Neu Wave AI Film Festival 2026: generation got cheap, directing did not

A talking dog that cost $12,000 to animate a decade ago was redone this year for a little under $100, and when the speaker asked who in the room uses Seedance, the answer was “all of them”. Making a shot is no longer what holds AI film back. The stage kept returning to three other things: keeping a character the same from shot to shot, finding people who can direct, and getting paid. For now the money is in vertical micro-dramas, where the host estimated that about 90% are made with AI since Seedance came out. I went to the Neu Wave AI Film Festival in Santa Monica on 14 October 2026, part of LA Tech Week, as an observer: I stopped building AI film generation two years ago and wanted to see whether I want back in.

The short version

• Generation is cheap; holding a film together is not. Bryan Michael Stoller’s talking dog was redone for under $100 in under three days, and Stoller still named continuity the biggest problem left. TopView built its whole pitch around it.
• Micro-dramas are where AI film makes money now. TopView’s CEO sees the biggest opportunity in Asia, especially China, where the film’s fact card puts the 2024 market at about ¥50B (~$7B), bigger than the country’s box office.
• Viewers stopped minding. By host Tony Shyu’s account, platforms that refused AI in October 2025 now run mostly AI dramas.
• The people who pay for films want hybrid. A VFX veteran will not pay for an AI feature but will for a hybrid one, where AI replaces the effects work and not the actors.
• My verdict: the shortage is directors. The editors I saw there do what I built two years ago, with newer models. Generated or shot, a video still needs someone to direct it, and that is where I am working, in Life2Film.

The film runs 19 minutes, with English and Russian subtitles. Jump to: consistency · China and micro-dramas · the money panel · the films · Gum Shan · what I took home.

This was the festival’s third edition, from 4 to 9 pm on the ground floor of the Lionsgate building at 2700 Colorado Ave, across the road from the Water Garden campus where Amazon has offices. A ticket was $100 (event page). The founder and host, Tony Shyu of Himalaya Entertainment, set the tone with a line from an earlier interview: in two or three years “there won’t be any AI film festivals. They’ll just be film festivals”, because every film will use AI (0:00).

Consistency is still problem number one

Stoller, author of Filmmaking For Dummies, ran the opening seminar. The example was The Amazing Wizard of Paws, Stoller’s film from about ten years ago in which the dog talks. For the remaster, the dog shots took under three days instead of three weeks and cost a little under $100 instead of $12,000 (1:46). In the full talk, not in the cut, Stoller said the 30 to 35 talking shots were done with Hailuo on MiniMax, from a still of the dog plus an actor’s recorded performance and an emotion in the prompt, and that a usable take came on about the third try. To keep costs down, Stoller buys credits on fal instead of paying for subscriptions.

Then the problem Stoller raised unprompted: continuity, “making sure that your character stays the same from shot to shot” (2:20). Tony Shyu said consistency was the reason for switching from Higgsfield to TopView (2:35); in the full recording the reason given is TopView’s early option to extend 15 seconds of Seedance footage from exactly where it ended. Dorothy from TopView’s LA team listed the failures they see over and over: one video spread across separate tools, faces, products and logos that drift between shots (they lock them with character sheets and reference images, 2:48), and small fixes that force a full regeneration.

Asked about favorite models, Stoller answered for the room: Seedance is the number one and has become quite a bit more expensive since it got popular. Then Stoller asked the room who uses it and answered: “all of them” (3:05).

One line from the full recording that did not make the cut is a different kind of consistency. Hallelujah Johnson, a finalist with the micro-drama Sire, said the generators tend to lighten Black and brown characters every time they generate, which means many extra iterations.

Micro-dramas: the money, as the speakers told it

Jensen Wu, CEO of TopView, named three problems for AI drama teams: production efficiency (many teams have one to three people for a whole series), cost, and steady income for creators. The opportunity Wu sees is micro-dramas in Asia, especially China, already a market of billions of dollars, and Wu expects the US market to follow (3:21). TopView’s answer is Drama Studio for full episodes and, next, distribution partnerships with TikTok and other micro-drama platforms (3:54).

For scale, the film’s fact card (3:48): China’s micro-drama market in 2024 was about ¥50B (~$7B), bigger than China’s box office, with 576M viewers. Those figures were not said on stage; the card takes them from CNSA via CNBC and Caixin.

Tony Shyu told the turning point (4:12). In October 2025 the platforms were all real actors and real shoots, and they all said no to AI because nobody would watch it. After Seedance came out in February, about 90% of micro-dramas are AI, in Tony’s estimate, not a measured share. Friends of Tony’s watch them knowing it: “I know it’s AI, but I just keep watching.” Tony’s conclusion: with a good, engaging story, people watch, AI or not.

Distribution came from Vanessa Mariveles, founder of AI M0V1E, a streaming app only for AI films. It started vertical-only for micro-dramas and opened to 16:9 to reach more users. The newest way to earn is unlocks: in a 10-episode series the viewer watches 4 free and buys the rest (5:06). From the full talk: creators keep 60% of ad revenue (70% once a video passes 100 million views), ads switch on at 1,000 views and 500 followers, and creators keep 95% of merch sold through a Shopify store inside the app.

The micro-drama block was TopView’s competition, with $5,000 for the winner (5:44); my recording does not include who won. Hallelujah Johnson said a minute of Sire used to take a week and takes a few hours with TopView, and that all three distribution offers so far were turned down, holding out for better positioning. “I don’t care that it’s AI,” Johnson added; what counts is a great story that resonates (6:05).

The money panel: hybrid, a GitHub for filmmakers, the end of technical slavery

The panel was called “Making AI Filmmaking Sustainable”. Nick, a visual effects veteran introduced with Aquaman and Star Wars among the credits, said almost every major VFX company now has an AI branch that is growing while every other department shrinks (6:30). Nick’s example was Crooks, a Singapore series whose budget was cut by 40% at the last minute. Rather than cut the story, the team kept the cast, shot live action in an empty room with makeshift props, and did all the relighting and backgrounds with AI, saving 80 to 90% on set lighting, as Nick told it (6:42). Then the line of the panel: “I still don’t want to watch your AI feature. And I’m not going to pay to see it.” A hybrid, yes, because there AI replaces the effects work, not the human side (7:13).

Tony Shyu had made a version of a feature in pure AI and a second as a hybrid, and argued that human actors’ imperfections are the part AI will never remove, so actors need not fear for their jobs (7:29).

Bryan Liu, a principal at Alumni Ventures and, by Liu’s own account, an investor in Higgsfield, split AI entertainment into four layers: foundation models, creation tools, distribution (maybe an AI-native Netflix or YouTube), and community. The last is where Liu wants to invest: a “GitHub for filmmakers”, where people share techniques and learn from each other (7:59).

Max Balter, formerly at Higgsfield and now at Magnific Creative Group, spoke as a director and producer of more than 20 years who was against AI for a long time. Now: “It’s the end of technical slavery.” In the full recording, scripts that sat waiting for funding and a crew now go from idea to motion picture in days rather than months. But it needs a creator behind the machine, and with a great story, “no one cares how it was made”; the worry is a lot of technically beautiful content with little meaning in it (8:19).

Justin from Loyola Marymount University’s film school teaches a course on producing with AI that some faculty dislike and that still fills fastest, with a wait list. The case study: of 12 grad students in the summer course, a 24-year-old pitched an animated series about the first town to get an AI mayor. Twelve weeks of notes later, the three-minute piece led to seven meetings, four job offers and a full-time job (9:19).

What the films look like now

Long films are arriving fast, by Tony’s count: the festival’s first year got two submissions over 10 minutes, last year five, this year 80, and 20 of them run over an hour (9:54). Tony’s prediction, which is a claim and not a fact: by this time next year there will be a blockbuster AI feature, full AI or hybrid (10:42).

Tony’s own Encore is the hybrid case: real actors, AI for everything else, made with TopView, directed through ChatGPT connected to TopView, which Tony calls the “first line producer” (11:04). In the panel Tony added a directing detail: when only one AI actor in a shot was acting, ChatGPT gave every actor a motivation, and the regenerated 30 seconds came out completely different.

The film that stayed with me was Gum Shan (Gold Mountain), directed by Max Acchione, about a Chinese miner in 1870s Antioch, California (11:29). Antioch’s Chinatown was burned to the ground in 1876; in 2021, 145 years later, the city formally apologized (both from the film’s fact cards). On the making of it, Acchione said that where others talk about three takes to one, Acchione’s ratio is closer to fifteen to one, and fifty to one on problem shots. Working with the models is like the best cast and crew ever, “but they’re all on crack”: between shots they change the set, the lighting, the clothes (12:25). In the full Q&A, Acchione said the team built its own harness around Seedance, and that the film came out of an open screenplay competition and doubled as a research project for UC Irvine on what the tools can do.

Also screened: Beyond the Sky, The Ferryman’s Boat (John Steele), MindPods by Jeff Synthesized, whose The Gifted won the Future Vision XPRIZE, The Girl Who Danced at Night, and Tower of One Hundred Demons (James Mitchell) (12:48).

Two craft notes from the directors’ Q&A. A filmmaker’s trick: act the scene out yourself or with an actor, as close to the action as possible, and the AI copies that performance instead of inventing one, which the filmmaker promised is a hundred times better (13:09). And the director of Road North, a solo filmmaker from London, had a story ten years in the making: Runway Gen-3 and Kling 1.6 were not up to it, Seedance was, and the film took two months working alone (13:37).

My verdict, from the car

I recorded this at 11 pm, driving home (14:01). It is in Russian in the film, with English subtitles.

Consistency first. Stoller’s problem is the one I was fighting two years ago. It more or less worked then, but not all the way, and it is still there and growing: the models get more powerful, but control between shots, so that a character and the objects do not change and the whole film holds together, has not kept up (14:29).

Then stories and directors. Max Balter’s point from the stage is the bigger one: the problem now is the people who will create and edit the stories. It is the same in software. You have the code, and what you get depends on you, not on the code (15:02). Resistance to AI is falling and the films are getting stronger.

A question I wanted to ask and found nobody to ask: when will a director win an Oscar with an AI film? My guess is two or three years; until then, short films and maybe documentaries (15:45). I agree with the host that the split will go away, the way nobody now separates films shot on digital from films shot on film. AI becomes a tool, mostly in production, a new layer like animation or 3D effects. Cinema did not replace theatre; theatre stopped being mass entertainment, and the two live side by side (16:03). One film said at its end that it would not have been possible without AI.

Then what to build (17:19). The editors I saw that day are competitors to what I built at SuperDuper AI two years ago: the same things, with new APIs and more powerful models. Back then I burned out and found it hard to finish the product, and I shut the project down, because what I wanted from AI films was not working, at least in that format. Now the desire is back (17:34).

But I already build project-life2film, an iPhone video editor that is in effect a director. Whether you generate video with neural nets or shoot it on a camera, you still need a director afterwards. People say there are no ideas and no directors; I think that is the main problem, and the one I want to focus on, whatever the footage comes from (18:06). So the answer to the question I came with: the interest is back, and it points at the director, not the generator.

That matches the trail I already have. My last public generation project was project-superstoryboard, a storyboard generator from the Supabase hackathon in November 2025 (Gemini images, Veo 3.1 clips, Figma sync). The latest on the director side is neon-personal-agents-life2film-director-2026, an agent that learns editing taste from a month of my footage. And this festival film was edited with Life2Film.

How this was put together

I filmed the evening on a DJI camera, 20 clips from the walk past Water Garden to the drive home, transcribed them locally with mlx-whisper, and edited the 19-minute film with Life2Film from the transcript. Times in this post are times in that film. Stage quotes follow the film’s English subtitles; numbers are as the speaker said them and attributed to that speaker, and predictions are marked as predictions. Details marked “full recording” or “full talk” come from the raw transcripts and are not in the cut, so they have no timestamp. The China market figures are the film’s fact card, citing CNSA via CNBC and Caixin, not a stage claim; the Antioch dates are fact cards too. People are named only where they were introduced by name on stage or credited in the film. Audience questions and conversations I had at the festival are left out. My recording does not include the micro-drama competition winner.

The next day, also in LA Tech Week: la-tech-week-video-agents-live-worlds-2026. Related: project-life2film, neon-personal-agents-life2film-director-2026, project-superstoryboard.