Runway unveils Gen-3 Alpha to improve AI-generated video

On June 17, 2024, Runway introduced Gen-3 Alpha, a model for generating video from text and images. On its research page, the company said the examples were generated without modifications and claimed better fidelity, consistency and motion; these are vendor claims that an independent test must turn into repeatable results.

The announcement matters because Runway was one of the first companies to bring generative video to creators and businesses. While OpenAI has kept Sora closed to general access and Google introduced Veo in May, Runway is seeking to hold on to its position with a model designed not just to produce eye-catching images, but to give users greater control over directing a scene.

Greater consistency from one frame to the next

Generating a still image with AI has become an everyday task. Creating a credible sequence remains considerably harder. Video requires characters, objects, lighting and perspective to remain consistent for several seconds. When the model fails, a hand changes shape, an object disappears or the movement seems to violate the laws of physics.

Runway says Gen-3 Alpha improves fidelity, consistency and motion compared with Gen-2, its previous generation. The model can work from written instructions and reference images, a combination that matters for audiovisual production: creators can provide a starting composition, character or aesthetic and then request a specific action.

The company is also emphasizing control. Rather than simply describing a scene—for example, a person walking down a rainy street—the goal is to let users approximate decisions typically made on a film set: how the camera moves, which element must be preserved and what action the subject should perform. That does not amount to replacing a camera or guaranteeing an exact result, but it reduces the unpredictability of early AI video tools.

The race accelerates after Sora and Veo

Runway’s move comes amid intense competition. OpenAI showed Sora in February with long, detailed sequences that made generative video one of AI’s major frontiers. In May, Google introduced Veo, its own model for creating high-quality video from text prompts.

Runway has a practical advantage: it has spent years integrating these models into an editing and creation product used by professionals. The company launched Gen-1 in 2023 to transform existing videos using prompts and introduced Gen-2 soon afterward, with generation from text or images. Gen-3 Alpha is, according to Runway, the first model in a new series trained on infrastructure designed for large-scale multimodal training—that is, capable of learning relationships between different types of content, such as images, text and video.

That integration matters as much as the model’s raw quality. A studio or agency needs more than a spectacular clip: it needs to reproduce a style, adapt formats, test variations and move the result into an editing workflow. The commercial value of these platforms will depend on whether they can perform those tasks with sufficient speed and predictability.

A powerful tool, but not an automated film set

Gen-3 Alpha does not eliminate the usual limitations of the technology. Video models still struggle with lengthy actions, complex physical interactions, legible text within images and completely stable identities for characters or objects. As a result, their most immediate uses are short shots, previsualizations, social media content, advertising and effects that can later be reviewed or edited.

Questions also remain over copyright and the provenance of training data. For professional customers, it will not be enough for a model to produce better videos: they will need to know under what conditions they can commercially exploit the material and how to protect their brands, characters and files.

Runway said the rollout would begin that week for its creative partners and enterprise customers, ahead of broader availability. Its reception will depend on a simple test: whether the model can turn relatively precise instructions into clips that require fewer attempts and fewer corrections than current systems. In generative video, that difference could determine which tool moves from demonstration to daily production.

A gallery of examples is not an evaluation

A launch page shows clips chosen by the system’s maker. Measuring quality requires defining a scene set in advance and retaining every attempt. Include human motion, crossing objects, text in the frame, camera changes, occlusion and an edit to material you own. Showing only the best output removes exactly the information needed to estimate reliability.

Evaluation can be divided into prompt fidelity, temporal coherence, control and production utility. Fidelity asks whether subjects, actions and setting appear. Coherence follows identity, anatomy and geometry across frames. Control checks whether a modification changes only what was requested. Utility adds waiting time, resolution, duration, cost, export options and hours of manual correction.

Failures belong in the report

A clip may look correct on one playback and reveal jumps when paused. Inspect frames, edges, reflections, hands, text and partly hidden objects. Repeat the same prompt to learn the distribution and alter one variable at a time to measure control. The minimum record preserves text, input image, version, parameters, seed when available, time and every output.

Technical capability must also be separated from permission to use it. A service can generate a persuasive aesthetic without adequately explaining data provenance, commercial terms or the handling of uploaded files. Vendor claims about safeguards and provenance credentials need to be checked in the exported file and publication workflow, not merely read in a description.

The transferable skill is to turn a video demo into a repeatable trial. Define scenes, keep failures, score each axis and count the work that follows. If a claim survives only by choosing the most favourable clip, it describes promotional selection, not the model’s normal quality.

A comparison needs a baseline

“Better than the previous version” is testable only when both receive the same inputs under equivalent conditions. Preserve resolution, duration, aspect ratio and attempt count; if one tool accepts a reference image and another does not, record it as a product difference instead of hiding it. Results should show success rates and failure types, not one ranking.

For professional use, add a continuity test across shots. Generate the same character in different scenes and inspect clothing, features and scale; then attempt a local correction without regenerating the rest. Many models create an attractive clip and fail to sustain a world across a sequence. That gap between shot and project explains utility better than an isolated demo.

Finally, verify provenance. If the service promises Content Credentials, download the file, inspect its metadata and test what happens after editing or publication. A credential that disappears in the normal workflow does not provide viewers with the advertised transparency. Recording that loss is part of evaluation, not an export detail.

The evaluation record ends with a use decision: sketch, final shot or discarded material. Note which errors a person can correct and which force regeneration. That boundary prevents a system useful only for exploring ideas from being called a production tool.

Value appears when the method survives the least flattering example, not only the chosen best clip.