Runway announces Gen-3 Alpha

Background

Text-to-video models generate a sequence of frames from a written prompt. They learn from large collections of video paired with descriptions, and at generation time a prompt steers the model toward a matching clip. Two properties are especially hard to get right. The first is consistency: an object, person or scene should not change shape, texture or identity from one frame to the next. The second is motion: movement should look plausible rather than sliding or morphing. Runway had already released an earlier generation of its video model, Gen-2, and Gen-3 Alpha was presented as an improvement on it ^[1].

What was announced

On 17 June 2024, Runway announced Gen-3 Alpha ^[1]. The company said the model was trained jointly on video and images, on new large-scale multimodal infrastructure ^[1]. Training on images as well as video is a design choice with a practical rationale: still images are far more plentiful than video clips and carry fine visual detail, so a model that sees both can learn appearance and motion from different sources. The "multimodal" part refers to the infrastructure handling more than one kind of data — here, video and images — rather than to a single input format.

Key details

Runway described Gen-3 Alpha as improving on Gen-2 in fidelity, consistency and motion ^[1]. Fidelity covers how sharp and detailed the frames look; consistency covers whether subjects stay recognisable across frames; motion covers how believable the movement is.

The model also added fine-grained temporal control ^[1]. Temporal control means directing what happens at particular moments within a clip, rather than describing the whole clip with a single prompt and accepting whatever unfolds. Finer control matters because video generation is not only about producing a good-looking frame but about arranging events in time.

Two further features were part of the announcement. Gen-3 Alpha came with an in-house moderation system ^[1], meaning the filtering of prompts and outputs was built and run by Runway itself rather than left entirely to outside tools. It also carried C2PA provenance ^[1]. C2PA is an open technical standard for attaching signed metadata to media, recording how a piece of content was made and edited; for generated video this gives viewers and platforms a way to check that a clip came from a particular tool and was not altered afterwards without record.

Availability

At announcement, access was limited to a closed alpha with select testers ^[3]. By 1 July 2024 the model had become available to users ^[2], moving from that closed alpha to general availability on the same date ^[3].

Access required a paid subscription starting at 15 dollars per month and used a credit-based billing system ^[3]. In credit-based billing, each generation consumes a number of credits, so heavier or longer use draws down the balance faster than light use; the subscription buys an allowance of credits rather than unlimited generation.

Why it matters

Gen-3 Alpha sits at the point where video generation moved from producing short, obviously synthetic clips toward tools meant to be directed. The emphasis on consistency and motion addresses the two failures viewers notice first, and fine-grained temporal control is what turns prompting into something closer to shot planning ^[1].

Attaching C2PA provenance and running an in-house moderation system at launch reflects a second concern: as generated video becomes harder to tell apart from filmed footage, the question of where a clip came from becomes as important as how it looks ^[1]. Provenance metadata does not judge whether a clip is true, but it does record which tool produced it and whether it was subsequently edited.

The pricing structure is also informative. Charging by subscription with credits ties cost to usage, which reflects the fact that generating video is computationally expensive relative to generating text or a single image ^[3]. For users, that means the practical limit on how much video they produce is set by their credit balance as much as by their ideas.