MP4/MOV or WebM · up to 100 MB, 1080p, 60 seconds / 1,800 frames. Exports keep at most 30 seconds.

Build a faceless, explainer or product draft from 1–6 stills. Each scene gets a caption and slow zoom. Images come from your local image studio or uploads.

Up to 30 seconds total. Narration longer than the timeline is trimmed and flagged.

During a gap, the crop holds still for at most half a second. Recovery requires a unique close match to the original subject.

Draw a rectangle around textured detail on your subject. The crop follows visible motion and stops if tracking becomes uncertain. Existing pixels only; no generated edges. Up to 30 seconds.

Tracked mode uses no new diffusion. It needs one expanded still with this exact source frame and padding. It refuses large movement; tiny uncovered outer strips hold the initial still. Inspect parallax and frozen border content.

Ten percent padding creates a small border in the same shape. Large shape changes still need much more invented content.

Inspect the still in Image Studio. Then choose it under Saved appearance keyframe.

The guide must match this source frame, shape and padding. Its first frame is retained exactly; inspect the following generated frames for joins and flicker.

Select 0.6–2.1 seconds of source context. VACE generates two more seconds; the original frames are restored in a lossless master. Inspect the join. Added frames have no generated audio.

Prepared pose mode retains your colored skeleton video instead of extracting edges. It does not detect people or retarget body proportions. Upload a prepared pose clip and select a still appearance reference.

Choose a still that already shows the requested appearance and material. Conflicting still and prompt can steer the wrong look. Keyframes retain their review status.

Selected local VACE editor · 20 steps · 16 fps · up to 49 frames. Preview resolution follows the source shape. Model edits may drift or flicker.

Preserve mode transfers broad, neutral brightness from the AI proposal onto the original footage. Facial detail and colors come from the source. Amount 0 keeps the source; amount × limit bounds the shift. It does not add colored lights or guarantee new cast shadows.

Both export at twice the prepared source size. Resize keeps the source appearance; it cannot recover missing detail. AI enhancement may sharpen or invent details. Add sound afterwards with Edit this result.

Choose automatic clothing or paint only the material. Keeps its brightness, folds and texture; color may soften in bright highlights. Amount 0 retains the source. Up to 10 seconds, source shape retained.

The optional flow check flags fast or inconsistent movement. Holding uses an existing frame instead of an uncertain prediction; expect judder. It cannot detect every artifact.

2x frames at 2x playback rate: duration and audio stay the same. Keeps the original frames, protects detected hard cuts and holds the final frame for timing. Up to 3.4 seconds, 30 fps input and 0.5 megapixels. Inspect hands, thin edges and fast motion.

No target selected. The guide keeps the driver’s proportions.

Choose one person with clearly visible shoulders, elbows and wrists. Full-body fitting also needs hips, knees and ankles. Uses visible 2D limb lengths, so choose a front-facing pose with limbs apart. Our first generated comparison followed this fitted guide less accurately; compare with the original driver proportions. It does not recover depth, hidden joints or identity.

One fixed framing across the selected clip. Shoulders, elbows and wrists must stay visible. Missing or uncertain joints are omitted, not invented. The silent body guide does not contain facial or finger motion.

Neural object tracking

Click the picture to keep part of an object. Hold Shift and click to exclude hands or background. Inspect the colored preview before using it for a region edit.

Use the same original source, start/end range and crop as the selection preview. Mask mismatches are refused.

Draw the area to edit on the first frame of the selected range. Tracking follows visible texture and stops if confidence falls. Fast motion, occlusion and scene cuts can fail.

Replacement clears the model input inside the mask. The original footage remains protected outside your final soft mask.

Exposure and color adjustments are conventional finishing. AI Relight is a separate experimental workflow.