Runway’s Act-One changes how creators animate digital characters. The tool bypasses traditional motion capture hardware and complex facial rigging processes. Creators record a video of an actor using a single-point camera like a smartphone and then select an AI character to map the performance onto. This method replaces expensive multi-million dollar motion capture stages and the manual cleanup work performed by teams of professional animators. Act-One provides the most efficient path for digital human performance by using deep structural understanding to interpret spatial coordinates and facial topography. Project Luxo, a research initiative by Runway, aims to spark conversations about where AI-generated media stands with audiences in terms of realism and emotional resonance. Jamie Umpherson, the chief creative officer at Runway, says the technology reaches an inflection point where the technology no longer has to be the story. A single staff member produced a realistic ten-minute film about a boy lost at sea in three weeks for $4,000. In a different test, a staff member created an animated segment about pigeons in just four hours. Runway also released an AI-generated ad for a fake watch company that reached 100 million views on Instagram. You already know that traditional VFX requires massive budgets and specialized crews. The technology captures subtle human quirks like a nervous lip twitch, a sudden smirk, or a slight wince. This capability helps close the chasm between authentic and fake media, the uncanny valley.
Netflix uses generative AI in approximately 300 programs this year. The company used text-to-video technology to create a building collapse scene in "El Eternauta" ten times faster than traditional methods. While Netflix uses these tools to increase efficiency, Disney filed a lawsuit against AI company Midjourney alleging the creation of a "bottomless pit of plagiarism." Despite this public stance, Disney runs its own AI experiments behind closed doors. Netflix also acquired Ben Affleck’s AI studio InterPositive to speed up its use of artificial intelligence in filmmaking. The company operates an in-house studio called INKubator for AI-assisted animated short-form content. Netflix reported $12.56 billion in revenue for the second quarter of 2026, which grew 13% compared to the previous year. This growth comes as Netflix faces an 8% decline in daily engagement per subscriber. Disney manages much tighter margins, with its streaming segment operating at only 5% profit. These financial realities force studios to seek ways to reduce the soaring costs of visual effects in a $200 billion industry. The global text-to-video AI market might reach £1.33 billion by 2029. The industry must also compete with the 1.2 billion new internet users who access storytelling through vertical video formats. Netflix also expects to double advertising revenue to about $3 billion in 2026.
The technology works by analyzing facial geometry and speech timing to map nuances onto a target character. Users provide two inputs for the system to process.
| Input Type | Description |
|---|---|
| Driving Video | A recording of a human actor using a basic camera |
| Character Reference | A still image of a photorealistic or stylized character |
Act-One captures micro-expressions like a sudden smirk or a nervous lip twitch. Because the model interprets underlying geometry instead of a rigid 1:1 copy, an actor with one facial structure can drive a character with different proportions like a stylized monster, a fantasy creature, or a sci-fi cyborg. The technology analyzes the structural and temporal dynamics of the performance, mapping geometry, gaze, and muscle movement without needing traditional 3D bone rigging. This preserves the psychological weight of the original actor. However, facial capture tools still face technical failures. Users of Metahuman Live Link Face report that the left side of faces often fails to move or characters blink from the wrong eye. One user found that the Arkit Mapping poses remained incorrect even after twenty calibration attempts. This user had to manually rename all pose assets like EyeBlinkLeft and EyeLookDown to match the exact names used in the default Metahuman project to fix the error. Can the industry fully bridge the uncanny valley without more stable facial mapping?
