AVisualHistoryofAIFilm&Video

Uber engineer Phillip Wang launches a website that generates a new hyper-realistic human face on every page reload, powered by NVIDIA's StyleGAN. The site is visited 4.2 million times and becomes a cultural moment: the first time the general public confronts AI-generated imagery that is indistinguishable from photography. The concept showed that AI could generate photorealistic faces indistinguishable from real photographs, with no camera, no model, no studio. For creators, it planted a seed: if AI can imagine a face this convincing, what else can it imagine?

Topaz Labs releases Video Enhance AI, the first commercially available AI-powered video upscaling tool. Using deep learning models trained on thousands of video pairs, it can upscale footage up to 8K while adding genuine detail rather than just interpolating pixels. The tool quickly establishes Topaz Labs as the dominant player in AI video enhancement. While not a generative video tool, Video Enhance AI sets a critical foundation: when AI video models arrive years later with their low resolutions and noisy outputs, this is the tool creators will reach for to make that footage usable.

OpenAI reveals DALL-E, a 12-billion-parameter version of GPT-3 trained to generate images from text descriptions. Type 'an armchair in the shape of an avocado' and get a photorealistic image back. It is not released to the public, but the concept is proven: language can direct visual creation. For creators, this is the 'holy grail' moment. The idea that you could describe what you want and have AI create it transitions from science fiction to demonstrated reality. Everything that follows in AI video builds on this foundation.

Artist-mathematician Katherine Crowson (RiversHaveWings) and researcher Ryan Murdock combine two open-source models, VQGAN and CLIP, into a Google Colab notebook that anyone can run. Type a text prompt, get an AI-generated image. But the real breakthrough is animation: by applying zoom, rotation, and translation to each frame and feeding it back through the system, creators discover 'prompt travel', infinite zooms into AI-imagined worlds that shift and morph as the prompt changes. The aesthetic is surreal, psychedelic, unmistakably AI. Twitter accounts sharing VQGAN+CLIP art go viral. For the first time, a global community of artists is creating AI art together, sharing techniques, and pushing the medium forward. This is the birth of the AI art community that will later become the AI filmmaking community.

Somnai releases Disco Diffusion, the first tool to combine diffusion models with CLIP for text-guided image and animation generation. Unlike VQGAN+CLIP's iterative optimization, diffusion models generate images by gradually denoising random noise, producing higher quality results with more coherent structure. Somnai adds 2D animation techniques, and Adam Letts contributes 3D animation. The tool evolves rapidly through six major versions in eight months. Disco Diffusion becomes the training ground for the generation of AI artists who will later define the AI filmmaking movement. Riccardo Fusetti's AIFF 2023 Grand Prix winner 'Generation' will be made with this lineage of tools.

The first open-source pretrained text-to-video model. 9.4 billion parameters generating 4-second clips at low resolution. Blurry, dreamlike, barely coherent... but it works. For the first time, anyone with a GPU could type words and get moving pictures back. The quality was terrible by any standard. But the concept was proven: machines could imagine motion from language. Everything that followed started here.

Meta's research team demonstrates that text-to-image models can be extended to generate video. The technique leverages what AI already learned about images and teaches it how things move. Never released to the public, but the demos circulate widely among researchers and signal that every major tech company is now racing toward AI video. The implications for creators are clear: if this works at scale, anyone will be able to direct scenes with words.

Weeks after Stable Diffusion goes public in August 2022, Deforum launches as a Colab notebook that turns text prompts into animated video. Building on the lineage of Disco Diffusion, VQGAN+CLIP, and PyTTI, Deforum adds Stable Diffusion's dramatically improved image quality to the animation workflow. The signature technique is 'prompt travel': schedule different prompts at different timestamps and watch the AI morph between them, creating dreamlike sequences where one world melts into another. Camera controls let creators zoom, pan, rotate, and fly through AI-generated landscapes. Quickly evolves into an Automatic1111 WebUI extension, making it accessible to anyone who can run Stable Diffusion locally. Deforum animations are projected at concerts worldwide, synced to music. The tool becomes the creative backbone of a generation of AI artists who will later move into AI filmmaking. Many of the aesthetic signatures of early AI video, the morphing transitions, the surreal prompt travels, the infinite zooms, originate here.

Google Brain enters the race with a cascade of video diffusion models producing 1280x768 video. The quality leap over CogVideo is dramatic, but like Meta's work, it stays locked in the research lab. What matters for creators: Google is proving that high-resolution AI video is possible. The gap between 'research demo' and 'tool you can use' is the only thing left to close.

Topaz Labs rebrands Video Enhance AI as Topaz Video AI with a major v3.0 rewrite featuring improved stabilization, faster processing, and new AI models for denoising, upscaling, and frame interpolation. The timing proves prescient: as generative AI video models begin producing short, low-resolution, noisy clips, Topaz Video AI becomes an essential part of the creator pipeline. Filmmakers submitting to early AI film festivals and contests like Gen:48 rely on it to upscale and clean up their AI-generated footage, turning rough model outputs into something presentable on the big screen. For a community pushing tools beyond their limits to make longer-form work, Topaz bridges the gap between what the models can output and what audiences expect to see.

Google demonstrates something no one else has: AI-generated video that lasts over two minutes, driven by a sequence of text prompts that change over time. While other models max out at a few seconds, Phenaki shows that long-form AI video is theoretically possible. For filmmakers watching from the sidelines, this is the first hint that AI might one day generate not just clips, but scenes.

AI video leaves the research lab. Runway Gen-1 is the first commercially available AI video tool, letting creators upload real footage and transform it with AI — applying new styles and visual treatments guided by text or reference images. Not text-to-video yet, but the moment AI enters the creative workflow. Thousands of filmmakers, designers, and artists get their hands on it immediately.

A quick follow-up to Gen-1, Gen-2 improves upon the initial video-to-video capabilities with higher-fidelity video generation — and makes the leap to text-to-video. Type a prompt, get a 4-second video clip. The quality is rough, the clips are short, and the physics are wrong. None of that matters. What matters is that for the first time in history, anyone with a browser can generate original video from nothing but words. Runway Gen-2 creates the AI filmmaking community overnight. The experiments begin.

The Copyright Office formally opens a multi-year initiative to study the copyright implications of generative AI — registration of AI-assisted works, training data, digital replicas, and authorship. The Office issues a registration guidance the same day clarifying that material produced by generative AI is not on its own copyrightable. This becomes the legal spine that every later U.S. policy moment — Part 1 on digital replicas, Part 2 on copyrightability — hangs from, and the reference point every AI filmmaker eventually has to read.

Marvel's Disney+ series premieres with an opening title sequence made with generative AI, produced by Method Studios. Director Ali Selim defends the choice as thematically fitting for a show about shape-shifting alien Skrulls. The backlash is immediate and loud — concept artists, illustrators, and union members frame it as a high-profile studio normalizing AI work that displaced human jobs. For mainstream audiences, this is the moment generative imagery walks into a prestige franchise and forces a conversation that until then had largely been confined to art-Twitter.

The 148-day WGA strike ends with a tentative agreement whose AI provisions become Hollywood's first enforceable guardrails: AI can't write or rewrite literary material, AI output is not 'source material,' studios must disclose if any material handed to a writer was AI-generated, and writers cannot be required to use AI tools. The contract is ratified weeks later and immediately reframes the AI-and-creativity conversation from speculation to negotiated industry rule. Every studio AI experiment after this is shaped by what writers were and were not willing to live with.

Open-source video generation arrives. Stability AI releases models that turn a single image into a 2-4 second animated clip, and anyone can download and run them locally. For creators who want control over their tools, who want to fine-tune and customize, who want to run AI video without a subscription, this is the starting gun. The open-source AI video ecosystem begins here.

Javi López & Emilio Nicolás Riquelme launch Magnific AI, the first popular 'creative' image upscaler — a tool that doesn't just add resolution but reimagines and hallucinates new details into images. With sliders for 'creativity,' 'HDR,' and 'resemblance,' creators can control how much the AI invents, turning rough Midjourney or Stable Diffusion outputs at the time, into images with stunning detail and texture. Over 30,000 people sign up within the first 24 hours. For the AI filmmaking community, where image-to-video workflows with models like Gen-2 depend highly on the quality of the input image, Magnific feels like magic. It fills in the details that early image models miss — skin pores, fabric texture, environmental detail — transforming good images into extraordinary ones. The impact is immediate and the approach is quickly emulated by competitors. Magnific is later acquired by Freepik in May 2024.

Pika Labs — founded by Stanford PhD students Demi Guo and Chenlin Meng — opens early access to Pika 1.0, pairing text-, image-, and video-to-video generation with a chat-style web interface that anyone could use without a research-lab login. The launch arrives alongside a $55M funding round and quickly becomes the consumer face of AI video for late-2023, sitting next to Runway in nearly every creator's toolbox. For the AI filmmaking community, Pika was the first product that made browser-native generative video feel approachable.

Members ratify the 2023 TV/Theatrical contract, locking in the first union-negotiated framework for digital replicas of performers — covering consent, compensation, scope of use, and synthetic-performer disclosure. Together with the WGA's September deal, the contract turns abstract performer-likeness anxiety into bargained law and reshapes how studios can legally use AI-generated faces and voices. After this, every studio AI pilot that involves a performer's image has to answer to language in this contract.

OpenAI drops a preview that changes the conversation overnight. 60-second videos with coherent physics, consistent characters, and cinematic camera movements. Nothing publicly available comes close. The internet is split between awe and existential dread. For creators, the preview is both inspiring and frustrating: it proves AI video can be genuinely cinematic, but almost no one can actually use it. The gap between 'demo' and 'product' becomes the story.

YouTube rolls out a Creator Studio tool that requires uploaders to label realistic content that has been meaningfully altered or generated — including generative AI footage, voice cloning, and digitally altered faces of real people. Obviously fantastical and production-assistance use is exempt. For AI filmmakers, this is the moment the distribution layer starts formalizing what audiences see when an AI clip lands in their feed, and the first time a major platform turns disclosure into a workflow step.

The Cairnes brothers' '70s-talk-show horror film opens to glowing reviews — and a controversy that quickly eclipses them. The directors confirm that three brief interstitial graphics (a network bumper, a pumpkin, a skeleton) were generated with AI. Even though the live action and performances are entirely human, the disclosure drives a days-long online debate about where the line sits and how loudly it should be drawn. It becomes the canonical case study for how a tiny amount of generative imagery can become the loudest fact about a release.

OpenAI joins the steering committee of the Coalition for Content Provenance and Authenticity, the cross-industry group that maintains the Content Credentials standard for cryptographically signed metadata on media. Combined with announced plans to embed Content Credentials in Sora and DALL·E output, this is the moment provenance stops being an aspiration in white papers and becomes a default piece of the frontier-model stack. For AI filmmakers and the platforms distributing their work, it's the first concrete answer to "how do we tell where this came from."

At Google I/O, DeepMind extends SynthID — its imperceptible watermark for AI-generated content — into video, embedding a signal across every frame of Veo output that downstream tools can detect even after editing, compression, or re-uploads. Where C2PA attaches signed metadata as a label, SynthID hides the signal inside the pixels themselves. Together they sketch the two-track provenance future every AI-filmmaking tool will eventually have to support.

Google officially enters the consumer AI video race at I/O 2024, announcing a model that generates 1080p video over a minute long. What this meant in practice: full HD video at lengths that start to feel like actual filmmaking, not just clips. Google's massive compute infrastructure means they can offer quality that smaller companies struggle to match.

Luma Labs launches Dream Machine and 25 million users sign up. The appeal is simplicity: type a prompt or upload an image, get video back fast. For many creators, Dream Machine is their first experience with AI video generation. It becomes the gateway that brings filmmakers, animators, and content creators into the AI video world. The sheer scale of adoption proves the demand is real.

China's Kuaishou launches Kling, and within weeks it goes global. The model produces surprisingly good 1080p video that competes with Runway and Pika from day one. Kling represents something important: real competition in AI video. More models means more options, lower prices, and faster innovation. Kling will go on to iterate faster than any other model family in the space.

A major quality leap. Gen-3 Alpha produces 10-second clips with dramatically improved fidelity, motion coherence, and understanding of 3D space. For the first time, filmmakers start treating AI video as a real production tool rather than a novelty. The jump from Gen-2 to Gen-3 Alpha is the difference between 'interesting experiment' and 'I can actually use this in my work.' The era of AI filmmaking as craft begins.

Swiss director Peter Luisi releases a feature whose screenplay was written entirely by ChatGPT, billed as the first AI-written feature film. After backlash, the planned London premiere is cancelled and the film moves online for free. The cultural impact dwarfs the production: for a wide audience this is the moment the screenwriting fears that drove the WGA strike stop being hypothetical and become a poster, a trailer, and a play button.

The Copyright Office publishes the first part of its long-promised AI report, focused entirely on digital replicas — unauthorized AI re-creations of a real person's face, voice, or likeness. The report recommends new federal legislation: it concludes existing state laws and federal patchwork aren't enough to cover the harms generative tools now make trivial. It is the document the SAG-AFTRA contract spent the prior year arguing about, written in legal prose, and it sets the agenda for every subsequent likeness-rights bill.

The European Union's AI Act formally enters into force, with a staggered application schedule running through 2026. For generative video, the key piece is Article 50: providers and deployers of synthetic media — text, image, audio, or video — must mark output as artificially generated, and deepfakes must be labeled. It is the first comprehensive cross-border regulatory framework that any AI filmmaker distributing into Europe has to plan around, and the model other regions immediately start comparing themselves to.

MiniMax quietly launches its first video model. 720p, 25fps, up to 6 seconds. The specs look modest on paper, but the motion quality and realism surprise everyone. Hailuo represents the emergence of a new player that will quickly become known for photorealistic output. The Cat Olympics viral moment a few months later will prove that Hailuo's physics simulation is genuinely convincing.

Kuaishou introduces Motion Brush, a feature that lets creators paint directional motion onto specific parts of the frame. Want a character's hand to wave while the background stays still? Draw the motion path. For creators, this is a breakthrough in creative control: you are no longer just typing prompts and hoping, you are directing the motion. Camera movement controls follow, making Kling feel less like a slot machine and more like a tool.

Meta unveils Movie Gen — a frontier model that generates HD video with synchronized audio and can edit existing footage from natural-language instructions. The release pushes the field past silent shorts into synchronized media and production-style editing language, and within two weeks Meta partners with Blumhouse and filmmakers including Casey Affleck and Aneesh Chaganty to test it. Movie Gen never ships as a consumer product, but it sets the bar competitors have to clear and quietly reframes the race around audio and editability, not just length.

Adobe launches its Firefly Video model at Adobe MAX, entering the AI video race with a focus on commercial safety and professional workflows. Text-to-video with camera controls, image-to-video transformation, and Generative Extend for Premiere Pro that fills footage gaps and smooths transitions. Supports 100+ languages. The first AI video model explicitly designed for commercial use from day one, built on content that Adobe has the rights to use for training.

Robert Zemeckis reunites Tom Hanks and Robin Wright in a Forrest Gump-style decades-spanning drama, using Metaphysic's generative AI face-transformation system to age and de-age the leads on-set in real time. This isn't a viral demo or a creator short — it's a wide theatrical release from a major studio using generative AI directly in the finished image, with the work supervised inside a traditional VFX pipeline. After this, 'AI is in production' stops meaning experimental shorts and starts meaning marquee faces in front of cameras.

The first VC-backed studio built specifically for generative AI filmmaking. Co-founded by Dave Clark (Battalion, Borrowing Time), Fullscreen CEO George Strompolos, and former YouTube exec Jamie Byrne. Funded by Peter Chernin's The North Road Company and Andreessen Horowitz (a16z). Developing projects including 'NinjaPunk' and 'Extinction.' Building proprietary production software called MUSE. A signal that Hollywood money is betting on AI-native storytelling.

Tencent releases HunyuanVideo's inference code and weights — a 13-billion-parameter video model that's the largest open-source release of its kind to date. Anyone can download it, run it locally, fine-tune it, and ship workflows that don't depend on a subscription or a closed API. For researchers, ComfyUI builders, and creators wary of platform risk after seeing what happens when models get sunset, this immediately becomes the open-source anchor of the AI video stack.

Resolution jumps from 1080p to 4K, and the physics simulation takes a serious step forward. Released via VideoFX and later the Gemini app. Veo 2 becomes the model behind some of the most impressive creator work of the year, including Jason Zada's 'The Heist' (10M+ views) and Henry Daubrez's 'Kitsune.' Veo 2 is the first model where the output genuinely looks cinematic at feature-film resolution.

Ten months after the preview that broke the internet, Sora finally launches for ChatGPT Plus and Pro users. The hype has cooled considerably. Competitors like Veo 2, Gen-3 Alpha, and Kling have closed the quality gap. The lesson is sobering: in AI, a year is a lifetime. The model that seemed impossibly ahead in February is merely competitive by December.

Kuaishou claims a 195% overall improvement. Standard and Pro quality tiers give creators a choice between speed and fidelity. Better prompt adherence means the model actually does what you ask more often. More natural facial expressions and improved temporal consistency in sports and action scenes. For creators working on narrative projects, the reliability improvements matter more than any single feature.

Built on an entirely new multi-modal architecture with 10x the compute of the original Dream Machine. Ultra-realistic details, coherent fast motion, and logical event sequences. For creators who built workflows around Dream Machine, Ray2 delivers a generational upgrade without changing the interface. The quality jump is dramatic enough that side-by-side comparisons circulate widely on X.

Mid-January reporting reveals that two of the season's biggest awards contenders used Respeecher to refine performance audio: The Brutalist used it to sharpen Adrien Brody and Felicity Jones's Hungarian pronunciation, and Emilia Pérez used it to extend Karla Sofía Gascón's singing voice. Both films stay in the race, but the debate dominates Oscar coverage for weeks and expands the AI-in-film conversation from visuals into the most sensitive territory of all: the actor's own performance. After this, every awards-season disclosure form asks new questions.

Six months after the digital-replicas report, the Copyright Office publishes Part 2: the answer to the question every AI filmmaker has been asking — can the output of a generative model be copyrighted at all? The conclusion: AI-assisted work can be protected to the extent of the human contribution (curation, arrangement, edited human-authored material), but purely AI-generated material is not, and prompting alone doesn't qualify as authorship. For festivals, distributors, and rights desks, this is the legal handle they've been waiting for to evaluate which AI-assisted submissions they can actually license.

Promise, the a16z-backed generative AI studio founded by Dave Clark, George Strompolos, and Jamie Byrne, acquires Curious Refuge, the leading AI filmmaking school. Founded in 2020 by Caleb and Shelby Ward, Curious Refuge has educated over 4,000 people from 171 countries through courses in AI filmmaking, animation, advertising, and documentary. The acquisition creates a direct pipeline from education to production: Curious Refuge creators can now work on Promise film and TV projects. For the AI filmmaking community, this is a signal that the ecosystem is maturing. The school that taught thousands of people to use AI tools is now part of the studio that will put those skills to work.

Alibaba open-sources a 14-billion-parameter video model under Apache 2.0, and it immediately tops the VBench leaderboard. For the first time, a state-of-the-art video model can run locally on consumer hardware with just 8GB of VRAM. The ComfyUI community adopts it overnight. For independent filmmakers who want to run their own pipeline without subscriptions or API costs, this changes everything. The democratization of AI video takes its biggest step forward.

Amazon's biblical epic House of David, created by Jon Erwin, premieres on Prime Video with 72 generative AI-assisted shots in Season 1 — one of the first major streaming series to openly integrate AI into its VFX pipeline. The standout is Episode 6's Goliath origin sequence: photoreal digital characters, armies of 10,000+, rain, smoke, and angel wing feathers — all driven by generative AI in collaboration with Vū Technologies. What would have taken four or five months of traditional post-production was completed in weeks. Erwin's approach: never generate from scratch, always augment existing show assets. The tools — Runway, Kling, Midjourney, Magnific, and Topaz alongside After Effects — handled physics simulations better than conventional VFX software at a fraction of the cost.

The breakthrough feature is single-image character consistency: upload a reference photo and the model maintains that character's appearance across generated clips. For narrative filmmakers, this solves one of the biggest pain points in AI video. Before Gen-4, every generation was a roll of the dice on whether your character would look the same. Now you can build stories with recurring characters.

A major quality leap with Multi-Elements Editor, 60+ style presets, and cinematic lighting controls. The creative control is the real story: you can now specify multiple elements in a scene and control how each one behaves. The style presets give you instant access to looks that previously required post-production. Kling is becoming a director's tool, not just a generator.

Alibaba combines video generation, reference-based generation, repainting, area modification, and spatio-temporal extension into a single open-source model. For creators using ComfyUI workflows, VACE turns Wan from a generator into a full production tool. Edit, extend, repaint, and modify shots without leaving the pipeline. One model that does the work of five separate tools.

The landmark addition: native audio generation. For the first time, a video model generates synchronized dialogue, sound effects, and ambient noise alongside the visuals in a single pass. Lip-sync accuracy under 120ms. In practice, this collapses what was previously a multi-tool workflow into a single generation. No more generating video, then finding sound effects, then syncing dialogue. The model imagines the entire audiovisual scene at once.

Introduces tiered quality modes: Standard for fast iteration, High Quality for final output, and a Master Edition for premium results. The tiered approach is practical: draft in Standard mode to find the right prompt and composition, then re-generate in Master for the final shot. Faster iteration without sacrificing quality when it counts.

Disney and Universal file a joint federal copyright lawsuit against Midjourney, calling the service a 'bottomless pit of plagiarism' — the first time major Hollywood studios sue a generative AI company directly. The complaint includes side-by-side exhibits of Midjourney generations of Darth Vader, Yoda, the Minions, and Shrek alongside the originals. The case shifts the AI-and-rights conversation from individual artists' anxieties into a marquee studio-vs-platform fight with the kind of legal firepower that actually moves the field.

Director Eliza McNitt premieres ANCESTRA at Tribeca — a hybrid live-action short produced with Darren Aronofsky's Primordial Soup and Google DeepMind, using Veo, Gemini, and Imagen to generate the prenatal and surgical imagery the story needs but could never have been shot. The film is autobiographical, and the AI sequences exist not to replace cinematography but to render images of the unfilmable. It becomes one of the cleanest reference cases for hybrid live-action + generative-AI filmmaking on a real festival platform.

ByteDance enters the AI video race with native bilingual support (Chinese and English), unified text-to-video and image-to-video, and multi-shot narrative generation with subject and style consistency. For creators, the multi-shot consistency is the standout feature: generate multiple shots of the same character in the same visual style, making it practical to build sequences that feel like they belong together.

Native 1080p generation, 10-second clips, and state-of-the-art physics simulation. MiniMax's second model climbs to #2 on global benchmarks, surpassing Veo 3 on several metrics. Built with 3x more parameters and 4x more training data than Video-01. The leap in realism is immediately visible. Gymnastics, water physics, fabric movement: the things that used to look obviously fake now look convincing.

Moonvalley moves Marey into general availability and positions it as the first AI video model built entirely on licensed training data — fully cleared for professional production use. Alongside the model is a deliberate set of production controls: a Voyager motion-and-camera director, character embedding, and shot-level continuity tools. The category statement matters as much as the product: with studios suing model vendors and risk-averse buyers refusing to touch scraped-data outputs, 'commercially safe' becomes a real positioning lane that other vendors now have to answer to.

The first open-source video model with Mixture-of-Experts architecture. 27 billion total parameters but only 14 billion active per step, cutting compute by 50%. Cinematic-grade quality with precise control over lighting, color tone, camera angle, composition, and focal length. For creators, Wan 2.2 becomes the community favorite: it balances quality, versatility, and cost better than anything else you can run locally.

Framer launches Cartoon Hero 1.0, a hands-on course community teaching workflows for creating AI-animated shorts. Students build cartoons from start to finish in multiple styles — sitcom, anime — learning tools and techniques for consistent character creation, multi-shot scenes, dialogue, lip-sync, and sound design. Impactful for demystifying AI animation workflows and proving that structured process, not just raw model output, is what makes AI cartoons feel professional.

xAI releases Grok Imagine, offering text-to-image and image-to-video generation with sound. It produces short video clips from images — similar to Midjourney's video tool but lacking text-to-video capabilities. Elon Musk called it a "meme motherlode."

Google's image generation model takes off under the codename 'Nano Banana.' Within two weeks, 23 million new users generate 500 million images. For AI filmmakers, this is the missing puzzle piece: a fast, high-quality image generator that maintains character consistency across outputs. Generate your character once, lock in the look, then feed the image to any video model as a starting frame. The image-to-video workflow that defines professional AI filmmaking finally has its source-of-truth tool.

OpenAI's second attempt, released to select users in the US and Canada. Improved quality and longer outputs, but the competitive landscape has shifted dramatically since the original preview. Sora 2 is a solid tool that arrives in a crowded market. The window of Sora dominance has passed, but the model finds its audience among creators already embedded in the OpenAI ecosystem.

10 filmmakers given unlimited access to Google Flow plus mentorship. Henry Daubrez becomes Google Labs' first Resident Filmmaker. Google bets on individual creators, not just studios. A new model for AI film patronage.

Three months after the Midjourney filing, the major studios go after MiniMax — alleging that Hailuo AI generates downloadable images and videos of protected characters and markets itself as a 'Hollywood studio in your pocket.' Unlike the Midjourney complaint, this one is squarely aimed at a video-native consumer product. The legal pressure on AI video moves from image-only systems into the generation tools many filmmakers were actively using, and risk-averse studios start re-reading their pilots' tool stacks.

The world's first 'reasoning' video model. More than 2x the size of Ray2, it generates native 16-bit HDR video in ACES EXR format, bringing AI output directly into professional color grading pipelines. For creators working in professional post-production, this is a milestone: AI-generated footage that slots into existing studio workflows without quality compromises. The HDR output means AI video can finally sit alongside camera footage without looking out of place.

Enhanced physics simulation, multi-step instruction handling, and causal relationships between elements in the scene. Nearly 30% lower pricing. What this meant for filmmakers: you can now describe complex sequences ('the ball rolls off the table, bounces twice, and the dog chases it') and the model understands cause and effect. Pricing drops make iteration affordable for independent filmmakers.

Enhanced audio generation, improved realism, and powerful new editing features via Google Flow. Native 1080p HD with videos up to 60 seconds. The Flow integration is transformative: it creates a complete production environment around Veo, with editing, trimming, and refinement tools that turn raw generations into finished shots. The gap between 'AI generated clip' and 'production-ready shot' narrows significantly.

Enhanced body movement, facial expressions, and physical realism. Introduces a Fast variant for batch creation at 50% lower cost. For creators working on volume projects like social content or storyboarding, the Fast variant means you can generate dozens of variations quickly and cheaply. The quality of the standard model continues to push toward photorealism, particularly in human movement and expression.

Oscar-winning filmmakers Michael Govier and Will McCormack (If Anything Happens I Love You) announce All Heart, a nine-minute animated short produced with Natasha Lyonne's AI studio Asteria. The film tells the story of a father visiting a young man with an extraordinary connection to his late daughter. The production combines hand-drawn artistry by illustrator Jimmy Thompson with AI animation: Asteria trained a custom model on Thompson's original art to generate additional production assets in his style, then integrated it into a traditional keyframing and in-betweening pipeline. The approach sidesteps the copyright debate entirely — the AI learns only from the creators' own work. The short qualifies for 2026 Academy Award consideration in Animated Short Film, making it one of the first AI-assisted films to enter the Oscar race.

Best-in-class text rendering in images, multilingual text support, and integration with Google's world knowledge for accurate maps, diagrams, and infographics. Blends up to 14 images while maintaining consistency across 5 characters. For AI filmmakers, this becomes the 'virtual cinematographer': design your entire shot in a single image with precise character placement, lighting, and composition, then feed it to any video model.

Native simultaneous audio-visual generation arrives at Kling: dialogue, singing, physics sound effects, environmental audio, and mixed mode. Following Veo 3's lead, Kling proves that audiovisual generation is becoming table stakes. The implications were immediate: fewer tools, faster iteration, and better synchronization between what you see and what you hear.

Simultaneous audio-video generation with millisecond-precision lip-sync across 8+ languages and dialects. Cinema-grade camera controls including Hitchcock zoom, orbit, and tilt. For creators working in multilingual markets or dubbing, the language breadth is unmatched. The camera controls give directors vocabulary they understand: specify the exact camera movement you want, and the model executes it.

Runway's most advanced model delivers unprecedented visual fidelity and creative control. The culmination of the Gen-1 through Gen-4.5 evolution. For creators who have grown with Runway since Gen-1, this release represents how far the technology has come: from simple style transfer to photorealistic video generation with character consistency, all within three years.

China's first open-source reference-to-video model. Upload a character reference video with your appearance and voice, then generate new scenes starring yourself. This opened up open-source character consistency with audio sync. Anyone can now create long-form narratives with consistent characters using free, locally-run models. The combination of visual and vocal consistency opens the door to serialized AI content without subscription costs.

A specialized variant for hybrid AI workflows. Keyframe control lets you set start and end frames, Character Reference preservation maintains identity across edits, and video modification works while keeping consistency. For creators, this shifts Luma from pure generation to an editing paradigm: take existing footage, refine specific aspects, and maintain coherence throughout.

Runway expands the scope of its annual AI Festival, adding categories beyond film. A signal that AI-generated media is spilling into games, interactive experiences, and new formats yet to be named.

Native 1080p generation at 4x faster speed and 3x lower cost. This eliminated the quality-speed-cost tradeoff that defined earlier models. You no longer have to choose between fast drafts and quality finals. Every generation is both fast and good, at a price point that makes iteration painless.

xAI's "biggest leap yet" — Grok Imagine 1.0 unlocks 10-second text-to-video and image-to-video generation at 720p with dramatically improved audio. Instruction-following capabilities allow restyling scenes, adding/removing objects, and motion control. The API launched days earlier on January 28. By this point, Imagine had generated over 1.2 billion videos in a single month.

xAI runs an ad contest challenging creators to make commercials using only Grok Imagine 1.0, offering $1M for the grand prize. Over 4,000 submissions pour in. Andy Orsow wins $1M for "Understand the Universe. Faster." — a witty reimagining of Galileo's inquisition as a query to Grok. The winning ad airs during Super Bowl placements on X. Runner-up Dave Clark ($500K) and PJ Ace ($250K) round out the top three, with two $100K honorable mentions. A landmark moment for AI video contests in terms of prize money and mainstream visibility.

ByteDance's unified multimodal architecture accepts text, image, audio, and video as inputs. Up to 15-second clips at 2K resolution with extreme character consistency and audio-visual beat matching. For creators, the multimodal input is the breakthrough: feed the model a reference image, a voice clip, and a text description simultaneously, and it synthesizes all three into coherent video. Then copyright disputes from major studios pause the global rollout, raising questions about the training data behind the magic.

Unified omni-modal architecture spanning video, image, and beyond. From Kling 1.0 in June 2024 to 3.0 in February 2026, Kuaishou shipped 9 major versions in 20 months, the fastest model evolution in AI video history. For creators, Kling's trajectory proves that the best model today will be obsolete in months. The pace of improvement rewards creators who build adaptable workflows rather than betting on any single tool.

Magnific brings the magic of its image upscaler to video. The same creative upscaling approach that transformed still images — adding details, enhancing texture, and adding fidelity beyond what the source contains — now works on moving pictures. For AI filmmakers who have relied on Magnific for their image-to-video pipelines since 2023, this is the natural next step: upscale the final output with the same creative intelligence that made the input frames shine. Precision mode was added in March 2026, focusing on preserving the original footage, while still allowing video enhancement options like 4K upscaling, sharpness, grain control, and FPS boost

Pro-quality image generation at Flash speed, ranked #1 on the Artificial Analysis Image Arena at launch, at half the price of Nano Banana Pro. Maintains character consistency across 5 characters and 14 objects per workflow. For AI filmmakers, this completes the image-to-video production pipeline: generate consistent starting frames for every shot in your film, fast and cheap enough to iterate freely. Combined with the latest video models, like Seedance 2.0 or Kling 3.0, it enables cohesive visual storytelling.

OpenAI announces that the Sora app and API are being wound down, 25 months after the preview that broke the internet. The help center confirms the schedule: the consumer app and web experience will end on April 26, and the API on September 24. From the most hyped AI video model in history to a published end-of-life date in a single post. For creators, the lesson is stark: platforms come and go, but your creative vision and storytelling skills transfer to whatever tool comes next.

The day OpenAI named in its March announcement arrives, and the Sora app and web experience go dark. The API continues for one more development cycle and is scheduled to end September 24. The actual sunset clarifies something the launch hype obscured: even frontier AI products with cultural moments of their own can disappear inside two years. Every creator who built a workflow around a closed model now has a fresh case study for why portability — and open-source backups — matter.

Alibaba Cloud opens limited beta access to HappyHorse 1.0 through Model Studio, with native 1080p output, synchronized audio, and capabilities pitched directly at the frontier video tier currently occupied by Seedance 2.0 and Veo 3.1. The day-after-Sora-shutdown timing isn't subtle: while one major player exits, another arrives and immediately becomes the new comparison point in every creator's frontier-model lineup.

Krea launches Krea 2, a new foundation model built completely from scratch for aesthetic diversity and stylistic control — some say it has Midjourney-level aesthetic with Nano Banana-level control. The system can extract the stylistic components of any image and transfer them to your results. Moodboards reference a wider range of visuals and analyze the ideas and concepts behind the images you upload, not just style. Taste Profiles on top of moodboards, plus text and style guidelines, open up a degree of personalized customization that's been missing from the image-model tier. A real shot at the Midjourney crown.

ViggleAI launches PINOC, a model that turns a motion video into a clean skeleton animation exported as .fbx or .glb — drop-in ready for Blender, Maya, or whatever your 3D pipeline is. PINOC also takes a single character image and produces a 3D Gaussian splatting model you can preview the motion against. It's a meaningful bridge between the AI video workflow most creators already have and the traditional 3D animation pipeline, collapsing what used to be days of mocap cleanup into an upload.

TrueShort, founded by Nate Tepper, announces a $12 million funding round led by Keith Rabois at Khosla Ventures, with participation from Jeffrey Katzenberg of Dreamworks, Ravi Nandan of A24, Scott Belsky of Behance, General Catalyst, and LookingGlassCap Fund II. The company plans to hire filmmakers to build small teams for AI-powered film production. Tepper's framing — "We believe a kid in his bedroom is going to make the next Star Wars" — captures the bet under the round: that the next blockbuster studio looks less like a studio and more like a kid with a laptop and an AI stack, and that funding the filmmakers directly is the right way to get there.

Announced at AI on the Lot at Culver Studios, the GenAI Creators' Fund is a joint Amazon MGM Studios and AWS initiative giving creators access to professional-grade AI tools and funding — grants for proofs-of-concept and shorts aimed at digital creators who built massive audiences but never had access to studio-grade production. Alongside the fund, Amazon revealed its first three greenlit animated series, all bound for Prime Video: "Cupcake & Friends" from BuzzFeed Studios, a cupcake and her friends navigating the chaos of a sleepover; "Love, Diana Music Hunters" from Albie Hecht at pocket.watch (former Nickelodeon entertainment president and a SpongeBob developer), a band of K-pop space-traveling musicians racing to Planet Goo; and "Punky Duck" from Emmy-winning "The Book of Life" and "Maya and the Three" creator Jorge R. Gutierrez, a punk duck and his cat tearing through an exaggerated Los Angeles. Powering the work is Project Nara, Amazon MGM Studios' purpose-built AI production platform on AWS — a model-agnostic, collaborative workspace that routes each task to the optimal AI model and integrates with the tools creators already use (Maya, Blender, Nuke, Unreal Engine, Adobe Suite), with provenance tracking for IP protection. The framing throughout: "human creativity leads, and AI supports." It marks the first major Hollywood studio to formally fund and greenlight generative-AI productions, alongside the AI-curious creators willing to work this way.