Runway

What Runway is and how it is organised

Runway is an AI research company and creative production platform, also known as RunwayML. It develops its own models and provides access to other providers within the same environment. Its offering combines generation, video transformation, images, audio and project organisation.

Distinguish the model, tool and deliverable. Gen-4.5 generates clips; Aleph transforms existing material; an editor or agent can organise multiple assets. A feature appearing on the platform does not mean every model supports it. This profile describes the offering reviewed on 8 November 2026 and separates production tools from research projects. The distinction helps a team choose a workflow by its required output rather than by a model name alone.

Reference documentation: Runway current creative product.

Founding and company development

Cristóbal Valenzuela, Anastasis Germanidis and Alejandro Matamala Ortiz founded Runway in 2018 after meeting at New York University’s Interactive Telecommunications Program. Its origins combined artistic experimentation and machine learning before content generators became widely established.

The company moved from making creative models easier to use towards developing proprietary technology and production services. On 10 February 2026 it announced a $315 million round led by General Atlantic to expand its world-model work. Funding explains its research expansion but does not measure the accuracy of each tool. Product selection requires checking which parts of that research are available and which controls accompany them. An announced direction and an operational capability are different kinds of evidence.

Reference documentation: NYU: Runway founders · NYU: company founded in 2018 · Series E announcement.

Stable Diffusion and the evolution of Gen models

Runway collaborated with CompVis and Stability AI on Stable Diffusion, based on latent diffusion. That historical contribution does not make every current product the same model or imply publicly available weights. Gen-1, introduced in February 2023, transformed an input video; Gen-2 expanded text-guided generation and reached web and mobile in 2023.

Gen-3 Alpha followed in 2024 and Gen-4 in 2025, with increasing emphasis on motion and consistency. The central technical distinction is temporal: generating attractive individual frames is insufficient to preserve an object, action or scene throughout a shot. Later families offer different controls; an old tutorial may describe parameters unavailable in today’s selected model. Preserve model names when documenting a production method.

Reference documentation: CompVis Stable Diffusion README · Gen-1 research announcement · Path to Gen-2 · Runway changelog.

Gen-4.5: from text or image to motion

Gen-4.5 supports text-to-video and image-to-video. The reviewed guide documents clips of two to ten seconds, base 720p output and web access from Standard. For image-to-video, the reference establishes the visual starting point while text guides motion, camera behaviour and progression.

A useful brief distinguishes what the subject does from what the camera does. A person may remain still while the framing moves closer; confusing those actions produces a different narrative. Begin with one action and add complexity later. Review the complete interval because correct opening and closing frames may conceal an impossible transition. Duration should allow the action to be understood without acceleration that changes its meaning.

Reference documentation: Runway — Creating with Gen-4.5 · Runway — Image to Video Prompting Guide.

References, characters and consistent images

Gen-4 Image References combines up to three references to guide characters, objects or style. The broader reference system allows assets to be saved for reuse and shared with a workspace. Media types depend on the model; a view of a 3D object can serve as a visual reference without becoming a complete geometric simulation.

For a series, prepare a visual record of the character and environment. Compare clothing, proportions and objects between shots before animating them. Give each reference a clear role: identity, composition or appearance. Mixing too many intentions makes correction difficult. If a detail must repeat exactly, consider adding it as a controlled post-production element while generating the rest of the scene.

Reference documentation: Gen-4 Image References · Reference media guide.

Aleph 2.0 and Edit Studio

Aleph 2.0 modifies existing video through instructions and an edited reference frame. Edit Studio previews the change in an image before applying it. The current presentation documents clips up to thirty seconds at 1080p and changes across multiple shots.

Editing an existing piece is useful when its structure is already approved. Specify what changes and retain an earlier copy for comparison. Inspect areas where an object is obscured, crosses another element or leaves the frame. Preservation claims should be checked against the actual result. A colour change may appear local yet affect shadows or reflections. For corrections requiring repeatable precision, compare the effort with conventional editing rather than assuming regeneration is always quicker.

Reference documentation: Aleph2.0.

Act-One, Act-Two and conversational characters

Act-One appeared in October 2024 to transfer performances to characters. Act-Two extends that approach: it combines a performance video with a character image or video and supports up to thirty seconds. Performance-driven gesture control applies to character images rather than every input type.

Runway Characters is a different capability: real-time conversational avatars through the API, with configurable voice, personality and knowledge. This differs from producing a performance clip. For a corporate character, distinguish image and voice permission, approved scripts and responses it may improvise. A convincing expression does not establish that an answer is correct; in training, the content and its review must remain identifiable. Choose the format according to whether the message is fixed or interactive.

Reference documentation: Act-One announcement · Performance Capture with Act-Two · Characters API.

Narration and audio work

Speech synthesis turns a script into audio using a selected voice. Emotion, pacing and reference controls vary between models. Custom voices require a compatible arrangement, documented from Pro, and a model supporting them. Audio resources can also be incorporated into Agent and Workflows.

A multilingual video requires reviewing each complete narration. Acronyms, figures and equipment names may be pronounced unexpectedly. Prepare a short glossary and check that translation preserves the script’s action. Synchronise captions after accepting the audio. Retain a version without music when the mix needs adapting for a room, course or display with different acoustic conditions. Clear delivery should be assessed separately from the voice’s apparent realism.

Reference documentation: Generating speech from text.

Agent, workflows and assembly

Agent develops assets through conversation and can reuse Brand Kits, Skills and workflows. The environment added timeline assembly within Agent in July 2026. App Studio and Runway Ads subsequently expanded the offering towards creative applications and advertising variants.

For recurring production, define a brief including audience, duration and mandatory elements. A workflow can prepare versions, but separate creation, acceptance and publication. Review the complete assembly: an image, narration and headline that are individually correct may contradict each other when combined. Retain decisions the team wants to repeat and specify what requires review for each new brief. Conversation coordinates resources but does not replace shared editorial criteria or a clearly identified approved version.

Reference documentation: Runway Agent · Runway changelog.

Export and handover to other editors

MP4 is the usual output. Gen-4.5 in Tool Mode and Aleph 2.0 in Edit Studio offer ProRes or PNG sequences on Max, Unlimited Legacy and Enterprise. Format is selected before generation; Agent and Workflows deliver MP4. Edits limited to a specific range also have output restrictions.

Choose a format for the next production step. A compressed file may suffice for quick review; frame-based compositing may benefit from a sequence. Higher-quality export does not correct content inconsistencies. Supply duration, frame rate and approved version with the clip. Check sound, colour and synchronisation in the destination program before closing the project, especially when several providers contribute material to one edit.

Reference documentation: Professional video export.

Platforms, library and API

Runway provides web access and official iOS and Android apps; specific features may differ. Uploaded or generated assets are private by default, although projects may share them with members. Runway Dev offers image, video and audio APIs through projects, keys and credits.

A library needs understandable organisation: originals, rejected trials and approved deliveries. In an integration, retain outputs independently of temporary links and record the requested model. The API guide warns that removing a person from a project does not itself revoke their keys. Assign access and budget owners, and test what happens when generation fails or is interrupted before automating an entire collection. Technical delivery should remain distinct from editorial acceptance.

Reference documentation: Managing assets · Runway iOS listing · Runway Android listing · Runway Dev overview · Dev setup.

Plans, credits and enterprise controls

Individual offerings include Standard, Pro and Max alongside limited free access. Team groups two to nine people with pooled credits; Enterprise adds contractual arrangements, SSO and organisational controls. Its documentation provides administration of external models and commitments against training on Enterprise customer content.

Estimate the cost of a finished piece: trials, corrections and export. Runway explains that a completed generation consumes credits even if its result is disliked. Agent can show its plan and estimated cost before generation. For an adoption trial, set a budget and compare accepted shots, review time and rework. Do not extrapolate Enterprise conditions to a personal account or treat the first generation’s cost as the final production cost.

Reference documentation: Current pricing · Team and Enterprise selection · Enterprise third-party models FAQ · Output mismatch and credits.

Studios, festivals and audiovisual creation

The Lionsgate collaboration began in September 2024 around a model adapted to its catalogue. In June 2026 the companies announced an expansion, an equity interest and planned joint projects. AMC Networks announced marketing and television-development uses in June 2025. These are specific agreements rather than general permission to reuse their catalogues.

Runway also organises an AI film festival and the Gen:48 short-film challenge. It collaborated with Tribeca in 2024 and IMAX screenings in 2025. These initiatives show how scripts, direction, generation and editing combine. To learn from a work, analyse specific decisions: what the tool contributes, what editing resolves and which material comes from filming or other sources. Finished work alone does not reveal its revision process.

Reference documentation: Lionsgate partnership2024 · Lionsgate expanded partnership2026 · AMC Networks partnership · Tribeca collaboration2024 · AIFF IMAX2025 · Gen:48.

Game Worlds and interactive research models

Game Worlds launched as a beta in August 2025, with nonlinear narratives accompanied by images and Chat and Comic modes. GWM Worlds 2, shown in September 2026 as a research preview, generates interactive audiovisual environments. The company acknowledges memory limits and degradation under difficult camera movements.

Solaris explores visually generated interfaces and remains in early access; its documentation identifies text, coherence and accessibility challenges. These directions extend Runway beyond fixed clips. For OSH, an interactive experience would require validating objectives, actions and consequences before using it for training. A world responding to instructions does not establish that it reproduces a facility’s physical conditions or equipment behaviour. Research demonstrations should not be treated as validated workplace simulators.

Reference documentation: Game Worlds2025 · GWM Worlds2 researchpreview · Solaris research.

Rights, training-data debate and provenance

Runway’s FAQ permits commercial use of creations without requiring platform attribution. This does not remove third-party rights in references, voices or people. In July 2024, 404 Media reported a leaked document concerning collection of YouTube videos for training. That is journalistic reporting, not a judgment on every model or output.

The company documents input and output moderation and C2PA provenance signals. Metadata helps identify processes but does not turn a scene into a proven event. To publish a recreation, retain sources and versions, identify its nature and review authorisation for supplied material. Clarity about origin and purpose is particularly important when a video resembles an authentic recording. Commercial permission and documentary authenticity are separate questions.

Reference documentation: Commercial use FAQ · 404Media public report,25July2024 · Runway safety.

Practical applications in OSH

In occupational safety and health, it can support awareness pieces and fictional scenes used to discuss risk situations. The proposed role is illustrative: encouraging discussion about decisions and behaviour through a validated script and technical review of every scene, rather than claiming to reproduce a real event accurately.

It can support audiovisual drafts, discussions about observing the environment or openings for learning activities. The objective is for viewers to explain what they see and what remains unknown, rather than accept a clip as a recording of an accident.

It can also help plan shots that will later be filmed authentically. A draft lets teams agree pacing and focus before involving people or equipment, while sensitive operations remain assigned to reviewed real footage.

Worked example and result checks

Example: create a brief fictional sequence about an obstructed walkway. The team reviews the script, identifies the scene as a recreation and uses questions to ask participants how they would report the problem at their own site.

1. Define the discussion objective and a short script. Do not introduce physical consequences or injury figures as if they had been calculated.
2. Prepare a starting image without relevant defects and describe one understandable action. Separate camera movement from subject movement.
3. Review the entire clip, pausing at key points. Look for objects appearing, disappearing, passing through others or changing shape.
4. Reject a sequence communicating an incorrect manoeuvre. If used to discuss ambiguity, state that explicitly.
5. Edit the clip with the approved question and context. Check captions, reading speed and the absence of details identifying real people.

This is a proposed learning exercise. Connect the group’s answers with actual workplace conditions afterwards.

Limitations and conditions of use

Generative video may show impossible movements or change equipment between frames. Do not use it as accident reconstruction evidence or an operational demonstration without review. Check continuity, safeguards, instructions and sound, and explicitly distinguish fictional scenarios from documented facts.

Errors in a starting image may become more obvious when animated. Camera movement can also hide details needed to understand the scene. Do not inspect only the first and final frames: the intervening motion may change the action portrayed.

Visual continuity does not establish physical continuity. For a manoeuvre, evacuation or vehicle–pedestrian interaction, conditions should come from actual work and technical judgement, rather than whatever movement the generator produces. Preserve the distinction in captions and accompanying discussion.

Reference documentation: Runway — Creating with Gen-4.5 · Runway — Image to Video Prompting Guide.

When to choose it and how to complement it

Runway fits motion exploration and video storyboarding. Descript can support later editing of narration and recorded material. Google Flow offers another workflow for scenes and references.

Test the same script in the intended format. Evaluate how many shots need rebuilding and the technical review effort involved, alongside cinematic appearance. Keep reasons for rejecting shots so another editor does not inadvertently reuse them.

Frequently asked questions

Can it reconstruct how an accident happened?

Not by itself. It can create an illustrative recreation, which must be distinguished from a reconstruction grounded in evidence.

Are isolated frames enough for review?

No. Watch the complete sequence and check continuity, actions and intermediate details.

When is real filming preferable?

When the lesson depends on a particular manoeuvre, item of equipment or condition requiring verifiable accuracy.