Skip to content
TheoreticallyMedia

Flux 3 Is Here — and an AI Film Is Getting a Theatrical Release

Black Forest Labs finally shipped the video model teased since 2024: 20-second generations, up to 10 references, native audio, and open weights in the launch plan. Plus the landmark: Gossip Goblin's feature film is headed to actual theaters, and the workflow breakdown — Midjourney as the generator, Nano Banana as the Photoshop, first-frame image-to-video for camera control — is worth stealing.

Grab your map — Black Forest Labs is back in the forest with Flux 3, the video model they teased at the launch of Flux 1 way back in August 2024. Same week, AI filmmaking hit a real threshold: a theatrical release. Both halves are in the video.

Flux 3: the specs and the tests

The question on everyone’s mind is how it stacks up on the Seedance scale. Short answer: it’s there-ish. The specs are legitimately good:

  • Generations up to 20 seconds — and 10 seconds really doesn’t feel like it cuts it anymore
  • Multimodal inputs: up to 10 references, plus audio or video inputs for editing and remixing
  • All the standard aspect ratios, 9:16 through 21:9, and native audio
  • 720p in the current launch window, 1080p rolling out shortly
  • No pricing yet — and per a mid-edit update: open weights are in the launch plan

I’ve had early access (via Discord — Discord haters, I’m officially on your side now), with one recurring early-access problem: image references weren’t reliably sticking to prompts. I don’t expect that to survive to release, but it shaped what I could test.

Text-to-video: I asked for a trailer for a Netflix show called Flux 3: Return to the Black Forest — that was the entire prompt — and got something I’d probably watch. Upscaled in Topaz, because of all the 720-ness of it all. From how it handled a one-line prompt, I think we can surmise this is a reasoning/thinking video model.

Image-to-video: the standard FBI-agent-in-a-Pacific-Northwest-diner test, complete with a damn fine cup of joe and pie from the log lady’s owl recipe. The notable moment: the lady in the background who’s been waiting to order for a year of these tests finally got a line. That’s what the extra seconds buy you.

Prompt coherence: the brutal community “15 shots in 15 seconds” hyperpop frantic-day-loop prompt. Everything I asked for is present, which is the point of the test. What’s missing is the kinetic energy — the fisheye trick shots and whip-fast camera moves the Seedance version of this prompt delivers. So: probably not your go-to for heavy fast-action sequences, at least at launch.

Community outputs round it out: a dramatic two-hander with rapid-fire dialogue somewhere north of Gilmore Girls, and a cool alien-invasion action piece I can’t properly credit because, again, Discord.

The verdict from the video: does Flux 3 replace Seedance in your workflow? No, probably not. Does it augment it? Yes, I think it can — and I’d expect generation costs to come in under Seedance pricing, which nobody would argue is cheap. It’s day one for this model; there’s a lot still to discover, and a waitlist for early access in the meantime.

Gossip Goblin goes to the movies

This is the biggest AI-film theatrical release to date: Zack London — the guy behind Gossip Goblin, recently called “the George Lucas of AI filmmaking” by the Hollywood Reporter (does that mean he’ll go back and revise all his previous work 300 times?) — has a feature-length film coming out, in actual theaters. In the current age of Curry Barker’s Obsession and Markiplier’s Iron Lung, backing an AI filmmaker with a YouTube-pedigree cult following is the smart play, and Gossip Goblin is exactly that. Fable’s Edward Saatchi is producing; I haven’t dug into how distribution rolls out yet, but we may have him on to discuss it as release gets closer.

In the meantime: go watch Pomegranate, the half-hour short Gossip Goblin just released. It stands alone, you need zero backstory, and it’s very good. A game worth playing while you watch: don’t count the moments you clock something as AI-generated — take note of when you don’t. The final tally on that number is much more interesting. And watch it where you watch movies, on the biggest TV you have. Not your phone.

The workflow breakdown

Between PJ Ace’s deep-dive interview with Zack and a fresh AI Live interview, there’s a lot to mine here:

  • Nothing begins in Nano Banana. Midjourney is the generator — not vanilla Midjourney, but honed personalization plus SREF codes. (At the time of PJ’s piece, Zack was on Midjourney v7 rather than 8.1; that may have changed.) Nano Banana is the Photoshop — the manipulation layer, not the origin.
  • No omni- or multi-referencing in the video generations. Everything is first-frame image-to-video, and the reason is camera control. I have to admit I kind of agree with him — the omni and multi models are making us a little lazy in that department.
  • No master character sheet. Characters are handled modularly — front, side, and back images comped into each location, with Nano Banana doing the compositing.

For the previous Gossip Goblin film, The Patchwork, Kling did all the animation; Pomegranate is definitely Seedance. If you’ve seen neither, it’s an interesting exercise to watch the Kling-era film and see where — or whether — the older model takes you out of it. I don’t think it will.

Between the theatrical news and Blomkamp’s Seedance short, big movements are happening in AI filmmaking. I’ll be sharing more of my own tips at a free live webinar with Teachable’s AI Academy — link’s in the video description, and I hope to see you there.