Four Continuities Make the Fall-Through-Feed Effect Work

The “fall through the feed” effect works by making a person appear to pass continuously through successive social posts. Its visual logic is simple: the character enters each frame from above, exits below, and the social interface moves with the same scroll. What reads as one uninterrupted fall is assembled from separate generated scenes.

Four constraints make that illusion work:

1. Identity: the same recognizable person, clothing, and general appearance must survive across every environment.
2. Pose: the body needs to retain a credible falling orientation from one style frame to the next.
3. Motion: each scene must begin empty, receive the character from the top, and lose them beneath the bottom edge.
4. Velocity: the generated clips and the social-post interface must travel through the edit at the same rate.

The demonstrated result uses a park as the real-world anchor. The person falls into a hole, descends through illustrated and surreal posts, then lands back in the park. The styles vary; the person’s apparent trajectory cannot.

Start with a vertical full-body reference. If only a headshot or partial photograph is available, the workflow expands it into a full-body image by tagging the original asset in the prompt and asking the model to preserve facial likeness. The demonstrated setup uses Nano Banana 2, a 9:16 aspect ratio, and at least 2K resolution. A creator who already has a suitable full-body photograph can use it directly.

That reference image is not just a starting asset. It is the continuity source for every later still. The source creates two additional versions of the same setting: a landing frame with the subject on the floor after the fall, and an empty plate with the subject removed while colors and composition remain unchanged. Those paired images give the eventual landing animation a defined location and end state.

The stylized scenes should derive from the full-body reference, not from the landing image. Their prompts have three functional layers:

• Identity instructions tag the reference and retain likeness, hairstyle, clothing, and footwear.
• Pose instructions define the fall in directional terms: for example, head tilted left, legs toward the upper right, bent arms and knees, and a diagonal body angle.
• Style instructions define the scene’s medium, environment, and visual treatment, often followed by a negative prompt.

The key is not the specific aesthetic. The example uses a retro 1990s anime cel treatment—bold linework, flat shading, warm color, grain, and painted sky and hills—but the source treats the style as interchangeable. What should not change is the identity-and-pose portion of the prompt. Reusing that language across styles prevents the sequence from becoming a collection of unrelated characters in unrelated body positions.

GPT Image 2 is used for the illustrated frames because the presenter prefers its style output. The source recommends generating those stills at the highest available quality and resolution, then repeating the same identity and pose structure for however many environments the edit requires.

A generated style image alone cannot create the scroll illusion. The video must contain a specific movement pattern: an empty scene at the start, a character dropping in from directly overhead, and a complete exit below the frame. The source’s detailed animation prompt also calls for a locked camera and a fixed, linear descent, while allowing arms, hair, and clothing to move during the fall.

The important technical distinction in Seedance 2.0 is between a start frame and an image reference. Connecting a style frame as the start frame makes the model begin with the person already present in the composition. That defeats the empty-to-entry structure needed for scrolling. The style frame should instead connect to the image-reference input, supplying character, scene, and aesthetic context without determining the first visible frame.

The prompt then carries the motion requirements the reference image cannot: no camera movement; steady vertical travel; clear limb movement; fluttering fabric and hair; and an empty image after the character leaves. The source repeats the desired style and mood in text even though the image reference already conveys them, as an additional cue for consistency.

For the demonstrated scenes, the source chooses four seconds, the shortest available duration, so the fall remains quick. It recommends 1080p rather than 4K for social distribution: platforms will compress the result, and 1080p uses fewer Seedance credits. If a 4K deliverable is needed, the source suggests upscaling afterward.

The landing shot uses the empty park plate and the frame of the subject on the ground as the visual endpoints for a separate generation. The opening uses a similar approach: the person steps forward, a round hole opens in the ground, and they fall in while the camera remains fixed. The sinkhole is presented as an effective opener, not a required one; another entry effect could occupy the same structural role.

Premiere Pro turns separate falling clips into posts in a feed. The source captures Instagram’s lower interface—the account area, engagement counts, controls, and caption region—as a separate graphic layer. One screenshot can be reused, though multiple screenshots with different descriptions and numbers make the sequence feel more like distinct posts.

Every generated clip receives the same Position animation. At its first keyframe, the clip sits just below the visible program frame. At a later keyframe, it has moved up beyond the top edge. Played back, that upward motion makes the subject appear to be falling down through the screen.

Copying those keyframes to every style clip is essential. If different scenes move at different speeds, the character’s trajectory breaks at each cut. The Instagram UI must receive matching position keyframes for the same reason: a stationary interface would read as an overlay placed over moving footage, not as part of the feed itself.

The source arranges adjacent clips so one begins before the previous one finishes. Between them, it keeps an equal black gap, sized just large enough for the Instagram interface to slide through. Consistent gaps make the handoff feel intentional rather than like a cut between clips.

The beginning and end require the same alignment. The opening animation joins the scrolling sequence once the subject has left its frame. The outro uses matching keyframes as well, but animates into the final section rather than simply repeating the earlier exit.

A seamless loop is not made by placing the existing intro and outro together. The source generates a dedicated bridge in Seedance 2.0 using three static assets as image references: the relevant empty location, the landed subject, and the sinkhole context.

Its prompt asks for an empty frame, the person falling from the sky, getting up, taking a step forward, and then falling through a hole that opens in the ground. The attached images are connected as references rather than start frames, so they serve as contextual anchors for a newly generated movement.

The resulting clip is split in the middle. Its second half becomes the intro, and its first half becomes the outro. Replacing the original endpoint clips with those reciprocal halves produces the loop shown in the source: the person lands in the park, moves toward the hole, falls through the feed, and returns to the same place.