# Scene assembly — the cut as one track, condensed The assembly procedure with the film-specific defaults. Self-contained, and since the `picsart-video` plugin was retired on 2026-09-10 (its `video-from-clips` and `refine-transitions` skills went with it) this is the only version — a one-off cut of the user's own footage is this same sequence without the pipeline around it. ## Tool names Every compositor tool on this server is prefixed `picsart_media_`. `picsart_media_quickstart` returns recipes naming `mp_probe_media`, `mp_export` etc. — those names do not exist here; translate `mp_` → `picsart_media_` and ignore the discrepancy. ## Procedure per scene 1. **Real URLs only.** Takes come from selects as `https://` URLs already. A user file goes through `picsart_drive` (`action: "upload"`, from a `data:` URI) first. 2. **Probe every take** — `picsart_media_probe_media` per URL for exact `width`, `height`, `duration`. The assets carry them; guessing produces a scene that validates and renders wrong. Skip only if a prior result already gave exact numbers. 3. **Author the cut** — one scene, one layer, whose content is a **track** holding the shots as `clips` in playback order. Pass it directly — as `picsart_media_validate_scene`'s `scene`, as `picsart_scene_editor`'s `scene` — and read the exact field set from `picsart_media_get_scene_schema`: ```json { "version": "1.0", "composition": { "width": 854, "height": 480, "fps": 30, "background": "#000000" }, "assets": { "sc01_12A": { "type": "video", "uri": "https://…/sc01_12A.mp4", "width": 854, "height": 480, "duration": 8.0 } }, "layers": [ { "id": "cut", "start": 0, "content": { "kind": "track", "clips": [ { "id": "sc01_12A", "duration": 7.2, "content": { "kind": "media", "assetId": "sc01_12A", "trim": { "in": 0.4, "out": 7.6 }, "fit": "cover", "align": "center" } } ] } } ] } ``` Rules: never set composition duration (it derives from the track); an asset's `type: "video"` or the take is treated as a still; every clip carries an `id` (the shot's name), the handle a camera move on that one shot targets; a clip's `duration` is its window, `trim` is `{ in, out }` on the source; the same take twice with different `trim` windows is how a cutaway returns; hard ceiling 1920×1920. Transitions are **per seam**, on the track's `transitions: [{ between, id, duration }]`, where `between` is the index of the clip the seam follows. 4. **Validate** — `picsart_media_validate_scene`. Free; catches schema errors that otherwise fail after minutes of rendering. 5. **Review** — `picsart_scene_editor` (the scene editor) on the validated scene, passed as `scene` — the parsed document from step 3, with **all** the shots in playback order. One editor, never one per clip and never one per piece of a cut the user made in it. It plays in the browser and renders nothing. When the reply says `edited: true`, its `document` is the scene from here on — the user's trims, splits and deletions already applied; apply any comment-driven change to it (`picsart_media_patch_scene`, free), validate, and reopen if there is something new to watch. 6. **Export** — `picsart_media_export`, `mediaType: "mp4"`. **No credits** (2026-09-07), but a real GPU render: a failed or abandoned one still costs the wait and leaves nothing. Never retry speculatively; never present a successful render as an error. Free, instant look-check before rendering real frames: `picsart_media_query_layout` (where layers land at time t). `picsart_media_contact_sheet` renders real frames — **No credits**, costs render time: target the suspect joins, not all of them. ## Film defaults - **Within a scene: hard cut** — no `transitions` entry for that seam. Continuity is the cut's job; crossfading coverage smears what Stage 6 paid for. - **Between scenes**: hard cut or `crossfade` 0.3–0.6s. Anything decorative must be motivated by the film's grammar, not novelty. - Draft canvas 854×480 to match 480p drafts; the final canvas (≤1920×1920) exists only in `film-finishing`. - **The window is the model's cut until the user moves it** (the automatic rough cut is off — below). A run generated with handles is longer than its shots on purpose: `genSec` (what was ordered) minus the sum of `sec` (what the film runs) is material at the run's two ends for the user to trim in the scene editor, and it plays until they do. **Do not derive the handle as 2s.** Snapping to the model's duration ladder gives more than that most of the time and less at the ceiling; the numbers on the run are the truth. ### Where the trim window comes from **No window is derived (2026-09-04).** The automatic rough cut is off: a run is split at its delivered cuts, and each shot's window is the model's cut — `in`/`out` are the `cutTimesActual` either side of it — until the user moves it in the scene editor. Record `trimSource: "user"` beside a window the user set; nothing else is written, and nothing automatic ever moves a window. The one thing that still bites silently: **a run-edge crossfade of 0.3–0.6s consumes 0.15–0.3s from each side**, so if one is used (last resort — see `../../film-scenes/references/SEAMS.md`) both points move inward by half the window or the transition eats designed action. A seam with no `transitions` entry pays nothing, which is why the default is a hard cut and a window, if ever, is spent only at a run edge. ## Virtual camera moves — animating the layer's window A move over an existing clip is the layer's own transform, animated. Two ways, both free to stage: **The one-call way — the `ken_burns` motion preset** via `picsart_media_apply_motion_preset`, with the clip's `id` as `layerId`. Parameters, as the capability doc declares them: | Param | Range | Meaning | |---|---|---| | `zoom` | 1–2 (use ≤ 1.2) | final scale zooming in; initial scale zooming out | | `direction` | `in` `out` `n` `s` `e` `w` `ne` `nw` `se` `sw` | `in`/`out` are pure zoom; the compass values add a pan | | `pan` | 0–500 **composition pixels** | pan distance; ignored for `in`/`out` | So a slow push-in is `{ zoom: 1.15, direction: 'in' }`; a drift is `{ zoom: 1.12, direction: 'w', pan: 120 }`. **The precise way — keyframes** via `picsart_media_patch_scene`, setting the clip's `animations` (the clip found by its `id`). It is an object keyed by property, each holding `keyframes: [{ time, value, easing? }]`: - `scale` — a per-axis `[sx, sy]`, identity `[1, 1]`. A push-in is `[{ time: 0, value: [1, 1], easing: 'linear' }, { time: , value: [1.15, 1.15] }]`. - `position` — **absolute composition pixels** `[x, y]`, for pans and tilts. - `anchor` — layer-local pixels from the layer's top-left; this is what sets the **zoom centre**, so a push-in onto a face anchors on that face. - easing sits on the keyframe that starts a segment, so the last one carries none: `linear` (one steady speed), `ease_in`, `ease_out`, `ease_in_out`, `step` (a hard jump — a punch-in is two scale keyframes ~0.1s apart, which reads as a cut). Ken Burns on an animatic still is the same thing on an image layer. Keep `scale` ≤ ~1.2 — past that the resolution loss shows. The overscan IS the travel budget: nothing exists outside the source frame, so an offset can never exceed half the scale-up, or the pan reveals emptiness. Check the framing with `picsart_media_validate_scene` and `picsart_media_query_layout` (both free), or one contact-sheet frame at the move's midpoint (**no credits**, costs render time). Staging is free; the move renders inside the scene's existing export. **Nothing previews this before it renders** (2026-08-31). A widget recipe table (`crop-path.ts`) used to animate these same primitives on the shot's own frame, so what the user saw was what the export performed; it is deleted, along with the policy of answering a camera-change note with a crop instead of a retake. Staging a move here is still legitimate edit craft — a push on a held insert, a drift across a still — but it is now unsighted, so check it with `picsart_media_query_layout` (free) before committing to it. ## Transition quick table | Want | Use | Window | |---|---|---| | invisible | `crossfade` | 0.3–0.6s | | soft, dreamy | `blur_dissolve` | 0.5–0.8s | | directional energy | `slide` / `push` | 0.3–0.6s | | impact | `zoom_through` | 0.3–0.5s | | camera-whip energy | `movement_camera` (set `flashIntensity: 1`) | 0.6–1.0s | | hard cut | `""` | — | `shape_reveal` mattes are 3-second animations — the window must be ~3s or the effect reads broken. `movement_camera` starts and ends at rest, so it splices anywhere. ## Join QC checklist (run at every seam) **First: did the join do what it was written to do?** At a video edge the outgoing shot ends mid-movement and the incoming shot picks it up from a clearly different angle (`../../film-scenes/references/SEAMS.md`); a join where neither side is moving, or both sides are the same framing, failed at generation, not at assembly, and no trim repairs it. Then the continuity list: eye-lines match · action axis kept (or crossed via a cutaway) · costume/props/hair continuous · light and palette do not jump · movement tempo does not lurch · adjacent shots of one subject are not the same size. A failing join is fixed by re-order, a cutaway, a trim — or a reshoot order back to `film-scenes` when the coverage simply does not exist. **Weight them, do not just tick them.** The received order of importance (Murch's, and it is the one professional editors work to) is emotion first, then story, then rhythm, then where the viewer's eye lands, then screen direction — and **literal three-dimensional continuity last.** Every item on the list above sits in that bottom band. So a seam that jumps a prop but lands the beat is a better cut than a spatially perfect one that kills the moment, and when the two conflict the beat wins. This matters here more than on set, because our continuity errors are the ones we cannot fix — knowing they are also the cheapest to forgive is what stops a film being re-shot over a coffee cup. **The join is authored upstream.** The mid-movement ending and the new angle are written into the two shot blocks either side of the edge (`../../film-scenes/references/SEAMS.md`, *The video edge*), so a join judged here was designed before it was generated; a join that fails it is a re-roll of the take that missed its ending — not a trim. **How to actually SEE a join** (the checklist is uncheckable from memory): `picsart_media_contact_sheet` the outgoing clip's last frame and the incoming clip's first frame (two frames per join — no credits but costs render time, so spot-check the suspect joins, not all of them), then view both as inline images and verdict the checklist against real pixels. Fold these QC frames into the film's time budget. Two fix rounds maximum, then ship with one honest line (see the skill's bounded-verification rule).