Happy Tuesday. A lot moving this week.

The 79th Festival de Cannes opens today with five AI works in the Immersive Competition and a two-day AI for Talent Summit later this week. Google I/O is one week out and the rumor mill is already out of control. Krea AI dropped its first in-house foundation model this morning. Things are moving faster than dailies on a David Fincher film.

But the thing that genuinely caught me this week wasn't on X or YouTube. It was a piece of Chinese AI filmmaking we hadn't been watching, doing numbers that would put any Western AI account I follow into a coma.

Don’t miss our Must Watch AI Video. 2.9 million views on X in three days! The numbers on this one tell you something.

In today's AIography:

Read time: About 7 minutes

THE LATEST NEWS

INDUSTRY SHIFT
Cannes 2026 Opens Today With the Festival's Strongest AI Programming Yet

TL;DR: The 79th Cannes Film Festival opens today (May 12) and runs through May 23. Two AI-related programs are worth tracking this year. The Compétition Immersive selected five AI works for its 2026 program. Cannes Next is running a two-day AI for Talent Summit on May 15-16, with Darren Aronofsky, Google's James Manyika, OpenAI's Chad Nelson, and Luma AI's Josh DiCarlo on the speaker bench.

Key Takeaways:

My Take:

The story is the scale, not a new category. Cannes has had AI work inside the Immersive Competition before. The difference this year is the concentration of programming, and especially the speaker bench at the AI for Talent Summit.

When Darren Aronofsky is sitting on a Cannes panel about AI tools alongside the President of Research at Google and the head of DreamLab at Luma, the credibility line has moved. The conversation at the top of the industry shifted from "is this legitimate filmmaking" to "which of the five Immersive Competition works was the strongest this year."

For a working filmmaker, the practical signal: festivals are how the industry credentials new categories of work. Cannes is the loudest signal in the chain. Toronto, Venice, Berlin, Sundance, and Tribeca all watch what Cannes does. Expect the rest of the circuit to expand their AI programming over the next twelve months.

Try This Now:

Take ten minutes today to review the five Compétition Immersive selections at the URL below. They are the bar Cannes set for AI-collaborative work this year. Frame your own work in progress against what they did well or what they left untouched.

MODEL PREVIEW
Google I/O Is One Week Away. Veo 4 (Or A New "Omni" Model) Expected

TL;DR: Google I/O 2026 runs Monday and Tuesday next week (May 19 and 20). Google has used I/O for major Veo announcements in 2024 and 2025, so a 2026 reveal is expected. A pre-event leak from the Gemini UI shows a "Powered by Omni" label near the video generation tab, hinting at either a new wrapper for Veo or a brand new Gemini multi-modal model. Prediction markets give about 69 percent odds of a Veo 4 launch before June.

Key Takeaways:

My Take:

I am not going to rebuild a working pipeline on a leak. But I am going to wait until next Wednesday before locking any new Veo 3.1-dependent shooting plans. The math is straightforward. A Veo 4 launch shifts the routing equation. An Omni launch shifts it more.

If you are running Veo 3.1 in production this week, finish what you started. Hold any fresh long-horizon work that depends on the current model behavior. Park new projects for seven days.

Try This Now:

Make a one-line note in your project folder of any Veo 3.1 quirks you have learned to work around. When the new model ships, that note is your retest cheat sheet. Without it, you will spend hours rediscovering what you already know.

TOOL LAUNCH
KREA 2 Launches Today As Krea AI's First In-House Foundation Model

TL;DR: Krea AI announced KREA 2 today (May 12), its first foundation model built entirely in-house rather than fine-tuned on third-party weights from Stability AI or another lab. The pitch is aesthetic diversity and precise style control. K2 Studio, the new style-template feature, lets creators build a reusable look using style guidelines, comment guidelines, and a like-and-dislike feedback loop.

Key Takeaways:

My Take:

The style-template approach is the part working creators should pay attention to. Most image-gen platforms treat style as a one-shot prompt: type the right reference name, hope the model still knows it next week. K2 Studio treats style as a trainable instrument. Guidelines plus a feedback loop equals a reproducible look set.

That matters in production. If you need three weeks of social posts to feel like they came from the same hand, the current platforms cannot reliably deliver without either fine-tuning a LoRA (expensive, technical) or accepting drift across the campaign. KREA 2 is a bet that creators will pay for consistent style more than for incremental realism.

Krea building this in-house instead of fine-tuning on someone else's weights is also the right product call. It means they can iterate the model's training data toward their specific aesthetic-diversity bet, rather than inheriting whatever bias the base model came with. Expect more platforms to follow once they realize their differentiation is in the model itself, not in the UI wrapper.

Try This Now:

Open krea.ai. Use K2 Studio to write a style template for one of your active campaigns. Style guidelines, comment guidelines, ten generations, like or dislike each, then a fresh one. See how close the eleventh comes to your intended look. That is your test of whether the feedback loop is doing real work.

FOUNDATION MODEL UPDATE
GPT-5.5 Instant Is The New Default ChatGPT Model

TL;DR: OpenAI released GPT-5.5 Instant on May 5 as the new default model for ChatGPT. Faster latency than the prior default with the same reasoning floor. Every ChatGPT user is now running on it without opt-in.

Key Takeaways:

My Take:

A quiet release with a real practical effect. If your filmmaking workflow uses ChatGPT for script drafts, prompt engineering, or story development (most working creators do), you are now running on a different default. Quality should hold because the reasoning floor did not move. The speed change is what to track.

If you have a pipeline that times out at certain thresholds, retest. If you batch requests, the new latency changes your batch sizing math. A few minutes of retest now saves real debug time later.

Try This Now:

Open your most-used ChatGPT workflow today. Time one full pass. Note the result. Compare against your memory of last week. That is your baseline for the new default.

SHORT TAKES

ESSENTIAL TOOLS

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]

ONE MORE THING…

Must Watch AI Video

A three-and-a-half minute AI short called Zombie Cleaner (丧尸清道夫) by a Chinese creator named Mx-Shell crossed 2.9 million views on X in three days. His native Bilibili upload of the same short sits at 816,000 plays and is still climbing. Either side of that math, this is one short pulling numbers that the Western AI filmmaking accounts I follow do not pull.

It is a domestic answer to Love Death + Robots. Same anthology register, same horror cuts. This is a remastered version of an earlier piece, so Mx-Shell was already iterating in public before this one broke.

The tool stack is what working creators should track. He built it inside ByteDance's Xiaoyunque "Skylark" Creator Program, an invite-only track that pairs top-tier Seedance access with native Bilibili distribution. That is a vertical integration we do not have a Western analog for. The closest comparison is Adobe Firefly inside Premiere, which is one company at the post layer, not a tool-plus-platform-plus-audience play.

The Western AI filmmaking conversation found Mx-Shell through this one viral clip. Bilibili knew him before. If you only watch X and YouTube, you are reading the wrong half of the map.

Credit: Mx-Shell. Posts natively on Bilibili. The X version that crossed 2.9M is what put him in the Western feed.

FINAL THOUGHTS

First: the bar at the top of the industry is moving. Cannes putting AI work across two competition slots plus a high-level summit with Aronofsky on the speaker bench is the most important institutional signal of the year so far. The working AI filmmakers who shipped real work in 2025 and early 2026 are now eligible for recognition that did not exist when they started.

Second: the tooling layer is shifting from third-party fine-tunes toward in-house foundation models. Krea building KREA 2 from scratch is a leading indicator. Expect more platforms to follow once they realize their differentiation is in the model itself, not in the UI wrapper. If you are evaluating tools right now, ask which ones own their stack and which ones are renting someone else's weights.

If you have shipped something this year, the runway just got longer.

See you Thursday.

Stay sharp. Keep creating.

-Larry

AIography is the AI filmmaking newsletter for filmmakers, editors, and content creators navigating the biggest technological shift since digital.

Subscribe · Join the Community · Become a Founding Member

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!