Runway unveils Gen-3 Alpha

Runway announced Gen-3 Alpha, a new video-generation model the company described as “a major improvement in fidelity, consistency and motion” over its predecessor, Gen-2. The model powered text-to-video, image-to-video and text-to-image generation, along with tools including Motion Brush for localised animation, camera controls and a “Director Mode” for shot composition, and Runway said it was trained jointly on video and image data on new infrastructure built specifically for large-scale multimodal training.

Runway’s announcement emphasised training on “temporally dense” captions, which it said allowed finer control over how a generated clip changed over time and supported more precise keyframing and scene transitions than earlier text-to-video systems. It also highlighted the model’s handling of human characters, claiming improved photorealism in expressions, gestures and emotion, and said outputs required no post-production cleanup to reach a usable state — a claim the company did not support with quantitative benchmarks. The release included a visual-moderation system and adopted C2PA content-provenance labelling, intended to let viewers verify that a clip was AI-generated, alongside custom model versions offered to enterprise partners.

Access rolled out gradually: Runway made Gen-3 Alpha available first to a waitlist and its creative-partners programme, then to paying subscribers over the following weeks, generating clips of roughly five to ten seconds at 720p. Gen-3 Alpha arrived as competition in AI video intensified: OpenAI had unveiled Sora four months earlier without any public access, and rivals including Luma AI and Kuaishou’s Kling were shipping competing video models over the following months. Because Runway’s model reached paying customers within weeks rather than staying in an unreleased research preview, it became one of the more widely used commercial text-to-video tools of 2024, cited in subsequent debate over AI-generated video’s use in advertising, film pre-visualisation and the disputes over training data that followed generative-media releases more broadly.