Recreate a Viral Video With Your Own Character

Before you start: what your character needs

Replication is not a general text-to-video tool. It works by putting a specific, established identity into an existing piece of footage, which means the identity has to exist first and has to be described well enough for the system to reproduce it consistently.

Concretely: the character must be finished and ready, and it must have its reference photos configured — a face shot and a front-facing full-body shot at minimum. Those are what the system leans on to keep the same person across every frame. A character without them will not appear in the picker at all, which is the most common reason people find the list empty.

If you also want the finished clip to carry your character's voice rather than the source's audio, the character needs a cloned voice attached. That is optional, it is a separate toggle, and it costs a small amount extra — but it has to be set up in advance, not at replication time.

Faithful mode: your character replays the exact movement

Faithful mode runs in two stages, and knowing this is what makes the mode's strengths and limits obvious.

First, the system extracts the source video's opening frame — a single still — and replaces the person in it with your character, keeping the original clothing, pose, scene and framing intact. Only the identity changes. The result is one image: your character, standing exactly where the original person stood, dressed the same, in the same room.

Second, that image becomes the starting point for a motion pass that drives the entire source video onto it. This is the part people usually get wrong when they describe the feature. The source video is not discarded after the first frame is taken — the whole clip is consumed as a motion reference, and the movement it contains is what your character performs. The first frame determines what the scene looks like; the full video determines what happens in it. The original audio is kept by default.

The practical consequence is that faithful mode reproduces choreography precisely. If the outlier you picked works because of a specific gesture, a timed movement, or a transition that lands on an exact moment, this is the mode that preserves it. It is also the mode that keeps the original spoken language, because the audio comes through unchanged unless you swap it for your character's voice.

The limit is the mirror image of the strength: because the motion is transferred rather than reinvented, the output stays anchored to the source's framing and pacing. You cannot ask it to be shorter, to change setting, or to say something different. It is a re-performance, not a reinterpretation.

Adapt mode: the video is analysed, then rebuilt

Adapt mode does something categorically different. There is no frame extraction and no motion transfer. Instead, the entire source video is passed to a vision model that watches it and writes a detailed description of what it sees — the shot, the movement, the setting, the delivery, the pacing.

That written description then drives a video-editing pass, alongside your character's reference photos, which produces a new clip built to that description. The output is not tracing the original's movement frame by frame; it is a fresh render guided by a prompt that was derived from the original.

This buys you one thing faithful mode cannot do: language. Because the dialogue passes through a written stage, you can nominate an output language and get the same concept delivered in it — around thirty languages are available, and leaving the setting on the original keeps the source language. For anyone whose audience does not share the source creator's language, this is the difference between a usable format and an unusable one.

The trade-off is fidelity of movement. A description, however detailed, is a lossy encoding of a video. Broad action and framing survive; a precisely-timed gesture may not come back the way it went in. If the post you picked hinges on an exact physical beat, faithful mode is the safer choice. If it hinges on a concept, a setting, or something being said, adapt handles it and gives you the language control as well.

What it costs

Video replication is priced by the source clip's duration, scaling linearly from a five-second anchor and rounded up so there are no free seconds. There is a floor: anything under about four seconds is charged as if it were four.

Faithful mode at standard quality is anchored at 15 credits per five seconds. So a five-second clip is 15 credits, a ten-second clip 30, a fifteen-second clip 45. Faithful at high quality — which buys tighter motion reproduction — is anchored at 25 credits per five seconds: 25, 50 and 75 for the same three lengths. Adapt mode is anchored at 21 credits per five seconds regardless of quality setting, so 21, 42 and 63.

Applying your character's voice to the result adds 2 credits on top, in either mode.

Image replication is much cheaper and flat-rated: 3 credits per image, no duration to scale against. If the outlier you found is a still rather than a video, or if you simply want to test whether a scene suits your character before committing to a video, this is the cheap way to find out.

One thing worth knowing about the economics: if a replication fails, the credits are returned automatically. You are not charged for a job that did not produce a result.

Running the job

The flow is deliberately short. From the replication browser, each tile carries a small button in its corner that opens the replication panel for that post — videos and images each get their own version of it. You can also open a video in the player first and start the replication from there if you want to watch it once before committing.

Inside the panel you pick your character, pick the mode, set quality or output language depending on which mode you chose, and decide on the voice toggle. The credit cost is displayed live and updates as you change those settings, alongside the source clip's duration — so you always see the price before you commit, not after.

Submitting closes the panel immediately. The job runs in the background and you are notified when it finishes or fails; the result lands in your gallery like any other generation. Video jobs take meaningfully longer than image jobs, so start one and go do something else rather than watching the page.

Then compare the result against the sentence you wrote in the previous lesson. If the thing that made the original travel is present in your version, you have what you came for. If it is not, the usual fix is a mode change rather than a retry — a lost gesture means faithful, a concept that came out muddled means adapt with a clearer source.

Getting to the page

The replication browser lives at ofmai.ai/replicate. It is not listed in the sidebar or the mobile navigation, so browsing for it will not find it — go to the address directly and bookmark it if you plan to use it regularly.

Any signed-in account can use it. If you are not signed in you will be sent to the login page first.

Replicate a video end to end

1. Check your character is ready — The character must be finished and have its reference photos configured — face and front-facing full body. Without them it will not appear in the picker. Add a cloned voice too if you want the voice option.
2. Open the replication browser — Go to ofmai.ai/replicate signed in, and load the profile holding your outlier post.
3. Open the replication panel — Use the small button in the corner of the post's tile, or open the video in the player and start from there.
4. Pick your character — Only ready characters with reference photos are listed. Select the one whose look suits the source.
5. Choose the mode — Faithful if the post depends on exact movement or timing. Adapt if it depends on a concept, or if you need the dialogue in another language.
6. Set quality or language — Faithful offers standard and high quality. Adapt offers an output language — leave it on the original to keep the source language.
7. Decide on the voice — The voice toggle replaces the audio with your character's cloned voice for 2 extra credits. It requires a cloned voice on the character.
8. Check the price, then submit — The credit cost and the source duration are shown above the button. Submitting closes the panel; the job runs in the background and you are notified when it lands in your gallery.

Faithful vs adapt at a glance

The two modes consume the source video differently. This is what that means in practice.

Frequently asked questions

Does faithful mode only use the first frame of the source video?

No — that is a common misreading of how it works. The opening frame is used for one specific job: it is the still your character gets placed into, which fixes the scene, the outfit and the framing. But the complete source video is then used as a motion reference, and it is that full clip which determines what your character actually does across the whole duration. If only the first frame were used, there would be nothing to animate from and the mode could not reproduce choreography at all. This is also why the price scales with the source video's length rather than being flat: the whole clip is being processed, not a single frame from it. The original audio is carried through by default too, unless you switch on the character voice option.

Can I find the replication page in the sidebar?

No. The replication browser is at ofmai.ai/replicate and it is deliberately not listed in the desktop sidebar or the mobile navigation, so searching the menus for it will be fruitless. Navigate to the address directly and bookmark it if you expect to use it often — that is the intended way in. You do need to be signed in; visiting it while logged out redirects you to the login page and you can come back afterwards. Beyond being signed in there is no additional gate: any account can use it, and the credit costs described in this lesson are the same for everyone. Once you are on the page everything else lives there, including the profile loader, the post grid with its engagement counts, and both replication panels.

What happens if a replication fails?

The credits are refunded automatically and the job is marked as failed in your gallery. You do not need to contact anyone or claim anything back — the refund is part of the pipeline, and it fires whether the failure happened at the frame stage, the motion stage or the rendering stage. Because both modes run as multi-stage background jobs, there are several places a job can stop, and the refund covers all of them. You will be notified when the job resolves either way, so you find out without watching the page. The most common avoidable failure is a source video that cannot be fetched, which usually means the post is no longer publicly reachable; reloading the profile and picking the post again generally resolves it. Failures caused by the source itself are worth spotting early, since a different outlier is faster than a retry.

Should I replicate images or videos when starting out?

Start with images. At 3 credits flat, image replication is by a wide margin the cheapest way to learn whether a given scene, framing or style actually suits your character — and that judgement is the hard part, not the button-pressing. A video replication of a fifteen-second clip costs between 45 and 75 credits depending on mode and quality, so testing your instincts on video is an expensive way to discover that a setting does not flatter your character. Run a handful of image replications across different looks first, see which ones come back convincing, and let that inform which video outliers you commit to. It is also faster: image jobs resolve quickly, so you can iterate several times in the span of one video job. Once you have a feel for what works, move up to video.

Related reading

• Keeping an AI character consistent
• What an AI influencer costs to run
• AI influencer for Instagram