AI Video Generation Models Directory

HunyuanVideo 1.5 is Tencent's lightweight open-weight video generation model under the Tencent Hunyuan Community License, ranked #58 on the Artificial Analysis Video Arena with an ELO of 1,015. The official repository lists an 8.3B-parameter architecture, 14GB minimum GPU memory with offloading, text-to-video and image-to-video checkpoints, 480p and 720p model cards, and 1080p super-resolution. Hosted provider scope differs: FAL.ai exposes 480p output up to 121 frames at $0.075/sec, while WaveSpeed lists 480p at $0.02/sec and 720p at $0.04/sec for 5, 8, or 10 second clips. It has no native audio or lip-sync; use it for low-cost silent drafts, volume social tests, and self-hosted research workflows.

Pika 2.0 is VidScore's canonical record for Pika's public v2.2 API endpoints. Pika's own API page sends developers to FAL.ai, where Pika Model 2.2 supports text-to-video and image-to-video at 720p ($0.20 per 5-second clip, or $0.04/sec) and 1080p ($0.45 per 5-second clip, or $0.09/sec). WaveSpeed offers Pika v2.2 Pikaframes at $0.04/sec for 720p and $0.06/sec for 1080p. This is separate from Pika 2.5's consumer subscription pricing.

Kling 2.5 Turbo is Kuaishou's speed-optimized video model, ranked #13 on Artificial Analysis Video Arena (no audio) with an ELO of 1,210. It generates 5-10 second clips at 1080p with 30fps output, stronger prompt adherence, and improved camera control over Kling 2.1. On FAL.ai, the Pro tier costs $0.07/second ($0.35 for 5 seconds). WaveSpeed offers the Std tier at $0.042/second ($0.21 for 5 seconds). Also available on Replicate. Priced 30% lower than Kling 2.1, it is one of the most cost-effective high-quality video models available via API.

LTX-2 Pro is Lightricks' open-weight audio-video generation model under the LTX-2 Community License, ranked #37 on the Artificial Analysis Video Arena with an ELO of 1,130. The stable FAL.ai `fal-ai/ltx-2/text-to-video` endpoint is the pricing source for this record: 1080p at $0.06/sec, 1440p at $0.12/sec, and 2160p/4K at $0.24/sec, with synchronized audio, 25/50fps options, 16:9 output, and 6-10 second duration. LTX-2.3 is a newer family upgrade with native portrait support, 20-second clips, and open weights; keep its API pricing separate because public LTX/FAL pages currently use different endpoint and pricing scopes. Entities with at least $10M annual revenue need a paid commercial license.

Kling v3 Standard is Kuaishou's lower-cost Kling 3.0 tier for text-to-video and image-to-video generation. Verified API pricing lists Standard at $0.084/second without native audio and $0.126/second with native audio on FAL.ai and WaveSpeed; FAL.ai also lists voice control at $0.154/second. Current official and provider documentation supports 720p/1080p generation for this tier, so 4K should be treated as unverified rather than a published capability.

Sora 2 is OpenAI's second-generation video model with native audio output. OpenAI's current guide describes Sora 2 and Sora 2 Pro as supporting up to 20-second generations, while the API reference still lists 4/8/12 seconds as exact create-job values, so VidScore avoids treating longer durations as guaranteed endpoint defaults. OpenAI pricing lists Sora 2 at $0.10/second for 720p, Sora 2 Pro at $0.30/second for 720p, $0.50/second for 1024p, and $0.70/second for 1080p. Supports text-to-video, image-to-video, extension, and editing workflows. SUNSET NOTICE: OpenAI says Sora web/app experiences were discontinued on April 26, 2026 and the API shuts down September 24, 2026.

Veo 3.1 is Google DeepMind's flagship video generation model with native audio and lip-sync. Current Vertex AI documentation lists API generation at 720p or 1080p, 24 fps, and 4/6/8-second clips. On FAL.ai, the Standard tier costs $0.20/second without audio or $0.40/second with native audio; the separate Fast variant starts at $0.10/second without audio. Known for strong lip-sync quality and rich audio generation including dialogue, sound effects, and ambient sound. Available via FAL.ai, WaveSpeed, and Replicate.

Veo 3.1 Fast is the speed-optimized variant of Google DeepMind's Veo 3.1. Current Vertex AI documentation lists API generation at 720p or 1080p, 24 fps, and 4/6/8-second clips. On FAL.ai, it starts at $0.10/sec without audio and $0.15/sec with native audio. The Fast tier is cheaper than Veo 3.1 Standard, while keeping the same broad audio-visual model family. Supports text-to-video and image-to-video. Available via FAL.ai, WaveSpeed, and Replicate.

Kling v3 Pro is Kuaishou's higher-quality Kling 3.0 tier for cinematic text-to-video and image-to-video generation. Verified API pricing lists Pro at $0.112/second without native audio and $0.168/second with native audio on FAL.ai and WaveSpeed; FAL.ai also lists voice control at $0.196/second. Current official and provider documentation supports 720p/1080p generation for this tier, so 4K should be treated as unverified rather than a published capability.

Runway Gen-4.5 is Runway's flagship video model for text-to-video and image-to-video. Runway's current help center lists 2-10 second clips, 720p output, 24/25 fps, and six aspect ratios. Through Runway's official API, Gen-4.5 costs $0.12/second (12 credits at $0.01/credit) and is also available on Replicate. Treat audio, video-to-video, and higher-resolution claims as unverified for Gen-4.5 unless the endpoint explicitly documents them.

HappyHorse 1.0 is Alibaba's video generation model for text-to-video, image-to-video, reference-to-video, and video editing workflows. Alibaba Cloud Model Studio now documents HappyHorse API endpoints, including `happyhorse-1.0-t2v` for text-to-video and `happyhorse-1.0-video-edit` for video editing. FAL.ai lists live Happy Horse API endpoints under `alibaba/happy-horse/*` with 720p pricing at $0.14/sec and 1080p pricing at $0.28/sec for text-to-video, image-to-video, and reference-to-video. Replicate also lists `alibaba/happyhorse-1.0` with matching 720p/1080p output pricing. The FAL video-edit page currently has conflicting price copy, so VidScore does not use video-edit pricing as a canonical cost floor. Public weights, inference code, and open-source license terms are still not verified, so VidScore treats HappyHorse as API-available but not self-deployable.

Seedance 2.0 is ByteDance's flagship video generation model with unified multimodal audio-video architecture. ByteDance's Seed page says it accepts text, image, audio, and video inputs simultaneously — up to 9 reference images, 3 reference videos, and 3 audio files — and supports 15-second multi-shot audio-video output. Verified public provider pricing ranges from $0.10/sec (WaveSpeed Fast 480p) to $0.303/sec (FAL.ai Standard 720p). FAL.ai, WaveSpeed, and Replicate currently have the cleanest public evidence; Atlas Cloud now lists roughly $0.101-$0.127/sec, but VidScore keeps it outside verified-floor rankings until its API docs and endpoint behavior are independently rechecked. Also available through ByteDance/Dreamina-style consumer workflows. Strong for director-style camera control, native dialogue with lip-sync, and beat-synchronized audio.

CogVideoX-5B is Zhipu AI and Tsinghua University's open-source text-to-video model, ranked #72 on Artificial Analysis Arena with an ELO of 785. At $0.20 per clip on FAL.ai or $0.32 on Replicate (~$2.00-3.20/min), it is one of the most affordable API options available. The 5B-parameter model uses a 3D Expert Transformer with 3D Causal VAE and 3D Rotary Position Embeddings for spatiotemporal fusion. Native output is 8 fps at 720x480, with RIFE interpolation to 16 or 32 fps. Supports text-to-video, image-to-video, and video-to-video on FAL.ai. Trained on 35 million video clips, it supports LoRA fine-tuning and deterministic generation via seed control — positioning it as a budget-friendly, customizable baseline for developers building video pipelines.

FramePack is an Apache 2.0 image-to-video model by lllyasviel that compresses input contexts to a constant length, making the generation workload largely independent of video length. The official repository documents 6GB minimum GPU memory and a 60-second / 1,800-frame local example at 30fps on consumer GPUs. The hosted FAL.ai endpoint is much shorter: it is image-to-video, costs $0.0333/sec, supports multiple aspect ratios, and caps a generation at 180 frames. Treat long clips as a self-hosted/local capability, not a hosted API limit.

Grok Imagine Video by xAI is a fast, competitively priced video generation model built on the Aurora mixture-of-experts architecture. It ranks #6 on the Artificial Analysis Text-to-Video Arena (ELO 1,227) and #3 in Image-to-Video (ELO 1,330). At $0.05/sec (480p) or $0.07/sec (720p) on FAL.ai and $0.055/sec on WaveSpeed, it offers some of the lowest per-second API costs available. Generation speed is ~17 seconds from prompt to finished output — 2-4x faster than competitors. It supports text-to-video, image-to-video, video-to-video editing, and video extend with native audio including dialogue, near-perfect lip-sync, ambient sounds, and effects. Available on FAL.ai, WaveSpeed, and Replicate.

Hailuo 02 Pro is MiniMax's flagship video generation model, ranked #33 on the Artificial Analysis Video Arena with an ELO of 1,161. At $4.80/min ($0.08/sec) on FAL.ai, it generates 1080p video in 6 or 10 second clips at 25 fps. The NCR architecture enables 2.5x faster training and inference with 3x more parameters and 4x more training data than its predecessor. It excels at physics-accurate simulations including fluid dynamics, cloth, collisions, and acrobatics. Available via FAL.ai and WaveSpeed with both text-to-video and image-to-video modes.

Luma Ray 3 is Luma AI's third-generation video model and the world's first reasoning video model, ranked on Artificial Analysis Video Arena (text-to-video without audio) with an ELO of 1,204. Updated to Ray3.14 in January 2026 with native 1080p, 4x faster generation, and 3x lower cost. It is the first model to generate native 16-bit HDR video, outputting studio-grade ACES EXR format. Native generation maxes out at 1080p; 4K is achieved via a neural upscaler (Hi-Fi Diffusion mastering), not native generation. Luma's Dream Machine credit documentation lists Ray3/Ray3.14 credit tiers, but the public Luma API pricing page currently lists Ray 2 and Ray Flash 2 rather than a flat Ray 3 API price. Ray3 audio is not available. Not yet available on FAL.ai or Replicate. Supports text-to-video, image-to-video, video-to-video (Modify), keyframes, extend, loop, and the Ray3 Modify mode for hybrid acting workflows.

Luma Ray2 is Luma AI's second-generation video model built on a multimodal transformer architecture. Known for realistic motion physics and sophisticated camera control with 20+ preset camera movements including dolly zoom, crane, orbit, and tracking shots. On FAL.ai, it generates 5-second 540p clips at $0.10/second, with 2x multiplier for 720p and 4x for 1080p. Also available on Replicate in 540p and 720p variants, WaveSpeed at $0.08/sec (720p), and via Luma's official API. Luma's official API lists Add Audio at $0.02/sec on top of video generation; third-party Ray2 provider rows in this catalog are video-only unless they explicitly list an audio tier. Supports text-to-video, image-to-video, and video extension up to 30 seconds.

Minimax Hailuo 02 is MiniMax's second-generation AI video model, known for its ultra-low API pricing and strong cinematic motion quality. On FAL.ai, the Standard tier generates 768p video at just $0.045/second, making it one of the cheapest per-second API options available. The Pro tier outputs native 1080p at $0.08/second. It supports both text-to-video and image-to-video generation with durations of 6 or 10 seconds. The newer Hailuo 2.3 variant adds anime/illustration styles and a Fast efficiency mode.

Mochi 1 is Genmo's 10-billion-parameter open text-to-video model, ranked #61 on the Artificial Analysis Arena with an ELO of 1,000. The official GitHub repository releases the model under Apache 2.0, provides downloadable weights, and documents 480p generation with a 10B AsymmDiT architecture and 128x latent compression. FAL.ai lists `fal-ai/mochi-v1` at $0.40 per video, while Replicate lists `genmoai/mochi-1` as an available hosted model with hardware-time billing. It remains a strong research and self-hosting baseline, but it has no native audio, no lip-sync, and a 5-second / 480p practical hosted profile.

Pika 2.5 is Pika Labs' current consumer/web model for fast social video, creative effects, and keyframe workflows. Pika's official pricing page lists Pika 2.5 access inside subscription credits: Basic is limited to 480p, while paid plans unlock 720p/1080p text-to-video, image-to-video, Pikaframes, Pikascenes, Pikadditions, Pikaswaps, Pikatwists, and all Pikaffects. Pika's public API page points developers to FAL.ai, but FAL currently exposes Pika Model 2.2 endpoints rather than verified full Pika 2.5 API pricing. VidScore therefore treats Pika 2.5 as a consumer-platform model with no verified public Pika 2.5 API pricing, separate from the Pika v2.2 API record.

PixVerse V6 is PixVerse's sixth-generation model, launched March 30, 2026. It ranks #16 on the Artificial Analysis Text-to-Video Arena (ELO 1,209) and #4 in Image-to-Video (ELO 1,312). It stands out for its breadth of modes: text-to-video, image-to-video, transition, extend, and lip-sync — all available via API. On FAL.ai, pricing starts at $0.025/sec for 360p without audio; the cheapest audio tier is $0.035/sec at 360p, and 1080p with audio is $0.115/sec. V6 introduces 20+ cinematic lens controls (focal length, aperture, DoF, chromatic aberration), a physics simulation engine for realistic fluid and material interactions, and multi-shot generation with native audio. Available on FAL.ai, WaveSpeed, and the PixVerse Platform API.

Runway Gen-4 is Runway's fourth-generation video model featuring granular camera controls, Motion Brush for region-specific motion direction, and reference image consistency. It generates 5-10 second clips at 720p, with 4K handled as a separate upscale action rather than native generation. Through Runway's direct API, Gen-4 Turbo costs $0.05/second (5 credits/sec), Gen-4 Video costs roughly $0.12/second from Runway plan-credit math, and Gen-4 Aleph (gen4_aleph, video-to-video) costs $0.15/second (15 credits/sec). WaveSpeed also offers Gen4 Aleph at $0.18/second. Use Gen-4 for Motion Brush and controlled image-to-video; use Aleph for video-to-video editing. It does not have verified native audio or Gen-4 extend support.

Seedance 1.5 Pro is ByteDance Seed's joint audio-video model for prompt-following video with synchronized voices and sound effects. The official ByteDance page confirms native audio generation, wide language/dialect support, strong lip-sync and motion alignment, cinematic camera movement, and cohesive storytelling. FAL.ai lists the Pro v1.5 text-to-video endpoint at roughly $0.26 for a 5-second 720p clip with audio, with token pricing of $2.40 per million video tokens with audio and $1.20 per million without audio; controls cover 480p/720p, 4-12 second duration, audio on/off, six aspect ratios, and camera lock. Replicate also lists `bytedance/seedance-1.5-pro` as available. It is closed-source and not self-deployable.

SkyReels V4 by Skywork AI is a multimodal video-audio generation, inpainting, and editing model described in the public arXiv technical report. The report states that V4 accepts text, images, video clips, masks, and audio references, uses a dual-stream MMDiT architecture, and supports up to 1080p, 32fps, and 15-second output with synchronized audio. API access appears limited/invite-heavy, and the official pricing page is JavaScript-only in static crawls; keep the $0.12/sec official-API rate as reported until direct rate-card access is rechecked. V4 is not treated as verified open-source or self-hostable until public V4 weights, runnable inference code, and license terms are available.

Vidu Q3 Pro is Shengshu Technology's flagship video model, ranked #7 on Artificial Analysis Text-to-Video Arena (no audio) with an ELO of 1,223 and #9 on Image-to-Video (no audio) with ELO 1,288. It generates up to 16 seconds of native audio-video in a single pass at up to 1080p with lip-synced dialogue, the longest single-generation duration among leading models. On FAL.ai, pricing starts at $0.07/second for 360p-540p with a 2.2x multiplier for 720p/1080p ($0.154/sec). WaveSpeed offers text-to-video at $0.07/sec (540p), $0.15/sec (720p), and $0.16/sec (1080p). Also available on Replicate at matching per-second rates. Features native dialogue with lip-sync, SFX, and background music generation.

Wan 2.1 is Alibaba's open-source video generation model, released in February 2025 under Apache 2.0. Available in 1.3B and 14B parameter variants, it supports text-to-video and image-to-video generation at 480p and 720p resolution. On FAL.ai, Wan 2.1 costs $0.20/generation at 480p and $0.40/generation at 720p for clips up to 5 seconds. The VACE extension (released May 2025) adds video-to-video editing, video extension, and basic camera control. Wan 2.1 does NOT support native audio or lip-sync — those features were introduced in Wan 2.5. For higher resolution (1080p), native audio, and advanced capabilities, see Wan 2.7.

Wan 2.7 is Alibaba's hosted video generation suite for text-to-video, image-to-video with first/last-frame control, reference-to-video with subject and voice references, and instruction-based video editing. At $0.10/second on FAL.ai (~$6.00/min for 1080p), it outputs 720p/1080p at 30fps with native audio and 2-15 second durations depending on endpoint. Official Alibaba Cloud materials say Wan2.7-Video is available through Model Studio and the Wan website; no downloadable open weights or open-source model release is verified as of April 18, 2026.