Mirage combines Captions video editing with generated AI presenters

Mirage is the AI video company behind Captions. Its offer combines an application for editing and producing videos with its own video research and models. The useful buyer question is whether a recorded or generated presenter can become a consistent stream of approved videos, including the corrections and format changes that follow the first attractive draft.

01 / ProductCaptions is the application, Mirage is the company

The official company page says Captions rebranded as Mirage in 2025 as its work expanded to foundation models and an API. It still builds the Captions application. That distinction helps resolve older tutorials and subscriptions carrying the Captions name without inventing a second corporate identity. This blueprint focuses on the publicly documented creator workflow rather than assuming that every model or API has identical access terms.

The AI Edit product page describes applying styles to recorded footage and requesting changes in natural language. Its examples include added B-roll, transitions, music, captions and adjustments to recording quality. These are editing operations around a message the team must still own. A rapid sequence can look finished while changing what the original speaker appears to emphasize.

For videos that begin with a script, the avatar page identifies Mirage Avatar X as the underlying model. Captions offers a library presenter, a prompt-created character or a personal AI twin. Mirage describes the model as generating speech, movement and expression together. That is a vendor description of the generation approach, not independent evidence that every performance will look natural or maintain identity equally well.

02 / AudienceA recurring communication job is the strongest fit

A founder recording product explanations may want consistent captions and editing without relearning a timeline for every clip. A small marketing team may need several versions of a message with different openings or presenters. In both cases the useful starting point is approved source material and a known audience, rather than an instruction to make something viral.

An agency can evaluate the product around client review. The client may approve the overall style but object to a background image, a shortened sentence or an overenthusiastic delivery. The tool becomes useful when those specific corrections can be made predictably and a reviewer can recognize the latest version.

It is less suitable as an unquestioned substitute for documentary evidence. A synthetic presenter can explain an idea, but cannot truthfully serve as a customer who experienced a result that never occurred. For a technical demonstration, preserve real product footage where the viewer needs to understand actual behavior. Visual polish and evidentiary value are different properties of the finished video.

03 / WorkflowProposed pilot: revise a launch clip in two routes

Choose a short launch explanation with one factual claim, one demonstration and one next step. Write an approved script and mark any language that must remain exact. This proposed pilot uses the same message in two routes: editing a recording of the real speaker, then generating a presenter from the script. The purpose is to discover which route is easier to maintain for this specific recurring job.

For the recorded version, use AI Edit with a restrained style. Inspect the output before requesting extra effects. Listen for cuts that remove qualifications or change a sentence's meaning. Treat any automatically chosen B-roll as a suggestion: replace a generic illustration when an original explanatory visual or actual authorized footage would communicate more accurately.

For the generated version, select the presenter and approve the script before exploring cosmetic variations. The avatar documentation distinguishes a synthetic character from a twin made from a real person. Keep the presenter's identity clear in the production brief. A team using a colleague's likeness should have a durable approval process for what that likeness may say, rather than treating consent for one recording as approval for every future script.

Next, ask the reviewer to change an important sentence and a visual detail. Compare whether the change affects only the requested element or alters nearby timing, expression and captions. Maintain the original and revised exports together during review. The measurement is not simply how quickly a command completes; it is whether a reviewer can confirm that the intended correction arrived without introducing another error.

Finally, produce the intended platform format and view it on a phone. Check text legibility, safe placement, pronunciation and the consistency of the speaker across cuts. Record credits consumed by both accepted and rejected generations. No comparative quality or audience-response results were produced for this article; these are suggested checks for a real evaluation.

04 / PricingCredits describe actions, not a fixed number of finished videos

The Captions pricing page labels its public amounts as USD and explicitly says its features and prices reflect iOS plans only. Max is US$24.99 per month with 500 credits, while Frontier is US$69.99 with 1,400 credits. These figures should not be copied into a web or Android procurement estimate without checking that channel's offer.

The credit guide makes generation consumption more granular: AI Edit, media generation and chat-directed editing all draw from the allowance. Actual usage can vary with duration, model and action. A credit pool therefore cannot reliably be translated into an exact number of finished videos without specifying how those videos are made and revised.

The two official pages disagree on the free allowance: the main comparison labels free AI credits as absent, while the help guide describes a non-renewing introductory allocation. This blueprint does not promise a particular free generation entitlement. Inspect the actual account. Enterprise is custom-priced and the public card describes custom seats, bulk credit terms and training-data exclusion; do not assume those controls apply to every consumer plan.

05 / DistinctionsA choice between editing decisions and model access

Descript is an adjacent approach when the source is a long conversation and the main job is reshaping recorded speech. Mirage's Captions workflow emphasizes stylized short video and generated presenters. A useful comparison gives both tools the same correction, rather than comparing their most polished marketing examples.

HeyGen is a closer comparison for an organization standardizing presenter-led explanations or multilingual video. Examine the entire approval chain: source script, presenter, pronunciation, scene revision and export. The deciding feature may be control over a repeated brand detail rather than how striking a single generated face looks.

Mirage's company structure also matters for a developer evaluating its models. Access to a creator subscription is not evidence of API throughput, licensing or a supported production integration. An application buyer wants a usable editing surface; an integration buyer needs stable inputs, outputs and failure handling. Those purchases can share underlying technology while requiring separate commercial review.

06 / QuestionsThe consequential questions are about correction and identity

Can the chosen subscription support the real workflow?

Check the current product and platform before using an advertised credit allowance as a budget. Keep the iOS limitation visible in the purchase decision. If a team works mainly in a browser, its pilot should use that browser account from the beginning. A successful demonstration in a different application is not enough to establish the same entitlements.

What is being represented by the presenter?

The Captions terms describe user-provided input and place responsibilities on the person using the service. For an original production, decide whether the person on screen is a fictional narrator, an approved personal twin or an actual recorded speaker. Keep that distinction understandable in the context where the video will appear. A generated performance should not create a false impression of personal experience or endorsement.

Can the message survive a style change?

Test a conservative and an energetic version of the same clip. A fast transition can obscure a qualification, and a generated reaction can imply certainty beyond the script. Have the subject expert approve meaning separately from the designer approving appearance. That division keeps a visually successful edit from becoming the final factual check by accident.

07 / DecisionChoose a repeatable presenter and approval process

Mirage is worth evaluating when short video production is frequent enough that editing style and presenter continuity become operational concerns. Start with one message that matters, require a substantive revision and track the actual cost of the accepted result. Use the result to choose the recorded or generated route, then document who approves future scripts and appearance changes.