In partnership with
Welcome to Today’s AIography!
Happy Tuesday. Two big things happened in cinema this month. The Academy formalized AI rules for the 2027 Oscars, with the first submission deadline four months out. Steven Soderbergh defended the 10 percent AI use in his John Lennon documentary that premiered at Cannes on Saturday. Both moments point at the same question: who's getting credit for what?
Meanwhile, NVIDIA released a free open-source video model that runs on a workstation graphics card. Runway shipped a real-time AI talking-head tool. And the Financial Times reported that Chinese AI video models from ByteDance and Kuaishou have moved ahead of US rivals on quality and pricing.
In today’s AIography:
Oscars 2027 AI rules
Soderbergh defends his AI-assisted Cannes documentary
NVIDIA released free, one-minute AI video
Runway shipped a real-time AI talking head from one photo
Chinese AI video tools pulled ahead of US tools
Essential Tools
Short Takes
Today's Video
Final Thoughts
Read time: About 8 minutes
THE LATEST NEWS
INDUSTRY POLICY
AMPAS Just Formalized Its AI Rules for the 2027 Oscars
Image generated with ChatGPT Image 2
TL;DR: On May 1, the Academy of Motion Picture Arts and Sciences approved new rules for the 99th Oscars, set for March 14, 2027. Generative AI is permitted across most categories, but the Academy reserved the right to ask about AI use and weigh "the degree to which a human was at the heart of the creative authorship." Screenplays must be human-authored to qualify for writing categories. Acting nominations require performances "credited in the film's legal billing and demonstrably performed by humans with their consent." The first major submission deadline is September 17, 2026. That is four months out.
Key Takeaways:
Generative AI use is not banned at the Oscars. It is now subject to formal Academy review on any film submitted for consideration.
Screenplays must be human-authored. Acting performances must be human and consented. Both categories now have explicit AI clauses in the eligibility rules.
AI-rendered performances like the previously announced Val Kilmer recreation no longer qualify in acting categories. The "demonstrably performed by humans with their consent" language closes that door clearly.
Editing, cinematography, post-production, sound, and visual effects are not directly addressed by the new clauses. AI tools remain eligible in craft categories, but documenting their use is the prudent move.
Deadlines: shorts and documentaries close August 13, 2026. General categories close September 17, 2026. International features close September 30, 2026. Final categories close November 12, 2026. Ceremony: March 14, 2027.
My Take:
We led with this story two weeks ago when the rules first dropped, and I am leading with it again because the deadline math just changed. Four months from now, the first submissions hit. If you are working on a feature you intend to submit, you do not have a year to figure out what your AI documentation looks like. You have until September 17, 2026.
The Academy did not ban AI. It set a bar. The bar is a human has to be at the heart of the creative authorship, and the Academy can ask you to prove it. That is not a gotcha. It is the same direction Cannes is converging on this week, which is the next story in this issue. When the Oscars and Cannes agree, that is a signal. The rest of the industry follows.
Practically, what does an authorship paper trail look like? It looks like the same project logs working editors and writers have always kept. Who wrote which draft? Who directed which take. Which AI tools were used on which shots, by whom, with what prompts, and what reference material? The version-control habit we already use for our edits, extended to the AI step. The bar is not perfect documentation. The bar is "Could I answer the Academy's question if asked?" If yes, you’re fine.
Try This Now:
Lead action: Start an authorship log this week. One Google Doc or one Notion page per project. Three columns: date, AI tool used, what the human did with the output. Backfill the current project if you can. The habit takes ten minutes a day and becomes muscle memory in two weeks.
Lighter touch: Pull up the Academy's full rule text and read the new generative AI clauses. Bookmark it. Whether you submit a film or not, this is the language other festivals, guilds, and competitions will start adopting in the next twelve months.
Image credit: Kishin Shinoyama (via Cannes Film Festival)
TL;DR: Steven Soderbergh's John Lennon: The Last Interview premiered Saturday, May 17 at the 79th Cannes Film Festival. The 97-minute documentary, edited down from a 165-minute Lennon interview recorded hours before he was killed in December 1980, uses Meta AI for roughly 10 percent of its visuals. Meta partly financed the film and provided the AI tools. Critics in Cannes slammed the AI sequences. Soderbergh disclosed the Meta partnership earlier this year and defended the choice from the Cannes podium: "I knew what was coming. You don't say yes to Meta offering you these tools and offering to finish the film and not know you're going to come in for some heat." His broader argument: the problem is not that he used AI. The problem is the people using it without saying so.
Key Takeaways:
The film has no distributor yet. The Cannes premiere is the test run with the most adversarial press audience in cinema.
Roughly 10 percent of the visuals are AI-generated, used to render "dream spaces" around the audio of the Lennon interview itself.
Meta partnership was disclosed earlier this year. Soderbergh's argument is that disclosure is the new craft norm: not "no AI" but "be honest about which parts."
Critic reaction in Cannes split between technique judgment (the AI sequences read as visually weaker than the rest of the film) and the broader question of disclosure becoming a credit-chain expectation.
The disclosure argument lines up directly with the Oscars 2027 rules and with Cannes director Thierry Frémaux's proposal last week for a "non-AI" label on selected films. Two institutions, one direction of travel.
My Take:
I have been watching A-list directors test the AI water in public for the past year, and Soderbergh is the most interesting case to date because he is not pitching AI as a creative breakthrough. He is making a craft argument about transparency. That is a different conversation than the one most of the industry has been having.
The 10 percent number is the point. The Lennon audio is the entire film. The "dream space" visuals are the connective tissue. If you tell the audience up front that the visuals are AI and the audio is real, the audience can evaluate the work on those terms. If you do not tell them, they have no way to know what they are watching. Soderbergh's bet is that the second category is what is actually undermining trust in cinema right now, not the first.
TL;DR: NVIDIA released SANA-WM on May 16, an open-source video model that generates one minute of 720p video on a single consumer graphics card. The model is 2.6 billion parameters and uses a Diffusion Transformer architecture, which is the standard backbone for state-of-the-art image and video models. With NVIDIA's NVFP4 quantization (a method for shrinking the model so it runs faster without losing meaningful quality), the distilled version generates a 60-second 720p clip in 34 seconds on an RTX 5090 (the most powerful current consumer card). The code is free to download from GitHub and free to use commercially. Until this week, most free AI video tools maxed out at five to eight seconds. A full minute changes the math for narrative work.
Key Takeaways:
Free to download. Free to use commercially.
Runs on a consumer graphics card you would buy for $1,500 to $2,500. No cloud subscription needed.
60 seconds of 720p output is the threshold that makes narrative video viable on local hardware.
Joins four other free AI video tools (SwiftI2V, MOVA, Wan2.2, HiDream) that all shipped in the past two weeks.
Output quality is not quite top-tier cloud yet, but it is good enough for pre-viz, test renders, and iteration.
My Take:
What changed this week is not just another open-source video model. What changed is the floor, the minimum for one person to make AI video without paying a cloud service. As of May 16, the answer is a desktop graphics card you can buy today, (albiet an expensive one) plus a free model called SANA-WM.
For a year I've said the free, run-it-yourself tools would catch up to the paid cloud tools on the basics that matter to working creators: clips a minute long, 720p quality, runnable on one machine. This week, that line was quietly crossed. Whether SANA-WM looks better than Veo 3.1 or Seedance 2.0 on any given prompt is a different question. The real question was whether your laptop or workstation is a serious option yet. As of this week, it is.
The practical part is cost. If you've been paying $300 to $500 a month in cloud video credits, you now have a free local option for the work that doesn't need the best possible output. Save the paid credits for your final shots. Use the local tool for rough versions and prompt experiments you used to burn credits on. The math adds up fast.
Your prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.
TL;DR: Runway shipped its Characters product on May 4, the formal launch of what they announced in late March as a developer preview. The input is one reference image. The output is a real-time conversational video agent that performs lip-sync, facial expressions, and head motion at 24 frames per second in HD. The lag from your voice ending to the character starting is 1.75 seconds. The model behind it is GWM-1, Runway's General World Model, a diffusion transformer with a VAE decoder running in pipelined parallel. Available now in the Runway API and across Runway's web and mobile apps.
Key Takeaways:
1.75 seconds is the latency number that matters. Below that threshold, the character feels conversational. Above it, the audience reads the lag as broken.
One reference image is the entire input. No video training data, no character rig, no motion-capture session.
The output is real time, not pre-rendered. You can speak to the character and it responds with synced visuals while you are still talking.
Use cases Runway is highlighting: vision and webcam input, custom voices, tool calling, knowledge bases, embeddable widgets, and integrations with Zoom, Google Meet, and Microsoft Teams.
Available now via the Runway API plus web and mobile. No waitlist.
My Take:
Talking-head video is the format that AI was always going to crack first, and Runway just shipped the version that crosses the believability line for short-form work. The 1.75-second turn-around is the part to focus on. Real-time character work has been the holy grail of AI video since 2023 because every other version we have seen sits in the uncanny valley of delayed response, frozen expression, or out of sync mouth. Sub-two-second latency is conversational. That’s a different product.
For practical use right now: explainer videos, interview-format content, virtual hosts, and quick-turn reaction clips. Not for theatrical work where the character carries dramatic weight. Not yet. The expressive range still has the AI tells if you are watching for them. But for the broad swath of talking-head content that working creators produce every week, this is genuinely useful. One reference image into a character that holds a conversation. The cost calculation versus shooting a person changes immediately.
CREATOR ECONOMY
Chinese AI Video Tools Pulled Ahead of US Tools.
Image created with ChatGPT Image 2
TL;DR: The Financial Times reported Saturday that Chinese AI video models from ByteDance (Seedance 2.0) and Kuaishou (Kling 3.0) have moved ahead of US rivals on quality and usability, according to working developers and the public Arena.ai video leaderboard. The same week, Kuaishou announced plans to spin off Kling as a separate company and raise two billion dollars at a twenty-billion-dollar valuation, with Tencent reportedly in talks to participate. The shift is uneven. American labs still lead on large language models and coding tools. On video, the FT reporting confirms what working creators have been saying on X for months: Seedance and Kling are now the tools that hold up best on real production work, the pricing is more flexible than US tools for individual creators, and the gap is not closing in the West's favor right now. The context: in January, Google DeepMind CEO Demis Hassabis told CNBC that Chinese AI is "a matter of months" behind Western frontier labs, not years.
Key Takeaways:
The Arena.ai video leaderboard now ranks Kling, Seedance 2.0, and Alibaba's HappyHorse 1.0 among the top-rated text-to-video and image-to-video models, with Google's Veo 3.1 still competitive but no longer at the top.
Pricing model is part of the story. Chinese tools are mostly subscription or pay-as-you-go for individual creators, which lets you experiment quickly. American tools are mostly pay-per-generation credit packs, which makes iteration expensive. ByteDance does charge differently for enterprise: the FT reports two-million-dollar upfront commitments for some US enterprise clients.
Ben Chiang, founder of Director AI (a startup that produces AI animated content and short dramas), told the FT his team primarily uses Kling and switches between Seedance 2.0 and MiniMax's Hailuo depending on the task. He said most American models "are not very good at video generation" and that stricter content controls in the US produce less realistic outputs.
Kuaishou is exploring a separate stock listing for the Kling business to capitalize on the growth. The two-billion-dollar raise at a twenty-billion-dollar valuation would make Kling one of the higher-valued standalone AI video plays globally.
Hassabis's earlier "months not years" framing came with an important qualification: Chinese firms have shown they can catch up to the frontier and approach it closely, but he said they have not yet shown the ability to invent something fundamentally new beyond it. The implication for working creators: the tools work today, but the next foundational breakthrough may still come from the US or UK.
My Take:
The Financial Times reporting just made official what working creators have been quietly routing around for months. On AI video specifically, the practitioner evidence has been pointing at Kling and Seedance for at least six months. The institutional press caught up Saturday.
What matters for the AIography reader is not the geopolitics. It is the tooling question: should you be learning Kling and Seedance right now if you are not already? The honest answer is yes, at least enough to know what they can do on the kinds of shots you make. If you are working in narrative video, in advertising, in social-first short-form, the gap between what comes out of Kling or Seedance and what comes out of Veo or Sora is now consistent enough that the people doing real production work are routing to the Chinese tools for the harder shots. The advertising case is the loudest signal. Vincent Yang, the CEO of Firework (a video infrastructure company for ecommerce), told the FT his team is now generating tens of thousands of product videos at scale that previously would have been "prohibitively expensive." That is a different business case than experimenting on a hero shot.
The pricing-model difference is the part that flips this from "tracking" to "trying." If you have been priced out of iterating on Veo or Sora because the per-generation cost ate your budget on the first ten test renders, a Kling or Seedance subscription is a different cost calculation. You can fail faster. That matters more than top-of-leaderboard quality on any given prompt.
The cautious frame: data is part of why the Chinese platforms have the edge here. ByteDance and Kuaishou own the largest short-video platforms in the world, and that footage is the training material. American labs do not have equivalent access. That gap is not going to close quickly. Plan accordingly.
ESSENTIAL TOOLS
AI Filmmaking & Content Creation Tools Database
Check out the Alpha version of our AI Tools Database. We will be adding to it on a regular basis. Got a tip about a great new tool? Send it along to us at: [email protected]
NVIDIA SANA-WM: open-source 60-second 720p video on a single workstation graphics card. NVlabs/Sana on GitHub.
Runway Characters: real-time AI character from one reference image. Available in Runway API plus web and mobile apps.
Kling and Seedance 2.0: Chinese AI video models now leading practitioner rankings. Subscription and pay-as-you-go pricing for individual creators; enterprise terms differ.
Color Mode in Adobe Premiere (beta): built-from-scratch color grading for editors, free for all Premiere subscribers.
DaVinci Resolve 21 (Public Beta). Free upgrade. New Photo page + 9 new AI tools including Face Reshaper and Face Age Transformer.
SHORT TAKES
AI-native production companies launched at Cannes 2026 Marché du Film: Storyverse, amersia (Vertigo + Federation Studios, behind OpenAI-backed feature Critterz), Innovative Dreams (Wonder Project + Luma), and The Next Valley (nmatic.ai + Alibi Studios). The business model is forming in public. Tool companies becoming studios, and filmmakers who can direct AI systems becoming the new hires.
Open-source video releases hit five in ten days: SwiftI2V, MOVA, Wan2.2, HiDream, and now NVIDIA SANA-WM. All single-GPU runnable. The local stack now has a credible end-to-end pipeline for narrative-length video, not just the experimental short clips of six months ago.
Cannes week 2 disclosure pattern continues: Beyond Soderbergh, the festival has been quietly converging on disclosure as the new craft norm. Demi Moore's jury position, the Frémaux non-AI label proposal from last week, and the SAG-AFTRA contract conversation on synthetic performances all point in the same direction.
Diesol's Tairell screening at Runway's AI Festival: The fourth annual Runway AI Festival hits New York on June 11 and Los Angeles on June 18. Tairell is in the program. Theatrical screening of AI-native work is the milestone the field has been waiting for.
Veo 3.1 remains Google's current stable video model: If you see fresh-sounding "Veo 3 launch" headlines this week, those are reviews and re-coverage of the May 2025 model, not a new release. Standard clip length is eight seconds with Scene Extension for longer output. Veo 3.1 from October 2025 is the current stable.
SAG-AFTRA synthetic-performer contract talks continue: The union is still working through how synthetic performances slot into the existing collective bargaining structure. The Oscars rules announced this month make that conversation more urgent.
ONE MORE THING…
Today’s Video
“Clickbait” by Ryan Phillips (aka: Uncanny Harry)
Ryan Phillips better known on X as Uncanny Harry posted a 3-minute AI horror short today, made entirely inside InVideo's new Agent One tool. It's a sponsored video, but the film is genuinely good and the mood stays steady all three minutes. Phillips also posted a thread on X breaking down how he made it: he fed the script to an AI helper that asked him about visual style, then used the tool to generate storyboard frames, picked which AI model handled each animation step, and put the whole thing together in Adobe Premiere. The thread is the real value. A working filmmaker showing the exact steps he took.
FINAL THOUGHTS
Three stories this week point at the same conclusion, and a fourth one widens the frame.
The Academy and Cannes both arrived at "human authorship is the new eligibility test" within the same two-week window. That convergence is not a coincidence. The institutions read each other. They look at each other's drafts before publishing their own. When the two most prestigious bodies in cinema land on the same answer in May, the rest of the festival circuit and the guilds follow by year-end.
NVIDIA shipped a free, open-source video model that runs on a workstation graphics card. The tools are racing forward and getting cheaper at the same time the institutions are drawing the authorship lines. A working filmmaker now has more powerful AI in their pipeline than they had six months ago, while also having more formal rules about how to credit it. Both moving at once is the condition of the field right now.
The widening frame is the Financial Times reporting on Chinese AI video. For most of the past two years, the public conversation has assumed American AI labs were running the field. On video specifically, that is no longer the case. ByteDance and Kuaishou are doing the better work, the pricing model is more practical for working creators, and the gap is not closing in the West's favor. For the AIography reader, this is not a geopolitics story. It is a tooling decision. If you have not tried Kling or Seedance because they felt like the second-tier option, that assumption is out of date. Add them to the routing list this week.
The job for the next four months is straightforward. Build the documentation habit. Try the in-NLE AI you already pay for. Test the local open-source stack on something low-stakes. Try the Chinese tools your competitors are already using. The infrastructure is real on multiple fronts. The rules are forming. The window to be ready before the September submission deadlines is open right now.
Stay sharp. Keep creating.
— Larry
$992 Billion in Art Could Change Hands. Why Are These 71,105 Investors Paying Close Attention?
Deloitte ran the numbers. They project UHNW art and collectibles wealth -- already at $2.5 trillion -- to hit $3.47 trillion by 2030.
The institutional world has been quietly preparing for this. Back in 2011, 25% of wealth managers surveyed offered art-related services. In 2024, 51%. Family offices now average a 13.4% allocation to art and collectibles.
And it’s not just because they love art. It’s because they like the math.
These positions were built over decades through private dealer relationships most investors never had. The access just wasn't there.
Masterworks is changing that by allowing individuals to invest in shares of blue-chip artwork by artists like Basquiat, Warhol, and Picasso. Their track record to-date:
71,000+ investors
$1.3B deployed across 525+ artworks
29 closed sales
Net annualized returns like 16.5%, 17.6%, and 17.8%, not including those unsold.
Investing involves risk. Past performance is not indicative of future returns. See important disclosures at masterworks.com/cd.
What did you think of today's newsletter?
AIography is the AI filmmaking newsletter for filmmakers, editors, and content creators navigating the biggest technological shift since digital.
If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.
AIography may earn a commission for products purchased through some links in this newsletter. This doesn't affect our editorial independence or influence our recommendations—we're just keeping the AI lights on!
