Merge two clips
without the glitch.

RIFE transition creates new intermediate images to connect movement across the join. Automatic correction aligns and warps existing images, matches their color, and blends them briefly. Simple cut just plays Clip A followed by Clip B without visual smoothing or endpoint trimming. Start with the five-frame RIFE transition, then compare the other methods if the face, movement, or background looks unnatural.

RIFE is a trained frame-interpolation model. It takes two endpoint images, estimates how their contents move, and synthesizes images between them. Clip Merge uses it only for a short interval around the join. Source footage plays before and after that interval. This method can help small pose or position changes, but it cannot guarantee a correct face or seamless movement when the clips disagree substantially.

Automatic correction uses our own TypeScript implementation of classical computer vision: similarity-transform registration for camera alignment, RGB gain and offset matching for color, and pyramidal Lucas–Kanade optical flow for local movement. It warps Clip B toward Clip A, briefly blends their images, then gradually releases the correction. It runs in a browser worker without a trained AI model or OpenCV. The steps and their limitations are explained below.

Use Simple cut when the clips already match, when a visible edit is acceptable, or when smoothing introduces distortion. It preserves the untrimmed clip sequence and does not synthesize or warp images. The output is still fitted to Clip A’s dimensions and frame rate and encoded as an MP4, so this is not a lossless file concatenation. Audio receives brief fades at the join.

A crossfade gradually reduces one image’s opacity while increasing the other’s. If a face is in different positions, both positions can appear at once. RIFE estimates intermediate movement and synthesizes images instead. Automatic correction first tries to align and warp the images before blending them. There is no separate plain-crossfade setting in Clip Merge.

It compares small, 128-pixel-wide grayscale samples around the join, searching up to 0.5 seconds of trimming on each side when footage allows. It ranks candidate pairs by image difference plus a penalty for trimming. For the strongest candidates, it searches for a similarity transform: horizontal and vertical movement, uniform scale, and rotation. This aligns the overall image; it does not reconstruct a 3D camera or independently identify the face. A similar background can therefore outweigh a mismatch in the subject.

It measures the mean and standard deviation of the red, green and blue channels in the selected images. From these, it calculates a bounded gain and offset for each channel of Clip B: a contrast adjustment followed by a brightness adjustment. This can reduce small exposure or color-cast differences. It is a global statistical adjustment, not selective relighting or an AI color-grading model.

The initial compatibility check uses bidirectional block matching. The more detailed correction uses pyramidal, inverse-compositional Lucas–Kanade optical flow on samples up to 384 pixels wide. It tracks small image patches through progressively finer image scales and estimates movement with fractional-pixel precision. Forward/backward checks reject matches that fail to return near their starting point, and texture and image-error checks reject other uncertain matches. This estimates where image details moved; it does not recognize eyes, glasses, faces or body parts.

It combines motion estimates from three paired samples across the overlap, using a median-based consistency check and confidence weighting to reduce isolated tracking errors. Edge-aware diffusion then spreads reliable movement into nearby low-texture areas while discouraging it from crossing color boundaries. Here, diffusion means numerical smoothing of motion vectors, not a generative diffusion model. Three additional samples in Clip B help carry the correction along with moving content. There is no semantic subject mask, so similar-looking regions can still receive the wrong movement.

The camera transform and local displacement field are combined into one resampling of Clip B. The aligned images are blended over roughly four output frames, about 0.17 seconds at 24 fps, with a shorter overlap for very short clips. Pixels that still differ strongly switch between sources over a narrower part of that blend. After the overlap, a smooth easing curve releases the geometric and color correction over up to 0.8 seconds. Wrong correspondences can stretch glasses or cheeks, and pixels switching at different times can produce a local pop. A failed initial compatibility check uses a simple cut, but passing that check does not guarantee every facial detail is correct.

Every RIFE call receives the same two original endpoint images, A and B, plus a number telling it how far between them to place the result. For five frames, the positions are 1/6, 2/6, 3/6, 4/6 and 5/6: about 17%, 33%, 50%, 67% and 83% through the transition. The sequence is A, five new images, then B. None of the generated images becomes an input to another call.

RIFE can use the actual last and first frames. Our current method chooses endpoints slightly inside the retained clips so it can replace a short interval around the join, keeping the corrected timeline and audio timing. Inserting five new frames while keeping every original frame would add about 0.21 seconds at 24 fps and require an audio timing adjustment. The replacement approach changes a little original movement; it is a timing tradeoff, not proof that inside endpoints always look better.

It depends on the frame rate, transition length, existing overlap and any selected trims. With five generated frames at 24 fps and a four-frame overlap, the requests are roughly 0.21 seconds before the retained end of A and 0.21 seconds after the retained start of B. The exact times are rounded to the output frame grid, and the decoder uses the source frame covering each requested time. Clip Merge searches up to 0.5 seconds per side for candidate trims, subject to available footage; RIFE endpoints are then positioned around that selected join. There is no separate face-aware endpoint search.

Five is the default. At 24 fps, three generated images span about 0.17 seconds between the original endpoints, five span 0.25 seconds, and seven span 0.33 seconds. Five new images plus two original endpoints have six intervals between them, which explains the 0.25-second duration. Three makes a quicker transition; seven spreads the change over more time. More frames can make a jump gentler, but can also prolong a morph or make motion feel slow. Preview the face and the movement entering and leaving the transition.

No. In Clip Merge, RIFE replaces a few frames on the existing corrected timeline; it does not increase the output frame rate or add time to that timeline. Smoothing may already shorten the combined clips through endpoint trimming and overlap. Simple cut preserves the untrimmed sequence. If you want to increase the frame rate throughout a video, use the separate AI Video Frame Interpolator.

Not reliably. All smoothing methods work best when the clips already depict similar content. RIFE can still morph details or invent incorrect intermediate shapes; classical warping can stretch them. The compatibility rating is an estimate, not a visual-quality guarantee. If the initial correspondence check fails, both smoothing options use a simple cut. Inspect faces, glasses, hands and background edges in the preview even when the rating looks good.

No. Video decoding, RIFE inference, correction and MP4 creation run locally in your browser. The browser downloads the model and processing software; it does not upload your source videos for processing. Keep the tab open while the job runs. RIFE can use a compatible GPU and falls back to slower CPU processing when necessary.

The output uses Clip A’s dimensions, with Clip B fitted and padded to preserve its aspect ratio. The output frame rate is derived from Clip A. Downloaded files use H.264 video and, when present, stereo 48 kHz AAC audio. Audio follows the same trim and overlap timeline as the video, with a short crossfade for smoothing or brief fades for a simple cut. RIFE creates video frames only; it does not generate missing speech or repair audio continuity.

Yes. The preview uses the same endpoint times and transition setting as the download, but is limited to 640 pixels wide. The downloaded MP4 is created at the output resolution, so generated details can differ. Check the downloaded result when small facial details matter. If RIFE fails to load or run, try Automatic correction or Simple cut.