AI video generation is the use of artificial intelligence systems, predominantly diffusion transformers, to create video clips from text descriptions, still images, or other video, producing sequences of frames with coherent motion, consistent characters, physically plausible behavior, and increasingly synchronized audio. The leading models as of 2026 include OpenAI's Sora 2, Google DeepMind's Veo 3.1, Kuaishou's Kling 3.0, Runway Gen-4.5, and ByteDance's Seedance, which generate cinematic-quality footage at up to 1080p (with 4K via upscaling) and durations of roughly 8 to 20 seconds. The technology extends the principles of AI image generation into the temporal domain, and has progressed since 2022 from short, glitchy clips to footage that OpenAI describes as approaching "the GPT-3.5 moment for video." [9]

The field represents one of the most technically demanding frontiers of generative AI. Generating video requires not only the visual quality demanded of individual frames but also temporal consistency (objects and characters must look the same from frame to frame), physically plausible motion (objects must obey gravity, momentum, and collision), and narrative coherence (scenes must develop logically over time). These requirements make video generation substantially harder than image generation, and the rapid progress since 2023 has been one of the most impressive demonstrations of AI capability. Adoption has scaled accordingly: Google reported that users generated millions of Veo 3 videos within days of its May 2025 launch, and Kuaishou reported that its Kling AI tool reached more than 60 million creators worldwide who had produced over 600 million videos by December 2025. [10][11]

History

AI video generation has progressed through several technological eras, each expanding what is possible in terms of quality, duration, and controllability.

Early GAN-Based Approaches (2016-2021)

The earliest attempts at AI video generation used generative adversarial networks (GANs), extending the same adversarial training framework that had proven successful for image generation. Video GAN variants like VGAN (2016) and MoCoGAN (2018) attempted to separate motion and content representations, generating short clips of simple actions like faces turning or grass swaying in the wind.

These early models were severely limited. They could only produce a few seconds of low-resolution video (typically 64x64 or 128x128 pixels), struggled with complex scenes, and exhibited obvious artifacts including temporal flickering, object morphing, and physically impossible movements. The fundamental challenge was that GANs already had difficulty generating single high-quality images; extending them to generate temporally consistent sequences of images compounded every existing limitation.

Make-A-Video by Meta (2022)

A significant milestone came in September 2022, when Meta AI published Make-A-Video, one of the first systems to demonstrate convincing text-to-video generation [1]. Make-A-Video took a novel approach: rather than training a video generation model from scratch on paired text-video data (which was scarce and expensive to collect), it leveraged a pre-trained text-to-image model and extended it with temporal layers that learned motion dynamics from unlabeled video data.

The key insight was to decompose the problem: learn visual appearance from text-image pairs (abundant on the internet), then learn motion from video data (available without text labels). This approach allowed Make-A-Video to generate short video clips from text descriptions, image animations, and variations on existing videos. While the results were still limited in resolution and duration (typically a few seconds at low resolution), Make-A-Video demonstrated that the diffusion model framework could be extended from images to video.

Runway Gen-1 and Gen-2 (2023)

Runway, a startup founded by former NYU researchers, released Gen-1 in February 2023 and Gen-2 in June 2023 [2]. Gen-1 introduced video-to-video generation, allowing users to transform the style and content of existing videos. Gen-2 added text-to-video capabilities, generating clips of up to 4 seconds at 720p resolution. While the results were rough by current standards, Runway was the first company to offer AI video generation as a widely accessible commercial product, bringing the technology out of research labs and into the hands of creators.

Sora Preview (February 2024)

On February 15, 2024, OpenAI published a technical preview of Sora, a text-to-video model capable of generating up to one minute of high-fidelity video [3]. The preview videos stunned the AI community and the public with their visual quality, physical realism, and narrative coherence. OpenAI described Sora as a step toward building general-purpose simulators of the physical world, writing that "video generation models as world simulators" could learn physical dynamics from scale. [3]

Sora's preview represented a quantum leap over existing models. The generated videos showed complex scenes with multiple characters, realistic camera movements, accurate reflections and lighting, and physically plausible interactions. OpenAI did not immediately release Sora publicly, instead providing access to red teamers and creative professionals for evaluation. OpenAI later characterized the original Sora as "in many ways the GPT-1 moment for video," where "simple behaviors like object permanence emerged from scaling up pre-training compute." [9]

Rapid Competition (2024-2025)

Sora's preview triggered an intense competitive response across the industry. Throughout 2024 and into 2025, multiple companies released increasingly capable video generation models, and the pace of improvement accelerated dramatically. Runway shipped Gen-4 on March 31, 2025, with reference-image-based character consistency; Google unveiled Veo 3 with native audio at Google I/O in May 2025; ByteDance released Seedance 1.0 in June 2025; and OpenAI released Sora 2 in September 2025. [10][12][13]

Current Models (2025-2026)

The AI video generation landscape in early 2026 features a diverse range of models with varying strengths, capabilities, and access models.

Model	Developer	Release	Max Duration	Max Resolution	Audio	Key Strength
Sora 2	OpenAI	September 2025	20 seconds	1080p	Synchronized	Cinematic quality, realistic physics
Veo 3.1	Google DeepMind	October 2025	8 seconds (extendable)	1080p (4K upscale)	Native dialogue + SFX	Photorealism, audio quality
Runway Gen-4.5	Runway	2025	10 seconds	1080p	Yes	Physics accuracy, character consistency
Pika 2.1	Pika Labs	2025	16 seconds	1080p	Lip-sync	Scene ingredients, social content
Kling 3.0	Kuaishou	February 2026	15 seconds	1080p	Simultaneous A/V	Multi-shot sequences, subject consistency
Seedance 1.0	ByteDance	June 2025	~10 seconds	1080p	Yes (Seedance 2.0)	Multi-shot, leaderboard-topping quality
Hailuo Video-01	MiniMax	2024-2025	6 seconds	720p-1080p	No	Best value, strong text-to-video
Luma Ray3	Luma Labs	September 2025	10 seconds	1080p HDR	No	First reasoning video model, native HDR
Stable Video Diffusion	Stability AI	2023-2024	4 seconds	576x1024	No	Open-source, local deployment

What is Sora 2?

OpenAI released Sora 1.0 to ChatGPT Plus and Pro users in December 2024, initially limited to the US and Canada. Sora 2, released on September 30, 2025 alongside a dedicated invite-only Sora iOS app, represents a substantial upgrade with improved physics simulation, synchronized dialogue and sound effects, and stronger multi-shot consistency. [9] OpenAI framed the release in generational terms: "With Sora 2, we are jumping straight to what we think may be the GPT-3.5 moment for video." [9] The company emphasized world-modeling over surface realism, noting that a key test of a useful world simulator is the ability to "model failure, not just success": in Sora 2, a missed basketball shot now rebounds off the backboard rather than teleporting into the hoop. [9] The model excels at producing cinematic footage where light behaves as a real lens would capture it and motion follows believable physics. A signature feature, Cameo, lets a user insert a short selfie video of themselves into any Sora 2 scene. Sora 2 is available through ChatGPT subscriptions and the Sora application.

How does Google Veo 3 differ, and when was Veo 3.1 released?

Google DeepMind's Veo line emerged as a quality leader in benchmark and human-preference testing. Veo 3, announced at Google I/O in May 2025, was the first major model to generate native dialogue, sound effects, and ambient audio jointly with video in a single pass, eliminating the need for separate audio generation. [10] Demand outstripped supply at launch: DeepMind CEO Demis Hassabis said usage was so heavy it risked "our wonderful TPUs from melting," and Google Labs VP Josh Woodward described "way, way, way more demand than we expected." [10] Google reported that users generated over 40 million Veo 3 videos in the weeks after launch. [10]

Veo 3.1, released on October 14, 2025, roughly five months after Veo 3, added richer native audio, editing tools (Insert, Remove, Extend), and Flow integration features such as Ingredients to Video and Frames to Video. [14] Veo 3.1 outputs natively at 720p or 1080p at 24 fps with clip lengths of 4, 6, or 8 seconds, extendable well past two minutes via the Extend feature, with 4K available through upscaling; API pricing starts at $0.15 per second (Fast) and $0.40 per second (Standard). [14]

Runway Gen-4 and Gen-4.5

Runway has continued to iterate rapidly on its generation technology. Gen-4, released on March 31, 2025, introduced consistent characters, objects, and environments across multiple separate shots by conditioning generation on up to three user-supplied reference images, without retraining or extra compute. [12] Gen-4.5 then solved what Runway calls the "floaty physics" problem that plagued earlier models [4]. Objects now exhibit convincing weight and momentum, collisions look realistic, and characters move with natural biomechanics. Runway targets professional filmmakers and content creators with features for fine-grained camera control, character consistency across scenes, and integration with traditional video editing workflows.

Kling by Kuaishou

Kling, developed by Chinese technology company Kuaishou, has been one of the most innovative players in the space. Kling 2.6, released in December 2025, introduced "simultaneous audio-visual generation," creating visuals, natural voiceovers, sound effects, and ambient atmosphere in a single generation pass rather than requiring separate audio and video workflows [5]. Kling 3.0, released in February 2026, introduced multi-shot sequences of 3 to 15 seconds with subject consistency across different camera angles, a major technical breakthrough for narrative video generation. Kling has also become a commercial standout: Kuaishou reported that Kling AI exceeded USD 20 million in monthly revenue in December 2025, an annualized revenue run rate of USD 240 million, serving more than 60 million creators worldwide, over 600 million videos generated, and partnerships with more than 30,000 enterprise users. [11] Kling AI generated roughly USD 150 million (about RMB 1.04 billion) in full-year 2025 revenue. [15]

What is ByteDance Seedance?

Seedance 1.0 is a video generation foundation model from ByteDance, released through its Volcano Engine platform on June 11, 2025. [13] At launch it ranked first on both the text-to-video and image-to-video leaderboards of the third-party benchmarking site Artificial Analysis, outperforming Veo 3 and Kling 2.0 by over 100 points on the image-to-video task. [13] Seedance 1.0 generates 1080p video with seamless multi-shot transitions and strong motion stability, and its technical report notes it can produce a five-second 1080p clip in about 41 seconds on an NVIDIA L20, substantially faster than several commercial peers. [13] By early 2026, the follow-up Dreamina Seedance 2.0 led the Artificial Analysis text-to-video arena (with audio) with an Elo score around 1,218, ahead of Kling 3.0 and Veo 3.1. [4]

Pika

Pika, founded by former Stanford AI researchers, focuses on accessible, social-media-friendly video generation. Pika 2.1 features a "scene ingredients" system that maintains visual consistency across different scenes, with particularly strong image-to-video conversion that can transform a single static image into a dynamic narrative [6]. The tool produces clips of 10 to 16 seconds at up to 1080p resolution with improved lip-sync capabilities.

Hailuo by MiniMax

MiniMax's Hailuo video model offers strong text-to-video capabilities at an accessible price point ($14.99/month for comprehensive access), positioning it as a budget-friendly option without significant quality sacrifices [6]. The model is particularly popular in the Chinese market and has gained international traction through competitive pricing.

Luma Dream Machine / Ray3

Luma Labs made waves with what it describes as the world's first "reasoning" video model. Ray3, released in September 2025, is also the first model to generate native 16-bit HDR video, bringing AI-generated output into professional studio color pipelines [6]. Dream Machine covers text-to-video, image-to-video, and text-directed video editing capabilities.

Stable Video Diffusion

Stability AI's open-source Stable Video Diffusion provides a free alternative that can be run locally on consumer hardware. While it trails the commercial models in quality and duration (generating up to 4 seconds at moderate resolution), it serves as a valuable baseline for research and for users who prioritize privacy, customization, or cost savings over raw quality.

Which AI video model is best in 2026?

There is no single "best" model; rankings shift with each release and depend on the task (text-to-video versus image-to-video, with or without audio). The most-watched independent measure is the Artificial Analysis Video Arena, which ranks models by Elo rating from blind human votes on videos generated from the same prompt. In early 2026 its text-to-video arena (with audio) was led by Dreamina Seedance 2.0, followed closely by Kling 3.0 variants, with Veo 3.1 and Sora 2 also among the leaders. [4][13] Because the arena is updated continuously as new models are added, leadership has changed hands repeatedly between ByteDance, Kuaishou, Google, and OpenAI throughout 2025 and 2026.

How It Works

Modern AI video generation builds on the diffusion model framework that powers image generation, with critical extensions to handle the temporal dimension of video.

Video Diffusion Transformers

The dominant architecture for current video generation models is the video diffusion transformer (VDT), which combines diffusion-based generation with the transformer architecture. OpenAI's technical report on Sora provides the most detailed public description of this approach [3].

The pipeline works as follows. First, a video is encoded into a compressed latent representation. This latent representation is then divided into "spacetime patches," small chunks that span both spatial dimensions and time. These patches function analogously to tokens in a large language model, each representing a piece of the video in both space and time. A transformer network operates on these patches, learning the relationships between different spatial locations and different time steps simultaneously.

During generation, the model starts with random noise in the latent space and iteratively denoises it, guided by a text description encoded by a language model. The denoising process produces a clean latent representation that is then decoded back into pixel space to produce the final video frames.

This architecture differs from earlier approaches that used temporal convolutions or separate spatial and temporal processing. By treating video as a unified spacetime structure, the transformer can learn long-range dependencies both across the frame (spatial) and across time (temporal), producing more coherent and consistent results.

Temporal Consistency

The central technical challenge in video generation is temporal consistency: ensuring that objects, characters, textures, and lighting remain stable and coherent from frame to frame. Inconsistencies manifest as flickering textures, morphing facial features, objects that change size or shape, and backgrounds that shift unexpectedly.

Modern models address temporal consistency through several mechanisms:

Spacetime attention. The transformer's attention mechanism operates across both spatial and temporal dimensions, allowing each patch to attend to patches at other time steps. This lets the model maintain awareness of what happened in previous and subsequent frames while generating any given frame.

Temporal conditioning. Some models generate video autoregressively, conditioning each new chunk of frames on previously generated frames. This ensures continuity but can accumulate errors over long sequences.

Motion modeling. Advanced models explicitly learn motion dynamics, including how objects move under gravity, how fabrics drape and flow, how liquids splash, and how rigid bodies collide. This physics-aware generation is what distinguishes current models from earlier approaches that produced "floaty" or physically implausible motion.

Audio Generation

A major development in 2025 was the integration of audio generation into video models. Veo 3, Kling 2.6, and Sora 2 can all generate synchronized audio alongside video, including ambient sound effects, voiceovers, and even dialogue that matches lip movements [4][5]. Veo 3 was the first major model to do this jointly with video in a single pass; in prompts, dialogue is specified inside quotation marks while sound effects and ambience are described in plain text. [10] This is typically achieved by training the model on video with paired audio, allowing it to learn the correlations between visual events and their corresponding sounds (a ball bouncing produces a thud, a door closing produces a click, rain creates a patter).

Capabilities

AI video generation encompasses several distinct modalities and features.

Text-to-Video

The foundational capability is generating video from a text description. Users write a prompt describing the desired scene, characters, actions, setting, camera movement, and visual style, and the model generates a corresponding video clip. Prompt engineering for video requires describing not just static visual elements (as in image generation) but also temporal dynamics: what happens over the course of the clip.

Image-to-Video

Image-to-video generation animates a still image, adding motion, camera movement, and scene dynamics while preserving the visual content of the original image. This capability is particularly valuable for animating concept art, product shots, and photographs. The model infers plausible motion from the static image: a photo of ocean waves might be animated with rolling water, a portrait might gain subtle head movement and blinking, and a landscape might develop swaying trees and drifting clouds.

Video-to-Video

Video-to-video transformation applies stylistic or content changes to existing video footage while preserving the original motion and structure. A live-action clip might be transformed into animation, a daytime scene might become nighttime, or the visual style might be shifted to match a particular aesthetic. This capability builds on the img2img techniques developed for image generation, extended to operate consistently across video frames.

Camera Control

Advanced models provide explicit control over camera movement, including pans, tilts, dolly shots, crane movements, and tracking shots. Some models accept camera path specifications (defining the exact trajectory of the virtual camera through 3D space), while others respond to natural language descriptions of camera behavior ("slowly zoom in on the character's face" or "orbiting shot around the building"). Camera control is critical for professional filmmaking applications where specific framing and movement are essential.

Multi-Shot and Scene Transitions

Kling 3.0's introduction of multi-shot sequences in February 2026 represents a step toward AI-generated narrative video [5]. The model can generate multiple shots of 3 to 15 seconds each with consistent characters and settings across different camera angles. This capability moves beyond single-clip generation toward the production of short scenes with cinematic editing.

Limitations

Despite rapid progress, AI video generation in 2026 still faces significant limitations.

Duration

Most models generate clips of 4 to 20 seconds, far short of the minutes or hours required for full video production. While Sora's original preview demonstrated one-minute generation, production models are typically limited to shorter durations due to quality degradation, computational cost, and temporal consistency challenges over longer timescales.

Fine-Grained Control

Controlling exactly what happens in a generated video remains difficult. While text prompts can specify general actions and settings, achieving precise control over specific movements, timing, spatial relationships, and narrative beats is unreliable. Professional filmmakers often need exact control over blocking, timing, and composition that current models cannot consistently deliver.

Character and Object Consistency

Maintaining consistent appearance of characters and objects across different clips or within longer videos remains challenging. A character generated in one clip may look subtly different when generated in another, making it difficult to produce multi-scene content with the same characters. Runway Gen-4's reference-image conditioning and Kling 3.0's multi-shot consistency features address this limitation but do not fully solve it. [12]

Physics and Anomalies

While physics simulation has improved dramatically, current models still produce occasional physically impossible results: objects passing through each other, shadows pointing in inconsistent directions, liquids behaving unrealistically, and human hands or limbs contorting unnaturally. These anomalies have diminished in frequency but have not been eliminated.

Computational Cost

Video generation is extremely computationally expensive. Generating a single 10-second clip can take minutes on powerful GPU clusters and costs significantly more than image generation. Pricing exposes this directly: Veo 3.1 lists generation at $0.15 to $0.40 per second of video through the Gemini API, so a single eight-second clip costs roughly $1.20 to $3.20. [14] This limits both the accessibility of the technology and the ability to iterate rapidly on outputs.

Applications

AI video generation has found practical applications across multiple industries, though adoption is still in relatively early stages compared to AI image generation.

Advertising and Marketing

Brands and agencies use AI video generation to produce social media content, product demonstrations, and advertising concepts. The technology enables rapid iteration on creative concepts, generation of localized content variations, and production of video assets at a fraction of the cost and time of traditional video production. Short-form social media platforms like TikTok and Instagram Reels are particularly well-suited to AI-generated video, since the short duration and high volume of content align with current model capabilities.

Film Pre-Visualization

Filmmakers use AI video generation for pre-visualization (previs), creating rough versions of scenes before committing to expensive physical production. Directors and cinematographers can explore camera angles, lighting setups, and scene compositions using AI-generated footage, then use the results to plan actual shoots. This application leverages AI's speed and low cost for exploration while relying on traditional production for final output.

Education and Training

AI-generated video supports educational content creation, training simulations, and instructional materials. Complex concepts can be visualized dynamically, historical events can be recreated, and training scenarios can be generated without the expense of live production. Medical, military, and industrial training applications benefit from the ability to generate scenario-specific video content on demand.

Social Media and Content Creation

Individual creators use AI video generation for social media content, YouTube videos, and online storytelling. The low barrier to entry (requiring only a text prompt and a subscription) enables people without filmmaking skills or equipment to produce video content. This has expanded the creator ecosystem while raising questions about content authenticity and disclosure.

Game Development

Game developers use AI video generation for cutscenes, cinematic trailers, and concept visualization during development. The technology can produce placeholder footage early in development, helping teams align on visual direction before committing to full production.

Controversies

The Studio Ghibli Controversy (2025)

The most prominent controversy in AI video and image generation during 2025 centered on the "Ghiblification" trend triggered by GPT-4o's image generation capabilities in March 2025 [7]. Users flooded social media with AI-generated images mimicking Studio Ghibli's distinctive hand-drawn animation style. The trend provoked a sharp backlash from artists and animation professionals who viewed it as disrespectful to the studio's decades of painstaking hand-crafted work.

Studio Ghibli co-founder Hayao Miyazaki, who has publicly expressed contempt for AI-generated art, became a symbol of artistic resistance to AI generation. The irony that AI was being used to imitate one of the world's most famous opponents of AI art intensified the controversy. By November 2025, Studio Ghibli and other Japanese content creators formally asked OpenAI to stop using their work for training [7].

The Ghibli controversy crystallized broader tensions between AI capabilities and artistic rights. It demonstrated that even when AI generation does not copy specific works, mimicking a distinctive style can feel like theft to the artists and studios who developed that style over years or decades of work.

Deepfakes and Misinformation

As video generation quality improves, concerns about deepfakes and misinformation intensify. AI-generated video can be used to create convincing footage of events that never happened, put words in people's mouths, or fabricate evidence. Election cycles and political campaigns are particularly vulnerable to AI-generated video disinformation. Detection tools exist but are engaged in a continuous arms race with generation capabilities.

Labor Displacement

Professionals in video production, motion graphics, stock footage, and visual effects have raised concerns about AI video generation's impact on employment. While the technology currently supplements rather than replaces most professional video production (due to control and quality limitations), rapid improvement suggests that more tasks will become automatable in the near future. The advertising and stock footage industries are likely to be affected first.

Consent and Likeness Rights

AI video generation raises questions about the use of people's likenesses. Models trained on public video data learn to generate realistic human faces and bodies, and can potentially be directed to generate video featuring people who did not consent to their likeness being used. Sora 2's Cameo feature, which inserts a user's own likeness into generated scenes, foregrounded these questions by making personal likeness a first-class input. [9] This intersects with existing deepfake concerns but extends to any realistic human depiction in AI-generated video.

Market Size

The AI video generation market is growing rapidly from a smaller base than the broader AI image or text generation markets.

Metric	Value	Source
2025 market size	$716.8 - $788.5 million	Fortune Business Insights; Grand View Research
2026 projected	$847 million	Fortune Business Insights
2033-2034 projected	$3.35 - $3.44 billion	Multiple sources
CAGR (2026-2033)	18.8% - 20.3%	Fortune Business Insights; Grand View Research

While the market is still under $1 billion in 2025, strong double-digit growth is projected over the next decade as model capabilities improve and enterprise adoption increases [8]. The advertising, entertainment, and social media sectors are expected to drive the largest share of spending. Single-product revenue is already approaching these market estimates: Kuaishou reported a USD 240 million annualized run rate for Kling AI alone as of December 2025. [11]

Current State (2025-2026)

Several trends define the AI video generation landscape in early 2026.

Quality Leap

The most striking development is the sheer quality of generated video. Native resolution has reached 1080p across the leading models, with 4K available through upscaling on Veo 3.1. Physics simulation now produces believable real-world interactions, demonstrated by OpenAI's emphasis on Sora 2 correctly modeling a missed basketball shot rebounding off the backboard. [9] Light behaves realistically and human motion looks natural. The gap between AI-generated and traditionally produced video has narrowed dramatically, though it has not closed entirely.

Native Audio Integration

The arrival of synchronized audio generation in consumer video tools (Veo 3, Kling 2.6, Sora 2) represents a qualitative shift. Previously, AI-generated video was silent, requiring manual audio production. Veo 3 pioneered joint audio-visual generation at Google I/O 2025, and models now generate ambient sounds, sound effects, voiceovers, and even dialogue matched to lip movements, moving closer to complete video production in a single generation step [4][5][10].

From Clips to Scenes

The introduction of multi-shot capabilities (pioneered by Kling 3.0) signals a transition from generating isolated clips to generating coherent scenes with multiple camera angles and consistent characters [5]. This is a necessary step toward AI-generated narrative content and represents the early stages of AI video editing and directing capabilities.

Democratization

AI video generation is becoming accessible to non-professionals. Most models are available through web interfaces or chat applications at subscription prices ranging from free tiers to $30 per month. This accessibility is enabling a new wave of content creators who can produce video without cameras, studios, or production crews, and it shows up in usage figures: Kling AI alone reported more than 60 million creators worldwide by December 2025. [11]

Professional Skepticism

Despite impressive demos, professional filmmakers and video producers remain cautious. The lack of fine-grained control, limited duration, inconsistent character appearance, and remaining physics artifacts mean that AI video generation is currently more useful for ideation and pre-visualization than for final production. The technology is widely viewed as a powerful complement to traditional production rather than a replacement, though this calculus shifts with each new model release.

References

[1] Singer, U., et al. (2022). "Make-A-Video: Text-to-Video Generation without Text-Video Data." Meta AI. arXiv:2209.14792. https://arxiv.org/abs/2209.14792

[2] "Runway Gen-2: AI Video Generation." Runway ML, 2023. https://research.runwayml.com/gen2

[3] "Video generation models as world simulators." OpenAI, February 2024. https://openai.com/index/video-generation-models-as-world-simulators/

[4] "Text to Video Leaderboard." Artificial Analysis Video Arena, 2026. https://artificialanalysis.ai/video/leaderboard/text-to-video

[5] "15 AI Video Models Tested: Kling 3.0 vs Veo 3.1." TeamDay.ai, 2026. https://www.teamday.ai/blog/best-ai-video-models-2026

[6] "Ultimate AI Video Generation Models Guide 2025." UlazAI, 2025. https://ulazai.com/ai-video-models-guide-2025/

[7] "Studio Ghibli and other Japanese publishers want OpenAI to stop training on their work." TechCrunch, November 2025. https://techcrunch.com/2025/11/03/studio-ghibli-and-other-japanese-publishers-want-openai-to-stop-training-on-their-work/

[8] "AI Video Generator Market Size, Share." Fortune Business Insights, 2025. https://www.fortunebusinessinsights.com/ai-video-generator-market-110060

[9] "Sora 2 is here." OpenAI, September 30, 2025. https://openai.com/index/sora-2/

[10] "Google says Veo 3 users have generated millions of AI videos in just a few days." The Decoder, May 2025. https://the-decoder.com/google-says-veo-3-users-have-generated-millions-of-ai-videos-in-just-a-few-days/

[11] "Kling AI Annualized Revenue Run Rate Hits USD240 Million in December 2025." Kuaishou Technology, January 13, 2026. https://www.prnewswire.com/news-releases/kling-ai-annualized-revenue-run-rate-hits-usd240-million-in-december-2025-302659847.html

[12] "Runway's New AI Video Model Gen-4 Promises Character Consistency." PetaPixel, April 1, 2025. https://petapixel.com/2025/04/01/runways-new-ai-video-model-gen-4-promises-character-consistency/

[13] "Seedance 1.0: Exploring the Boundaries of Video Generation Models." ByteDance Seed, June 2025. arXiv:2506.09113. https://arxiv.org/abs/2506.09113

[14] "Google releases new AI video model Veo 3.1 in Flow and API." VentureBeat, October 2025. https://venturebeat.com/ai/google-releases-new-ai-video-model-veo-3-1-in-flow-and-api-what-it-means-for

[15] "Kuaishou Ramps Up AI Commercialization as Kling Revenue Hits $150 Million." Caixin Global, March 25, 2026. https://www.caixinglobal.com/2026-03-25/kuaishou-ramps-up-ai-commercialization-as-kling-revenue-hits-150-million-102427380.html

Editing: AI Video Generation