In 2025, AI video was mostly a case of "look what I made." In 2026, it is firmly embedded in client pitches, previsualisation, marketing content, and training assets. Working filmmakers are founding dedicated AI companies to solve highly specific production problems. Netflix is acquiring them. Amazon is building physical studios around them.
The core question has fundamentally shifted. It is no longer "should we use AI?" Instead, professionals are asking: has this category matured enough to belong in a serious, professional workflow?
AI is rapidly becoming infrastructure. It is no longer a tool you pick up and put down, but a fundamental layer of the pipeline that is already there when you arrive. While the individual models will keep changing, this complete workflow transformation is what will stick.
Here is exactly where the landscape stands today, as of June 2026.
Industry Timeline: The Shift to Infrastructure
Sora 2 Disrupts the Market
OpenAI launches Sora 2, introducing synchronised audio, multi-shot sequences, and the Cameos feature for opt-in digital likenesses. It marks the high-water mark for pure-play synthetic video applications.
Veo 3.1 & Granular Control
Google DeepMind releases Veo 3.1 with native audio, dual keyframe control, reference-image-driven generation, and native vertical video support.
Music Industry Collisions
Suno raises $250 million at a $2.45 billion valuation. Warner Music reaches a settlement with the platform, while Sony continues litigation ahead of a landmark summer 2026 fair-use ruling.
The Disney–OpenAI Deal That Wasn't
Disney announces a highly publicised $1 billion licensing partnership with OpenAI to bring legacy characters to Sora. As detailed below, this arrangement would unravel entirely by spring 2026 without money changing hands.
The Next-Gen Model Wave
Google DeepMind's Gemini 3 Pro Image (codename Nano Banana Pro) tops the LMArena leaderboard with native 4K output. Runway upgrades to Gen-4.5, and Kuaishou's Kling reaches version 3.0, introducing native multi-shot storyboard control.
The Sora Shockwave & Studio Consolidation
On 24 March, OpenAI announces the complete discontinuation of the Sora brand, app, and API — citing unsustainable compute costs estimated at $15 million per day and a strategic pivot toward enterprise software ahead of a late-2026 IPO. The app permanently shuts on 26 April. The Disney partnership is entirely dissolved, with no money having changed hands. Concurrently, Netflix confirms the acquisition of InterPositive — Ben Affleck's 16-person AI post-production startup focused on technical workflow fixes. On 20 April, iQIYI launches Nadou Pro in China, a consenting-likeness database allowing over 100 celebrities to license their digital counterparts for short-form AI dramas, sparking widespread creative union backlash.
The Studio Lot Integration
At the AI on the Lot summit at Culver Studios, Amazon MGM and AWS launch the GenAI Creators' Fund, greenlighting three AI-assisted animated series for Prime Video powered by Project Nara — a collaborative production environment combining proprietary studio IP with third-party models including Kling 3.0. Google I/O introduces Gemini Omni for conversational, iterative video editing, and embeds SynthID and C2PA verification directly into Chrome and Search. OpenAI and ElevenLabs join the C2PA steering committee shortly after.
The LED Volume Backlash: AI vs. Physical Infrastructure
One of the most quietly significant shifts of 2026 is the cooling of Hollywood's obsession with massive LED volumes. The multi-million-pound virtual production stages pioneered by The Mandalorian are facing an existential challenge — and AI combined with 3D Gaussian Splatting (3DGS) is precisely what is deflating the hype.
The primary argument for physical LED walls was always ambient light spill: if an actor stands in front of a digital sunset, the orange light naturally reflects off their skin. But the operational reality of these stages has hit a wall, allowing pure AI workflows and 3DGS to mount a structural takeover based on three core shifts:
The collapsing "fix it in pre" model. LED volumes require an immense Virtual Art Department (VAD) to build and optimise photorealistic 3D worlds in game engines before a single frame is shot. If an environment is not ready, a £200,000-a-day production stalls. AI and 3DGS shift the paradigm back to post-production — turning months of manual 3D world-building into a matter of hours spent scanning real-world locations with a smartphone or drone.
Real-time spatial efficiency. Gaussian Splats represent environments as millions of overlapping light coordinates rather than heavy polygonal meshes, rendering at ultra-high frame rates — 100+ FPS. This makes them exceptionally cheap to run on hardware, allowing independent productions to project high-fidelity digital backdrops onto affordable laser-projector screens rather than renting expensive LED tiles.
Neural re-lighting kills the green-screen flaw. The ultimate technical justification for LED screens has evaporated. Emerging neural re-lighting and matting frameworks can now take an actor shot against a basic green screen, extract them with flawless edge fidelity, and programmatically map the exact shadows, highlights, and colour reflections of a generative or GS-scanned backdrop onto their skin in post.
The industry is pivoting away from sprawling, bespoke physical stages towards micro-volumes — combining small, affordable projection setups for tight interiors with AI spatial pipelines that handle wide-angle tracking and environmental extensions at a fraction of the budget. Critically, if the show gets cancelled, you have not sunk millions into a specialised, single-use physical soundstage. The financial risk profile of virtual production has changed fundamentally.
AI Model & Infrastructure Reference
Studio Platforms & Acquisitions
Netflix / InterPositive
A specialised 16-person post-production unit acquired by Netflix. Rather than generating content from scratch, their technology focuses on technical fixes: generating missing pickup shots, automated lighting adjustments, and background correction. Positioned strictly as creative support.
Amazon Project Nara — GenAI Creators' Fund
Unveiled at the AI on the Lot summit at Culver Studios, May 2026. Built on AWS, Project Nara is an end-to-end collaborative workspace routing tasks across a model-agnostic architecture. It integrates industry-standard tools (Maya, Blender, Nuke) with custom video models trained on Amazon MGM's historical catalogue to solve complex cinematic consistency challenges.
Video Generation
| Model | Developer | Core Strengths | Pipeline Notes |
|---|---|---|---|
| Kling 3.0 | Kuaishou | Native multi-shot storyboard generation (up to 6 camera cuts); element reference voice binding; exceptional motion physics. | Heavily adopted by indie production houses for narrative consistency. |
| Veo 3.1 | Google DeepMind | Native audio-video generation up to 4K. Ingredients-to-video matching using up to three reference images. Scene Extension API supports coherent sequences up to 148 seconds. Native 9:16 vertical support. | Enterprise standard for marketing, commercial mockups, and portrait content. |
| Seedance 2.0 | ByteDance | Multi-input grid control — 9 images + 3 clips + 3 audio files per prompt. Highly stable 15-second narrative sequences. Strong cinematic camera movement. | Excellent for complex, multi-subject interactive scenes. |
| HappyHorse-1.0 | Alibaba ATH | Joint audio-video pipeline with native lip-syncing across seven languages. | Currently leads the Artificial Analysis Video Arena leaderboard. |
| Runway Gen-4.5 | Runway | SOTA world consistency models; hyper-granular cinematic camera motion controls. | Established baseline benchmark for independent narrative filmmakers. |
| Sora 2 — Legacy | OpenAI | Multi-shot sequences, flawless physics, Cameos feature for digital likeness insertion. | Discontinued. App closed 26 April 2026. API phasing out September 2026. |
Image Generation & Concept Art
Gemini 3 Pro Image (Nano Banana Pro) — Google DeepMind
Available in Flash and Pro variants. The Pro tier offers native 4K output and commands the top spot on the LMArena image quality leaderboard.
Midjourney v7 / v8
The undisputed leader for raw aesthetic coherence, world-building, and highly detailed theatrical concept art.
Flux 2 — Black Forest Labs
Renowned for photorealism and rapid inference speeds. The foundational backbone for open-source, local ComfyUI pipelines.
Ideogram v3
The strongest model in its class for typographic accuracy — rendering crisp, complex text layout directly within generated images.
Adobe Firefly
The enterprise standard for commercially safe generation, trained exclusively on licensed or public-domain content to guarantee legal safety for corporate workflows.
Audio, Voice & Localisation
ElevenLabs
The industry leader in multilingual text-to-speech and voice cloning. Notable for high-profile estate-consented legacy projects, including the official recreation of Stan Lee's voice. Fully integrates SynthID watermarking.
Suno
A dominant force in full-song generation, with roughly 2 million paid subscribers and $300 million ARR. Positioned to reshape commercial music licensing following its Warner settlement, with Sony litigation and a landmark fair-use ruling still pending in summer 2026.
HeyGen
Built for enterprise scale — corporate AI avatars, automated video translation, and high-fidelity video lip-syncing for global localisation.
Metaphysic
Hyper-realistic, real-time face replacement and de-ageing designed for live broadcast and Hollywood performance capture.
Provenance, Integration & Hardware
Synthetic output tracking and workflow automation have matured into standardised, enterprise-grade protocols.
Model Context Protocol (MCP)
Donated by Anthropic to the Linux Foundation's Agentic AI Foundation in late 2025, MCP has rapidly become the universal connective tissue of the production pipeline. Rather than engineering bespoke, brittle plugins for every AI model to communicate with an NLE or VFX suite, MCP standardises how creative agents securely query studio asset managers, read timeline metadata, and execute automated tasks directly inside software like Blender, Maya, or Nuke. It is the unheralded infrastructure layer shifting AI from a chat interface into an active pipeline assistant. Elgato Stream Deck, the hardware controller widely used by video editors, broadcast operators, and live producers to trigger actions across their entire software stack, is now integrating MCP — a sign of how far into the physical creative environment this standard is already reaching.
Synthetic output tracking has split into two complementary layers: C2PA handles the context, SynthID survives the compression.
C2PA — Content Credentials
Supported by Adobe, Microsoft, OpenAI, and the BBC. Injects cryptographic metadata detailing exactly what is human and what is synthetic. Fully integrated into Chrome and Search — though it remains vulnerable to simple screenshot stripping.
SynthID — Google DeepMind
An invisible, robust watermarking system embedded deep within the pixels and audio waves of synthetic outputs. Built to survive heavy re-encoding, editing, and compression. Over 100 billion assets already tagged.
Apple Silicon — Mac mini M4 / MacBook Pro
Increasingly popular as a dedicated local AI workstation in creative industries. The key advantage is Unified Memory Architecture — CPU and GPU share a single high-bandwidth memory pool, meaning a Mac with 64GB or 128GB can run large models that would otherwise require enterprise-grade dedicated graphics hardware. Apple's MLX framework runs these models with minimal power draw and near-silent operation, making the compact M4 Mac mini a cost-effective AI node for independent professionals.
NVIDIA RTX Spark
The Windows ecosystem's answer to Apple's unified memory advantage. Combining Blackwell GPU architecture with an Arm-based CPU via NVLink, RTX Spark supports up to 128GB unified memory and delivers 1 petaflop of AI performance — enabling on-device runs of 120 billion parameter models without cloud dependency. Adobe is already re-architecting Photoshop and Premiere Pro to run natively on this hardware.
A Final Note
This reference is accurate as of June 2026. New model versions will drop, companies will consolidate, and several tools on this list will simply become default plugins in the editing software you already open every day.
But that is precisely the point.
The core question was never about which specific model is best. The real question is whether you understand what this technology is doing to the pipeline you are entering — and whether you are going to be the person making those pipeline decisions, or the person they are made for.
The framing has changed. These are no longer tools you evaluate, trial, and decide whether to adopt. They are infrastructure — already embedded in the pipeline before you arrive, already shaping the decisions being made around you. The question is no longer whether to engage with them. It is whether you understand them well enough to direct them.
AI is not coming to replace creative practice. It is already fundamentally inside it. The practitioners who will thrive are those who can tell the difference between a tool that genuinely elevates the craft, and one that merely averages it out.
James Cameron, known for his passion for new technology while also warning us of its potential dangers, said generative AI performance puts "all human experience into a blender." Perhaps he's right. A blender in the right hands creates something extraordinary. In the wrong hands, it just makes an incredible amount of noise at high speed.
Last updated 06.06.2026.