How to Make a Fancy Cafe AI Video: Complete Guide

To prepare a Fancy Cafe AI video, choose a clear photo for the left visitor, the right companion, or both; check the reference scene; then review the intended casting before generation. The complete scene target is 21 seconds. Today this site lets you watch the reference and select photos locally. A new model is still being evaluated, and video generation and payments are not open yet.

This guide separates the actions you can try now from the output checks you should make when generation becomes available. It does not imply that choosing photos creates a video, or that the reference player is showing a result made from your pictures.

Step 1: Watch the complete scene

Open the Fancy Cafe AI reference player and watch from the beginning through the final reaction. The first guest is already perched on a wall-mounted metal table. A companion enters from the right, sits on a low ledge, and places a drink on the table. Seeing that, the first guest realizes the mistake and moves down.

The setup and the ending both matter. A crop that hides the ledge makes the reveal difficult to understand. A clip that stops before the guest changes seats loses the payoff. Our supplied source runs approximately 21 seconds, and its current audio can be heard by turning on sound in the player.

Step 2: Decide who plays each role

The Left person card represents the first visitor on the wrong seat. The Right person card represents the arriving companion. These are reference positions, not fixed gender requirements. Prepare one separate photograph for each role you want to change.

1. Choose only the left card if you want to play the mistaken visitor.
2. Choose only the right card if you want to play the companion.
3. Choose both cards to prepare a two-person cast.
4. Check the labels before moving on; the order in which you select files does not change the card labels.

The intended service will replace only the selected roles. The current local selector displays your choice, but actual role preservation and tracking still need verification with the new model. Read the scene explained if you are unsure which person is which.

Step 3: Prepare clear source photos

Use a sharp photograph of one consenting adult. The face should be easy to distinguish, with even light and no strong shadow across the eyes. Include the shoulders and upper body so the appearance is not defined by a tiny face crop alone. A head-to-waist or head-to-knees view is a useful starting point.

The local selector accepts JPG, PNG, or WebP, up to 10 MB per photo. These are the site’s selection limits, not a claim about every possible model. Avoid sunglasses, hands covering the face, heavy beauty filters, and motion blur. For a group photograph, prepare a separate crop containing only the person who agreed to participate.

Do not use celebrities, public figures, minors, or unauthorized photos. Permission to possess a picture is not the same as agreement to appear in an AI video. Check our acceptable use rules before preparing a cast.

Step 4: Try the local photo cards

Select Choose photo on either role card and pick a supported file. The card shows your selected photo with its role label. Use Change photo to choose another image, or Remove to clear that role. You can test both cards independently; choosing a right-side photo does not move it to the left.

This interaction happens in your browser. It does not upload a file, generate an AI image, start a video task, or spend a credit. The adjacent video remains the supplied reference. If your chosen photograph looks different from the sample actor, that is expected: the card is displaying your photograph, not a rendered result.

Step 5: Check availability and pricing

The generation button is currently disabled and checkout is closed. The site is evaluating a new model for the complete 21-second scene, character identity, role assignment, audio, and processing cost. There is no confirmed production processing time or generated-video quality guarantee.

Current displayed prices remain $9.99 for one video, $19.99 for three, and $29.99 for five. These packs cannot currently be purchased. Our free preview and pricing explanation describes exactly what is available without payment today.

What should you check in a future generated result?

When generation opens, review more than the first frame. A convincing face at the beginning does not prove that the same identity holds during the companion’s entrance or the final movement. Use this checklist before treating any result as ready to share:

• Identity: each selected person remains recognizable throughout, with no merging of the two faces.
• Roles: the intended visitor takes the wrong seat; the intended companion arrives on the correct side.
• Actions: hands, cups, limbs, and the move to the ledge remain physically plausible.
• Scene: the table and lower seat remain visible enough to explain the joke.
• Duration: inspect the actual downloaded file, rather than assuming a label establishes 21 seconds.
• Audio: listen to the whole clip and check whether it contains source audio, generated audio, or silence.
• Delivery: verify the file opens and that any download expiry is clearly explained.

These are acceptance criteria for the planned service. They are not claims that the current reference demonstrates a personalized generation passing those checks.

Why does the model choice matter?

Video generators and motion-transfer models solve different tasks. A general reference-to-video model may reconstruct a scene from several inputs. A motion-transfer model uses a driving video to guide how characters move in a prepared image. Neither a duration setting nor a model name alone guarantees the correct two-person reveal.

For context, fal’s motion-control API documents separate image and video inputs and an original-audio option. Seedance 2.5’s API documents a different multimodal reference workflow. These are technology examples; this website has not committed to either production model.

Common preparation problems

The wrong person is in a card: remove the photo from that card and choose it under the intended role. Check the visible Left person and Right person labels rather than relying on selection order.

The file is rejected: use JPG, PNG, or WebP and keep it at or below 10 MB. Renaming a file extension does not convert the image format. Export a supported copy from your photo editor if needed.

The sample does not change after selecting a photo: this is the current behavior. The sample is shared reference footage. Local selection does not run a model or modify its actors.

There is no generation or payment button available: generation and checkout are intentionally closed during model evaluation. Repeatedly signing in, refreshing the page, or changing photos does not unlock them.

Start with the reference scene and role cards. Once the model and workflow are verified, the availability notice and guide should be updated to reflect the actual generation and delivery behavior.