F
Frank Houbre
Guest
I spent far more time than I expected on a single line reading in Voidborne, my solo-made animated short that went on to pick up more than a dozen festival selections, including a Chroma Award. Not on the animation. On watching a character say one line, over and over, trying to get the read right. The model could draw the face. It could move the mouth. What it could not do was tell me whether the character sounded afraid or just tired. That call was mine, every time, however many versions it took to get there.
AI video models can now copy a performance. They still cannot tell you which performance is the right one, and that judgment call is what directing actually is. That gap, not the resolution or the frame rate, is the real story of where AI filmmaking stands going into 2026.
Not in the sense anyone in a theater or on a set would use the word. What AI can do, convincingly, is transfer a performance. Runway’s Act-One shipped in October 2024 and let a filmmaker record a single-camera take of their own face, no motion-capture rig, and drive a generated or animated character with it. Runway’s follow-up, Act-Two, released to paid plans in July 2025, extended that to full body and hand tracking at 1080p. Runway’s own description is blunt about what moves and what doesn’t: the driving performance “contributes only motion and timing,” not the character’s appearance.
Read that sentence again, because it says the whole thing. Motion and timing means the model reproduces a performance someone already gave it. It does not generate one from a script page, and it does not know whether the take it just reproduced was the right choice out of the ten someone recorded on their phone before picking the one that worked. A human actor made that decision, or a director picked the take afterward. The model is a very good camera operator for a performance that already happened.
Three things, and none of them show up in a model card:
Runway itself frames Act-Two around this handoff, not around replacing it: the pitch is that a director can shoot a rough phone take of a performance to communicate intent on set, fast, rather than skip the human performance step altogether. That is a tool for previs and communication between people who are already directing each other. It is not a director.
Probably closer than people assume, and further than the marketing suggests. Worth separating the two:
Faster iteration on the same reproduction loop is not the same claim as a model that reads a scene and decides what the character should feel.
The confusion is understandable. A clip where the emotion lands feels directed, the same way a photograph that’s well composed feels like someone framed it. But feeling directed and being the output of a directing decision are different claims, and right now the second one still requires a person who can tell good from almost-good.
It moves the bottleneck. For years the limit on a one-person production was execution: could you animate it, light it, composite it, alone, in a reasonable amount of time. AI video tools have genuinely collapsed that part. What’s left standing is judgment, and judgment doesn’t scale the way rendering does. Regeneration being nearly free doesn’t make the right take easier to recognize. It makes it easier to drown in options and lose track of which one was already good.
This is where a lot of the structure work has to happen before a single frame generates, not after. On Voidborne and on later projects I built with ScreenWeaver, the tool I’ve been building for this exact problem, the shot breakdown and beat-by-beat outline happen in the writing and storyboard stage, before generation starts, specifically so the question “what should this character be doing emotionally in this shot” gets answered once, on purpose, instead of forty times by accident in a model’s output queue. The software doesn’t direct the performance for you. Nothing does yet. It just makes sure you’ve actually made the decision before you start paying for regenerations to find out you hadn’t.
Can AI generate an emotional performance from text alone, with no reference video?
Text-to-video tools like Veo 3 will attempt an emotional read from a prompt, and results vary a lot by how specifically the emotion is described. Performance-transfer tools like Act-Two are a different approach: they need an actual human performance, on camera, to drive from.
Will AI replace the director’s job in indie filmmaking?
Not based on anything shipped as of late 2026. It has removed the need for a crew to execute a shot. It hasn’t removed the need for someone to decide what the shot should feel like and whether a given take delivers that.
Is performance-transfer AI useful for solo filmmakers right now?
Yes, concretely so for previs and communicating a performance idea fast, which is closer to what Runway pitches Act-Two for than to full automated acting.
I’m not making an anti-AI argument here, and I’d rather be clear about that than let the framing do it for me. I built a company on the belief that these tools are the reason someone who couldn’t make films before now can. Watching a model reproduce a face, take after take, taught me where the actual work moved to. It didn’t teach me that the work disappeared. It moved from execution to judgment. If you’re making something alone right now, that’s the skill worth building on purpose: not better prompts, but a sharper, faster, more honest opinion about which take is actually the right one.
If you want to see how that shot-by-shot judgment gets built into the writing stage before generation starts, ScreenWeaver’s screenplay and storyboard tools are free to try.
AI video models can now copy a performance. They still cannot tell you which performance is the right one, and that judgment call is what directing actually is. That gap, not the resolution or the frame rate, is the real story of where AI filmmaking stands going into 2026.
Can AI actually direct a performance now?
Not in the sense anyone in a theater or on a set would use the word. What AI can do, convincingly, is transfer a performance. Runway’s Act-One shipped in October 2024 and let a filmmaker record a single-camera take of their own face, no motion-capture rig, and drive a generated or animated character with it. Runway’s follow-up, Act-Two, released to paid plans in July 2025, extended that to full body and hand tracking at 1080p. Runway’s own description is blunt about what moves and what doesn’t: the driving performance “contributes only motion and timing,” not the character’s appearance.
Read that sentence again, because it says the whole thing. Motion and timing means the model reproduces a performance someone already gave it. It does not generate one from a script page, and it does not know whether the take it just reproduced was the right choice out of the ten someone recorded on their phone before picking the one that worked. A human actor made that decision, or a director picked the take afterward. The model is a very good camera operator for a performance that already happened.
So what is a director actually doing that the model isn’t?
Three things, and none of them show up in a model card:
- Choosing the read. The same line, “I’m fine,” can mean six different things depending on pace, pause placement, and what the eyes do half a second before the mouth moves. A director’s job is picking which of those six is true for this scene. A generation tool will give you a version. It will not tell you it’s the wrong one.
- Judging consistency across a scene. One good take of one line is easy. A performance that holds together across a dozen shots, where the character’s emotional state at shot seven has to logically follow from shot three, is a continuity problem that nobody has handed to a model yet.
- Knowing when to stop. Regenerating is free in a way reshoots never were, and that changes the job. The hard part on Voidborne was never getting a version that worked. It was trusting that an earlier take had already been the right one instead of generating past it out of habit, because nothing was stopping me from trying one more.
Runway itself frames Act-Two around this handoff, not around replacing it: the pitch is that a director can shoot a rough phone take of a performance to communicate intent on set, fast, rather than skip the human performance step altogether. That is a tool for previs and communication between people who are already directing each other. It is not a director.
Does this mean AI acting direction is coming, just not yet?
Probably closer than people assume, and further than the marketing suggests. Worth separating the two:
- What’s shipped, today: Act-Two’s motion and expression transfer from a reference video, and Veo 3’s dialogue-synced generation from a prompt. Both are reproduction and timing.
- What’s promised: Higgsfield has talked publicly about moving toward live, in-generation adjustment, changing a character’s expression or timing while a clip is still rendering, instead of regenerating it from scratch. That is a stated direction, not a feature shipped as of late 2026.
Faster iteration on the same reproduction loop is not the same claim as a model that reads a scene and decides what the character should feel.
The confusion is understandable. A clip where the emotion lands feels directed, the same way a photograph that’s well composed feels like someone framed it. But feeling directed and being the output of a directing decision are different claims, and right now the second one still requires a person who can tell good from almost-good.
What does this actually change for indie filmmakers?
It moves the bottleneck. For years the limit on a one-person production was execution: could you animate it, light it, composite it, alone, in a reasonable amount of time. AI video tools have genuinely collapsed that part. What’s left standing is judgment, and judgment doesn’t scale the way rendering does. Regeneration being nearly free doesn’t make the right take easier to recognize. It makes it easier to drown in options and lose track of which one was already good.
This is where a lot of the structure work has to happen before a single frame generates, not after. On Voidborne and on later projects I built with ScreenWeaver, the tool I’ve been building for this exact problem, the shot breakdown and beat-by-beat outline happen in the writing and storyboard stage, before generation starts, specifically so the question “what should this character be doing emotionally in this shot” gets answered once, on purpose, instead of forty times by accident in a model’s output queue. The software doesn’t direct the performance for you. Nothing does yet. It just makes sure you’ve actually made the decision before you start paying for regenerations to find out you hadn’t.
A short FAQ
Can AI generate an emotional performance from text alone, with no reference video?
Text-to-video tools like Veo 3 will attempt an emotional read from a prompt, and results vary a lot by how specifically the emotion is described. Performance-transfer tools like Act-Two are a different approach: they need an actual human performance, on camera, to drive from.
Will AI replace the director’s job in indie filmmaking?
Not based on anything shipped as of late 2026. It has removed the need for a crew to execute a shot. It hasn’t removed the need for someone to decide what the shot should feel like and whether a given take delivers that.
Is performance-transfer AI useful for solo filmmakers right now?
Yes, concretely so for previs and communicating a performance idea fast, which is closer to what Runway pitches Act-Two for than to full automated acting.
The part that doesn’t change
I’m not making an anti-AI argument here, and I’d rather be clear about that than let the framing do it for me. I built a company on the belief that these tools are the reason someone who couldn’t make films before now can. Watching a model reproduce a face, take after take, taught me where the actual work moved to. It didn’t teach me that the work disappeared. It moved from execution to judgment. If you’re making something alone right now, that’s the skill worth building on purpose: not better prompts, but a sharper, faster, more honest opinion about which take is actually the right one.
If you want to see how that shot-by-shot judgment gets built into the writing stage before generation starts, ScreenWeaver’s screenplay and storyboard tools are free to try.