Runway Gen 4 AI Video is Blowing My Mind! First Impressions

Runway ML’s Gen 4 arrives with a clear jump in video realism and control—especially for character motion, physics-like effects, and background consistency—while still showing familiar failure modes like occasional anatomy drift and imperfect object interactions. In early demos and hands-on tests, the model produces scenes where lighting behaves plausibly (light passing through fabric), motion reads cleanly (cloak movement, bird wing flaps), and backgrounds stay stable without the “mushy” artifacts common in earlier generations.

The most striking improvements show up in how Gen 4 handles complex movement and coherence across shots. A character in a sandstorm keeps believable physics in clothing and lighting, while a low-depth-of-field forest walk pairs a moving subject with a readable environment. Animal motion stands out too: a vulture-like creature spreads and contracts wings with a level of timing that feels more physically grounded than typical AI motion. Even abstract sequences—like jellyfish-like movement or large-scale transformations—tend to look sharper and more defined, with less of the scribbled, hallucination-heavy texture that can undermine credibility.

Hands-on testing adds nuance. Gen 4 supports image uploading and offers multiple aspect ratios (including 16:9, 21:9, 4:3, and portrait options) with generation lengths up to 10 seconds. When prompted to animate a person from an uploaded image—staring into the camera, sprinting away, and pulling the camera upward—Gen 4 follows the intent well, including dust kicked up during the run. But it still struggles with strict continuity: an “armless” test subject gains the missing arm once the character sprints away, and a 10-second run repeats the same issue. Upscaling to 4K can improve detail, yet it doesn’t fully fix motion artifacts; warping can still appear when frames are paused.

Vehicle and physics prompts show both progress and limits. A “car speeding away with smoke and a fire trail” prompt generally works better in Gen 4 than in Gen 3, with more consistent fire and overall quality. Still, camera rotation can fail to match the exact instruction, and motion can look slightly washed as the subject shrinks. A “truck smashing through a wall” test is hit-or-miss: sometimes the truck materializes inside the room rather than breaking through, then later generations do manage a more convincing smash—suggesting that precise physical causality remains difficult.

The model also leans into stylized and narrative-friendly animation. VHS-like “creepy footage” prompts maintain the look while delivering unsettling close-ups (a lemon held to the camera). In 3D animation tests—like a robot riding a rocket to the moon—Gen 4 can preserve character consistency and add emotion through body language, even when the prompt is relatively simple. Where 2D animation appears weaker, the broader takeaway is that Gen 4 is carving out a stronger niche in realism, 3D character performance, and cinematic motion, while leaving room for future refinement in strict anatomy control and deterministic physics.