• LTX-2.5 Lightricks

  A 22B-parameter diffusion transformer that generates synchronized video and audio in a single model, using a Gemma-class 12B text encoder and separate video and audio VAEs. Shipped as a full

• LTX-Video 13B 0.9.8 (and distilled / 2B variants) Lightricks

  The first-generation LTX-Video family: a 13B DiT plus a distilled 13B and a 2B distilled model, all built for speed via multiscale rendering. It is one of the few open video models fast enou

• Luma Dream Machine / Ray2 Luma AI

  Luma's original consumer video product (Dream Machine) and the Ray2 model family behind it, with API model IDs ray-2 and ray-flash-2 and resolution options of 540p, 720p, 1080p and 4k. Luma'

• Luma Ray3.14 Luma AI

  Ray3 line model that preceded Ray3.2. Luma's field guide describes it as the default workhorse with 5-10 second text-to-video, 5 second image-to-video, draft through 1080p native output, HDR

• Luma Ray3.2 Luma AI

  Luma's current video model, released June 2026. Generates 5 or 10 second clips at up to 1080p from text or images, with six aspect ratios, native HDR in 16-bit colour, EXR export in ACES2065

• Meta Movie Gen (weights unreleased) Meta

  A research model family announced by Meta in October 2024 covering text-to-video, personalized video, instruction-based video editing and synchronized audio generation. Meta published a deta

• MiniMax H3 MiniMax (Hailuo AI)

  MiniMax H3 is a single multimodal model that accepts text, images, video and audio in one content array and returns 4 to 15 second video at 768P or 2K with native stereo sound. Three request

• MiniMax Hailuo 02 MiniMax (Hailuo AI)

  MiniMax-Hailuo-02 is the 2025-generation Hailuo model and the only MiniMax V1 model that supports first-and-last-frame generation as well as text-to-video and image-to-video. Resolutions are

• MiniMax Hailuo 2.3 (and 2.3-Fast) MiniMax (Hailuo AI)

  MiniMax-Hailuo-2.3 is the previous-generation Hailuo video model, still served on MiniMax's V1 text-to-video and image-to-video endpoints. Output is 6 or 10 seconds at 768P, or 6 seconds at

• Mochi 1 (preview) Genmo

  A 10B-parameter text-to-video diffusion model using an Asymmetric Diffusion Transformer (AsymmDiT) architecture, released as a research preview under Apache-2.0. At release it was the larges

• Moonvalley Marey (Realism v1.5) Moonvalley

  Marey is Moonvalley's foundational video model, marketed as the first commercially safe video model because it is trained only on licensed high-resolution footage with no scraped content or

• NVIDIA Cosmos 3 (Super / Nano / Edge) NVIDIA

  A family of omnimodal world foundation models aimed at physical AI - robotics, autonomous vehicles and simulation - rather than creative filmmaking. Cosmos 3 ships as Super (64B), Nano (16B)

• Open-Sora 2.0 HPC-AI Tech (Colossal-AI)

  An 11B-parameter open video generation model released with weights, inference code and training code under Apache-2.0. It generates at 256px and 768px in frame counts expressed as 4k+1 up to

• Pika 2.5 Pika

  Pika's current generation model, available as both text-to-video and image-to-video on the Pika API and in the Pika consumer app. Outputs 720p or 1080p; Pika's pricing page states maximum vi

• Pika 2.5 Keyframe Pika

  Keyframe-driven variant of Pika 2.5, exposed as an image-to-video model on the Pika API. It interpolates between supplied frames rather than generating freely from a prompt, and is the API s

• PixVerse C1 PixVerse

  C1 is PixVerse's first model aimed at film production, announced in April 2026. PixVerse states it generates up to 1080p and 15 seconds with native audio-visual output, and adds storyboard-t

• PixVerse V6 PixVerse

  V6 is PixVerse's general-purpose flagship, launched 30 March 2026. It covers text-to-video, image-to-video, transition, video extension and fusion (reference-to-video), and PixVerse highligh

• Pyramid Flow (miniFLUX / SD3) Pyramid Flow (academic collaboration)

  An autoregressive video generation method based on pyramidal flow matching, trained entirely on open-source datasets and accepted to ICLR 2025. The larger checkpoint produces 10-second video

• Runway Act-Two Runway

  Performance transfer model. You supply a character as an image or video plus a driving reference video of a person performing, and Act-Two maps that performance onto the character. Exposed v

• Runway Aleph 2.0 Runway

  In-context video editing model. It takes an existing video plus a text instruction and optional keyframe images, and re-renders the footage to add, remove, restyle or relight elements while

• Runway Gen-4 Aleph Runway

  First-generation Aleph video-to-video editing model, which introduced prompt-driven in-context editing of existing footage at Runway. Discontinued on 2026-07-30; Runway directs users to Alep

• Runway Gen-4 Turbo Runway

  Fast, low-cost image-to-video model. It requires a starting image rather than accepting a bare text prompt, and supports a wider set of output aspect ratios than Gen-4.5. Still listed as an

• Runway Gen-4.5 Runway

  Runway's current flagship generation model, producing video from text alone or from a starting image. Announced 2025-12-01 for the web app and added to the developer API on 2026-02-10. Runwa

• Seedance 1.0 Pro (closed weights) ByteDance

  ByteDance's commercial video generation model, available in Pro, Pro Fast and Lite tiers via API. It supports text-to-video and image-to-video at 480p, 720p and 1080p, 24 fps, with clip dura

• SkyReels-V2 (Diffusion Forcing / T2V / I2V) Skywork AI

  An open video model family whose Diffusion Forcing (DF) variants generate effectively unbounded-length video by autoregressive extension rather than producing a single fixed clip. Sizes are

• Sora 2 OpenAI

  Generates video with synchronised audio from a text prompt or from a reference image used as the opening frame. Accessible only through the OpenAI Videos API; the consumer Sora app and websi

• Sora 2 Pro OpenAI

  Higher-fidelity tier of Sora 2 aimed at production output, and the only Sora tier that can export 1080p. Same text-and-image inputs and synchronised audio output as sora-2, priced substantia

• Stable Video Diffusion (SVD-XT) Stability AI

  An image-to-video latent diffusion model that animates a single still image into a short clip at 576x1024. The XT variant is a fine-tune of the original 14-frame model and produces 25 frames

• Veo 3 Google

  Previous-generation Veo model with native audio, superseded by Veo 3.1. Google's Gemini API pricing page marks Veo 3 as deprecated with a scheduled shutdown of 2026-06-30, so it should be tr

• Veo 3.1 Google

  Generates 4, 6 or 8 second clips with natively generated audio from a text prompt, an image, or up to three reference images. Supports first-and-last-frame interpolation and extension of an

• Veo 3.1 Fast Google

  Lower-latency, lower-cost variant of Veo 3.1 with the same feature surface: native audio, 4K support, up to three reference images, first/last frame and video extension. Model ID veo-3.1-fas

• Veo 3.1 Lite Google

  Cheapest Veo 3.1 tier, capped at 1080p. Unlike the standard and Fast variants it does not support video extension, 4K output, or reference images. Model ID veo-3.1-lite-generate-preview.

• Vidu Q1 Shengshu Technology (Vidu)

  Vidu Q1 is the legacy model kept for existing integrations. It produces exactly 5 second clips at 1080p only, across text-to-video, image-to-video, start-end-to-video and reference-to-video,

• Vidu Q2 Shengshu Technology (Vidu)

  Vidu Q2 is the previous generation and the model that established Vidu's reference-to-video positioning: up to seven reference images covering faces, gestures, scenes or props combined into

• Vidu Q3-mix Shengshu Technology (Vidu)

  Vidu Q3-mix is the reference-to-video-only Q3 variant that Shengshu describes as the strongest overall balance of visual quality, dynamics and intelligent scene transitions, with audio-video

• Vidu Q3-pro Shengshu Technology (Vidu)

  Vidu Q3-pro is Shengshu's top-quality Q3 variant for text-to-video, image-to-video and start-end-to-video, producing 1 to 16 second clips at 540p, 720p or 1080p. Q3 models generate audio tog

• Vidu Q3-turbo Shengshu Technology (Vidu)

  Vidu Q3-turbo is the fast, low-cost Q3 variant covering text-to-video, image-to-video, start-end-to-video and reference-to-video with 540p to 1080p output. Durations run 1 to 16 seconds for

• Wan 2.1 T2V (1.3B and 14B) Alibaba Tongyi Lab (Wan)

  The first widely adopted Apache-2.0 video generation family. T2V-1.3B generates 480P and fits in about 8.19GB of VRAM; T2V-14B handles 480P and 720P. Both produce 81-frame clips at 16 fps (r

• Wan 2.2 I2V-A14B Alibaba Tongyi Lab (Wan)

  Animates a still image into video, optionally guided by a text prompt, using the same 27B/14B-active mixture-of-experts architecture as the text-to-video sibling. Outputs 480P or 720P. Widel

• Wan 2.2 S2V-14B Alibaba Tongyi Lab (Wan)

  Audio-driven video generation: takes a reference image plus an audio track and produces a speaking or performing subject synchronized to that audio. Output length adapts to the length of the

• Wan 2.2 T2V-A14B Alibaba Tongyi Lab (Wan)

  Generates video from a text prompt using a mixture-of-experts diffusion transformer with 27B total and 14B active parameters, split into a high-noise expert for layout and a low-noise expert

• Wan 2.2 TI2V-5B Alibaba Tongyi Lab (Wan)

  A 5B-parameter dense model that handles both text-to-video and image-to-video in one checkpoint. It uses the Wan2.2-VAE with a 16x16x4 compression ratio to generate 720P at 24 fps on a singl

• Wan-Animate-2 Alibaba Tongyi Lab (Wan)

  Character animation and replacement driven directly by a video. It consumes the driving video inside a redesigned diffusion transformer rather than going through an intermediate pose/motion

Other categories

• Image generation 20
• Voice and speech 29
• Music and sound 20
• Editing and post 11
• Upscaling and restoration 8
• Lip sync and avatars 8

Everything in the directory