Compare video AI through revision: references, sound, and editing

A good shot is followed by a revision request

A team makes a short product video. The generated clip looks polished and receives a good response in a concept meeting. Then someone asks for a brighter background. Regenerating changes the product label and the shape of the container too. This hypothetical example captures a practical question in video AI selection: producing an attractive first clip differs from correcting it while preserving the parts worth keeping.

This article compares material published by OpenAI, Google, Adobe, Runway, and Luma by March 28, 2026, through reference inputs, sound, changes to existing footage, and editing handoff. It compares features and limitations, rather than measured rankings of image quality or production speed. My judgment is that obtaining an acceptable revised version should be the unit of work being evaluated.

The person selecting the product image, the person creating movement, and the person checking dialogue and labeling need different things preserved. Faster generation does not translate proportionally into faster production if every revision requires everyone to check everything again. Start with what should change, what should remain, and who must review it.

Distinguish generation mechanics from production revision

OpenAI's 2024 Sora technical report describes compressing video into smaller units of space and time, then using a model to recover video from noisy inputs. Video generation must address continuity of objects and motion over time, alongside individual frame appearance. However, a model's internal refinement process differs from an editing operation that lets a creator change only the background.

Runway's Gen-4.5 announcement describes improvements in motion and consistency while acknowledging limitations, including effects preceding causes and objects disappearing. Neither source establishes the internal design or revision success rate of every product. For the hypothetical product clip, examine the whole shot: does the container retain its shape after being hidden by a hand, and does the action of lifting it remain coherent?

Specifying appearance with a reference image, preserving that appearance within a clip, and preserving it in the next shot are separate checks. Treating all three as “consistency” hides the precision production requires. Define which details must persist, and across which moments.

Can references reduce repeated explanation?

The Runway Gen-4 announcement describes combining visual references and instructions to maintain subjects, locations, and styles. For three shots of the same product, sharing an asset gives a team an alternative to repeatedly describing its shape and color. The existence of references, however, does not guarantee matching label text or dimensions.

Version and operation matter too. Runway's changelog records first-frame image input for Gen-4.5 on January 21, 2026. A higher-tier model's launch does not imply that every control from earlier models was available immediately. Record the operation being tested, such as image-to-video, references, or keyframes, alongside the model name.

When revising the background, assess the requested brightness change separately from the product shape, color, and labeling that should remain. Did the reference reduce repeated input, and what detailed repair still remained? Its value includes making preservation requirements easier to communicate and inspect, alongside producing a similar-looking image.

Sound and aspect ratio create their own revision tasks

Google's October 2025 Veo 3.1 and Flow update describes adding audio to generation using ingredients, first and last frames, and clip extension. These features were described as experimental. Generating environmental sound and dialogue with the picture provides a reason to test audiovisual concepts.

Separately, test whether correcting a number in the dialogue preserves the picture. Joint generation does not establish independent replacement. Compare changes to the speech, lip movements, and the way the product is held, including alternatives using recording or conventional editing.

The January 2026 update describes native 9:16 output from ingredients and upscaling to 1080p and 4K. Generating a vertical format and increasing resolution are different operations; the upscaling options are limited to Flow, the API, and Vertex AI. If cropping a horizontal clip would cut off the product, assess native vertical generation separately from keeping the same product across multiple formats.

Keep an existing performance while changing its appearance

Changing existing footage offers an alternative to generating movement from scratch. Luma's Ray3 Modify announcement describes video-to-video transformations of wardrobe, environment, and lighting, using an original performance with keyframes and character references. If the team can film the product being lifted, retaining that timing as input becomes an option worth comparing.

Using original motion still does not establish that every detail will survive. The announcement discusses adjusting modification strength and possible degradation when aspect-ratio conversion crops input footage. Check occlusion by hands, subject boundaries, and the duration of the original movement.

The Ray3.14 announcement claims fourfold speed and threefold lower cost against Ray3 at 720p. This is a vendor comparison under specified conditions, rather than a measurement of total cost to obtain an acceptable revision. The launch user guide also explicitly lists unsupported character-consistency controls and reference images. A shared Ray3 family name should not be used to bundle features from different versions.

Separate generation, modification, extension, and assembly

Adobe announced general availability of Generative Extend in Premiere Pro in April 2025, extending existing video and audio. Holding an expression slightly longer or filling a few missing transition frames is a different task from generating a whole new scene.

The December Firefly update describes text-based editing of existing clips with Runway's Aleph and a public beta of a browser video editor with multiple tracks. The Firefly product includes Adobe models, partner models, and an editing interface. Comparing brands alone can obscure which feature and usage conditions are being assessed.

This classification organizes the production tasks discussed here. One product may cover several; it is not a proposal to buy four products. Identifying the actual bottleneck can also reveal occasions when conventional editing serves the task better than adding another generator.

Continued availability is also a production requirement

OpenAI's Sora 2 launch presented video and audio generation alongside creation and sharing involving users' likenesses and voices. Model capabilities need to be distinguished from the availability of the product using them. On March 24, Sora's official account announced its intention to discontinue the app, saying that further app and API timelines would follow.

This article's cutoff is March 28, so shutdown dates established later are not treated as information known then. Nevertheless, a product with an announced discontinuation cannot be recommended as a new adoption destination with confirmed continuity. Reusing saved footage in another editor and carrying reference assets and instructions forward also become selection criteria.

A model making a good clip differs from a team being able to revise the same project next month. Lost assets or an inability to reproduce revisions can add migration work to production. Examine storage and editing handoff alongside generation features.

Usable source material and visible provenance are distinct

Adobe's Firefly Video Model announcement describes training on permitted material, including licensed and public-domain content, and adding Content Credentials to outputs. Commercial safety is Adobe's positioning; this article does not guarantee the permissibility of an individual asset. Selecting a partner model inside Firefly also does not establish identical training material or usage conditions.

OpenAI's Sora safety description similarly distinguishes consent mechanisms for likeness and voice from provenance signals. Provenance information does not establish that usage conditions for an uploaded person, logo, or product image have been met. The production team should establish the input conditions and check that required disclosures remain after export and sharing.

The hypothetical product video needs both accurate labeling and suitable source imagery. Identifying reviewers during asset selection can help avoid discovering an issue just before publication and rebuilding the work. This is a production proposal, rather than a determination of legal permissibility.

Compare the work required to obtain an acceptable revision

For a trial, generate a short shot using the same product image, duration, and aspect ratio, then change only background brightness. Define preservation criteria for shape, color, labeling, motion, and dialogue beforehand. If a feature difference prevents equivalent inputs, record that gap rather than mixing results from different tasks.

Keeping only the best clip hides discarded generations and repair attempts. Repeat the task under a fixed budget or attempt limit and retain rejected outputs in the record. Include generation waits, the editor's repairs, the reviewer's checks, and final assembly. Comparing conventional edits of existing footage also separates generation price from the work of completing the production asset.

A concept clip communicating a movement direction and a public asset requiring precise product labeling may lead to different choices. A rough clip may align intentions for the former; filming or compositing may better satisfy the latter. The question is whether the necessary conditions are met and the next revision can be handled.

These trials are proposed, rather than measured time savings from the products discussed. Turning advances in video AI into production value requires looking beyond how quickly an idea appears to the work of revising and retaining the chosen one. The ability to absorb the next correction deserves a central place in selection.