Seedance 2.5 Advice with Tim Simmons "Theoretically Media" | fal Podcast

The most useful thing in fal's first podcast episode is not how good Seedance 2.5 is. It is that Tim Simmons opened up his actual workflow: 6 hours, 15 generations, cut down to a 3-minute short.

The test he runs matters more than the specs

Every time he gets early access he asks one question: can this hold a scene? Can it hold a performance?

His reasoning is the clearest thing in the episode. Every video model is great at eye candy. Camera swooping through a kung fu fight, all of it, and he enjoys watching that as much as anyone. But 90 percent of narrative filmmaking is two people talking. That is what a story is. People in conflict. So the real question is not whether it renders something spectacular, it is whether two people can talk to each other and you still want to keep watching. Seedance 2.5 handled it surprisingly well.

The workflow worth stealing

Most people who get a 30-second generation think in one of two directions: cram ten shots into 30 seconds, or generate one long 30-second take. He does neither. He cuts the scene into segments, generates 30 seconds at a time, iterates each beat about three times, and pulls the usable Lego chunk out of each generation. Fifteen generations became a three-minute short in about six hours.

And the line I keep repeating: at the end of the day you still have to edit. You generate, then you edit. AI video as it currently stands is really an editor's medium.

AI filmmaking favors editors and directors

The host pushed that further. AI filmmaking now favors editors and directors. Generating a lot of footage is easy. Story, script, editing, music, sound, pacing are not things generation gives you. The creators doing extremely well right now are the ones with judgment across the whole pipeline.

Image-first versus omni reference

He was honest about the tradeoff. Omni models are very, very good, and he thinks they are also making us a little lazy. Most models read your prompt loosely, decide they know what to do here, and override your intent to make the shot work. And a lot of the time they are right. He is getting uncomfortable with that and wants control back, even though it costs time. For this film, speed won and he used a lot of omni referencing anyway.

Shots as software patches

when a shot fails or a line is flubbed, he takes a screenshot from somewhere else and runs it as image-to-video to patch the shot in. He calls them software patches. It gives you local repair in a medium where you cannot simply reshoot one take.

He still has not seen an AI edit that impressed him

He tried it on the most linear, procedural YouTube content he makes, with a script he wrote alongside GPT and Claude, and it still could not pick the right clips. Consumer-side auto-edits are a different story and genuinely useful.

Three takeaways

The question changed. Not "does it look spectacular," but "can it hold a two-hander."
AI video is an editor's medium. Generation is not the scarce part.
Speed and control are a real tradeoff. Just because the machine can do it does not mean it should.

Asked what he would do if AI turned out to be a bubble, or got so good it did not need creators: I'd still be making videos. It's what I did before all of this and what I'll do after. I can't not.

---

Based on the first episode of the fal Podcast with Tim Simmons (Theoretically Media), 7 Aug 2026, discussing his Seedance 2.5 short "Death Walks Into a Bar." Quotes are paraphrased from the episode.