--- title: Editing video — Runway Aleph 2.0 | Runware Docs url: https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video description: How to use Aleph 2 to edit existing footage with localized changes that preserve the rest of the clip. Covers the edit prompt formula, frame-image anchors for precise visual control, and multi-shot propagation across cuts. --- ### [Introduction](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#introduction) Most video models start from a blank canvas: text in, generated footage out. That works when the shot doesn't exist yet. Commercial production starts the other way around, from **footage that's already been shot and color-graded**, that now needs a single specific change. The seasonal variant of one campaign film. The wardrobe swap a client signs off on three days before launch. The old answers were a roundtrip through compositing software or a reshoot. Both are slow, and both lock the rest of the clip into whatever the first edit sets in motion. Aleph 2 changes the default. You hand it footage you already have, name one specific change, and it returns the same clip with that change applied and **everything else preserved**. The framing, the cuts, the lighting math, the parts of the frame you didn't ask it to touch. [Watch video](https://runware.ai/docs/assets/source-jet.DwJ8gUHs.mp4) *Source clip* > **Prompt**: A sleek black single-seat fighter jet with smooth angular bodywork banks dramatically against a clear deep blue sky above a vast sea of golden cumulus clouds at sunset. Slow side tracking shot following the jet from a slight distance, the wing dips toward the camera revealing the underside, soft volumetric light catching the canopy and fuselage, faint vapor trails curling from the wingtips. Cinematic wide shot, photoreal, high-end aerial cinematography, warm golden hour color grade. No people visible. Ambient wind sound only, no dialogue. [Watch video](https://runware.ai/docs/assets/output-hero-color.DCXJBUjK.mp4) *Repainted as a white air-ambulance* > **Prompt**: Repaint the jet's bodywork as a glossy pure white air-ambulance with a bold red cross marking on the fuselage and tail. Keep the jet's shape, motion, banking angle, the sky, the clouds, the lighting, and the vapor trails exactly as in the source. [Watch video](https://runware.ai/docs/assets/output-hero-environment.wTO8v0K_.mp4) *Relocated to the edge of space above an alien planet* > **Prompt**: Replace the cloud bank with the view from the very edge of a planet's atmosphere: the curving horizon of a deep indigo planet stretches below with ribbons of green and violet aurora rippling along its surface, the black void of space fills the sky above scattered with sharp pinpoint stars, a vast cratered moon hangs close in the upper frame with its grey rocky surface clearly lit. Keep the jet, its color, its banking motion, its framing, and the camera move exactly as in the source. [Watch video](https://runware.ai/docs/assets/output-hero-restyle.DIvZ5D1O.mp4) *Restyled as a 1940s newsreel* > **Prompt**: Restyle the clip as a 1940s WWII black-and-white newsreel: monochrome film grain, slight gate weave, soft contrast, the look of period combat-aviation footage. Keep the jet, its motion, its banking angle, the clouds, and the framing exactly as in the source. Where this fits in production: seasonal variants of one campaign film, wardrobe swaps after a client's late note, SKU updates that would otherwise rebuild a product video library, restyled drops without a second studio day. The same source clip becomes every version the brief now calls for. This guide covers how to drive Aleph 2 for film and advertising work: the edit prompt formula the model honors, when to reach for frame-image anchors to lock a specific visual target, how a single edit propagates across the cuts of a multi-shot reel, and the source-video constraints to plan around. ### [Request shape](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#request-shape) An Aleph 2 edit takes a source video and a positive prompt. Frame-image anchors are optional and unlock the precision workflow described later. **Request**: ```json [ { "taskType": "videoInference", "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890", "model": "runway:aleph@2.0", "inputs": { "video": "https://example.com/source.mp4" }, "positivePrompt": "Change the sedan's bodywork color to a deep glossy crimson red. Keep the chrome trim, glass, headlights, road, and surroundings exactly as in the source.", "outputType": "URL", "outputFormat": "mp4", "numberResults": 1 } ] ``` **Response**: ```json [ { "taskType": "videoInference", "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890", "videoUUID": "9c1b2d3a-4e5f-6789-abcd-ef0123456789", "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/9c1b2d3a-4e5f-6789-abcd-ef0123456789.mp4" } ] ``` Two required fields, the rest optional: - `inputs.video` accepts a public URL or a UUID from a previous generation. The clip must be **between 2 and 30 seconds** at up to 1080p. - `positivePrompt` is required and capped at 1000 characters. Describe only what changes. See [Driving with a prompt](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#driving-with-a-prompt) below. - `inputs.frameImages` is optional. Up to **5 anchored reference images**, each pinned to a position in the clip via `frame` (`first`, `last`, `0`, `-1`) or `timestamp` (seconds, 0.01 precision). Each image accepts a public URL, base64 string, data URI, or a UUID from a previous generation or the [Image Upload API](https://runware.ai/docs/platform/image-upload) . See [Frame anchors](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#frame-anchors) below. ### [Driving with a prompt](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#driving-with-a-prompt) The single rule that carries every Aleph edit: **name only what changes**. The clip already contains the framing, the cuts, the lighting, the subject, the camera, the motion. Describing any of those tells the model they are fair game, and the part you wanted preserved starts drifting. The verbs the model honors cleanly are the ones a colorist or compositor would reach for: *change*, *replace*, *swap*, *add*, *remove*, *restyle*, *relight*. Lead with one of those and name the target attribute. End the prompt with a short clause naming what the model must hold in place for anything you might otherwise leave open to interpretation. Below, the same source clip drives three independent edits. Tap through the tabs to see the baseline and each transformation. Each prompt names a single change and explicitly pins the rest of the scene in place. [Watch video](https://runware.ai/docs/assets/source-sedan.CnBxiqDj.mp4) *Source clip* > **Prompt**: A sleek black luxury sedan with polished bodywork driving slowly down a wide sunlit upscale city avenue lined with modern glass-fronted boutiques and tall trees in soft afternoon light. Slow tracking shot from the side at hood height, shallow depth of field, warm golden hour light from the left, faint reflections of the buildings sliding across the bodywork. [Watch video](https://runware.ai/docs/assets/output-color-swap.B1soYIlt.mp4) *Color swap: deep crimson* > **Prompt**: Change the sedan's bodywork color to a deep glossy crimson red. Keep the chrome trim, glass, headlights, road, and surroundings exactly as in the source. [Watch video](https://runware.ai/docs/assets/output-background.DFuhH9aF.mp4) *Background replacement: coastal road at golden hour* > **Prompt**: Replace the urban avenue setting with a winding coastal cliffside road at golden hour, ocean visible on the horizon, warm late-afternoon light. The sedan, its color, framing, and motion stay exactly as in the source. [Watch video](https://runware.ai/docs/assets/output-restyle.BAFlCHW1.mp4) *Restyle: 1980s neon-noir grade* > **Prompt**: Restyle the clip as 1980s neon-noir cinema: vivid magenta and cyan rim lighting on the sedan, deep cinematic shadows, soft volumetric haze drifting through the avenue, slight film grain. Keep the motion, framing, and composition exactly as in the source. Three independent runs against one shared source clip. The crimson run changes the paint and leaves the road untouched. The coastal run keeps the car identical and redraws the environment around it. The neon-noir run re-grades the look without touching the geometry. The shared closing pattern across all of them is the same: a short clause naming what the model must hold in place. That clause is what stops a "change the color" prompt from also rebuilding the road. > [!NOTE] > Prompts that read like art direction notes tend to outperform prompts that read like image captions. "Change the bodywork to deep crimson, keep the chrome and surroundings" is a direction. "A red sedan drives down a sunlit avenue" is a caption, and the model has to guess which parts to honor and which parts to redraw. ### [Frame anchors](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#frame-anchors) Prompts handle most edits. They struggle when the target look is something you cannot describe precisely in words: a brand identity, a custom typography lockup, a specific pattern, a packshot design that has to match a print campaign. Spelling those out in a prompt risks the model approximating them, and approximation in advertising is the wrong answer. The model's answer is the **frame-image anchor**. You pull a still from the source video, edit that still externally with whatever tool gives you the precision you need, then send the edited frame back as a reference. The model treats the anchor as the **target appearance** at a specific moment in the clip and carries that look through every other frame. #### [The workflow](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#the-workflow) The sneaker source below is a plain white leather shoe sitting on a black matte surface while the camera orbits around it, no logos, no stripes, no stitching of any kind. The goal is to brand it: a bold red side stripe running from heel to laces, the wordmark "STRYDE" on the heel counter in thin black sans-serif type, red contrast stitching along the panel seams, and a pale blue accent line along the upper sole edge. Prompts can describe the shape of that identity, but they cannot reproduce the exact stripe geometry. The letterform drifts and the stitch pattern wanders between frames. [Watch video](https://runware.ai/docs/assets/source-sneaker.CFKEZjQO.mp4) *Source clip* > **Prompt**: A pristine plain white leather sneaker with no logos, no branding, and no markings sits centered on a smooth black matte surface in a softly lit minimal studio. Slow cinematic camera orbit around the stationary sneaker at eye level, revealing every angle of the shoe in one continuous smooth motion. Even soft rim light from above and behind, gentle reflections of the studio lighting on the leather. Macro product cinematography, photoreal, cinematic premium footwear commercial. Plain dark backdrop. The first frame of that clip is the starting point for the anchor. Extract it, edit it externally with whatever tool gives you the design control you need, then send the result back as the reference image. ![First frame extracted from the source clip: a plain white leather sneaker on a black matte surface, no markings](https://runware.ai/docs/assets/source-sneaker-first-frame.Bmk3pvDW_S94pD.jpg) *Source first frame* ![Edited version of the same first frame: the same white sneaker now with a red side stripe, the wordmark STRYDE on the heel counter, red contrast stitching, and a pale blue sole accent](https://runware.ai/docs/assets/anchor-sneaker.k1y5JTHa_3rG7k.jpg) *Anchor: same frame, edited externally* > **Prompt**: Edit image 1. Apply a specific brand identity to the plain white sneaker: a single bold solid red stripe running along the side panel from the heel to the laces, the wordmark "STRYDE" rendered in clean thin black sans-serif type on the heel counter, fine matching red contrast stitching along the panel seams, and a thin pale blue accent line along the upper edge of the white rubber sole. Keep everything else identical to image 1: the sneaker shape, the camera angle, the black matte surface, the lighting, the reflections, the framing, the depth of field. The anchor matches the source's framing and lighting exactly, with only the brand identity added on top. Now the same prompt is sent to Aleph twice, once without the anchor and once with it. [Watch video](https://runware.ai/docs/assets/output-anchor-without.ZKx6Iy-f.mp4) *Prompt only* > **Prompt**: Apply a brand identity to the plain white sneaker: a bold red side stripe running from heel to laces, the wordmark "STRYDE" on the heel counter in thin black sans-serif type, matching red contrast stitching along the panel seams, and a thin pale blue accent along the upper sole edge. Keep the sneaker shape, the camera orbit, the matte surface, lighting, and framing exactly as in the source. [Watch video](https://runware.ai/docs/assets/output-anchor-with.BslTMerB.mp4) *Prompt + first-frame anchor* > **Prompt**: Apply the brand identity shown in the reference image to the plain white sneaker: the red side stripe, the "STRYDE" wordmark on the heel counter, the red contrast stitching, and the pale blue sole accent. Keep the sneaker shape, the camera orbit, the matte surface, lighting, and framing exactly as in the source. The prompt-only run produces a plausible brand identity, but the stripe geometry shifts as the camera orbits the shoe and the letterform on the heel counter drifts between frames. A brand team would catch the inconsistency and reject the take. The anchored run treats the design in the reference image as the truth and carries that exact identity through the full orbit. #### [Pinning the anchor in time](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#pinning-the-anchor-in-time) Each anchor is pinned to a position in the clip. The schema accepts two ways to do it: - `frame` takes a named position. `"first"` or `0` pins the anchor to the opening frame. `"last"` or `-1` pins it to the closing frame. This is the right call when you want the look locked across the whole clip from one of its endpoints. - `timestamp` takes a number in seconds with 0.01 precision, up to the duration of the clip. Use this when the change happens mid-clip, for example a brand reveal at three seconds, or when the look needs to evolve and you want to mark a waypoint. The most common pattern is a single anchor at `frame: "first"`. The model takes the opening look as the target and renders the rest of the clip to match. Reach for multiple anchors when **the look itself needs to change over time**: one anchor at `frame: "first"` and one at `frame: "last"` lets the model interpolate between two target states, for example a daytime scene that transitions to night, or a product that visually transforms across the runtime. Up to five anchors total are allowed. ```json [ { "taskType": "videoInference", "model": "runway:aleph@2.0", "inputs": { "video": "https://example.com/sneaker.mp4", "frameImages": [ { "image": "https://example.com/anchor-brand.jpg", "frame": "first" } ] }, "positivePrompt": "Apply the brand identity shown in the reference image to the plain white sneaker. Keep the sneaker shape, the camera orbit, the matte surface, lighting, and framing exactly as in the source.", "outputType": "URL" } ] ``` > [!NOTE] > The anchor frame does not have to be a literal extraction of the source. As long as the framing matches the source clip at the pinned moment, anything that communicates the target look works: a designer's mockup, an image-edit pass, a hand-painted reference. The closer the framing matches the moment you pinned it to, the cleaner the propagation. ### [Multi-shot footage](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#multi-shot-footage) A 30-second window often holds more than one shot. Brand reels cut between a close-up of the product, a mid shot of the person using it, and a wide of the location. Aleph applies an edit across **every cut inside the source clip** without you splitting the timeline. The model treats the cuts as part of the source structure and carries a single edit through all of them, holding the cuts in place. The reel below is three shots stitched into one twelve-second clip: a hero studio shot of a black designer handbag, a tracking shot of a model walking down a cobblestone street carrying the same bag, and a top-down shot of the bag being set down on a boutique counter. A single Aleph prompt changes the bag's leather across all three. [Watch video](https://runware.ai/docs/assets/source-multishot.NL6obD3K.mp4) *Source reel: three cuts in one clip* [Watch video](https://runware.ai/docs/assets/output-multishot.BHSVOfNZ.mp4) *One edit propagated across all three cuts* > **Prompt**: Change the handbag's leather from black to a deep glossy crimson red across every shot of the clip. Keep the bag's silhouette, the gold-toned hardware, the marble pedestal, the cobblestone street, the woman in the cream coat, the boutique counter, the cuts, and all motion exactly as in the source. The bag is now crimson on the studio pedestal, crimson at the model's side as she walks, and crimson as it lands on the boutique counter. The gold hardware, the model's coat, the street, and the marble surfaces stay exactly as they were. The cut points stay where they were. The same edit applied shot by shot would have needed three separate API calls and three manual stitches. > [!WARNING] > Multi-shot propagation depends on the model recognizing the subject as continuous across cuts. Sharp identity jumps between shots (a different actor or a different product variant) can break the propagation and produce inconsistent edits per shot. If the shots in your reel are visually disjoint, edit them as separate clips and stitch afterwards. ### [Source video](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#source-video) Four constraints set the boundaries of what Aleph can accept and what it returns: - **Duration is 2 to 30 seconds.** Anything outside this range is rejected. For longer footage, split the clip externally, edit each chunk, and stitch the outputs. - **Resolution is up to 1080p.** The output preserves the input aspect ratio. A 16:9 source returns 16:9, a 9:16 social cut returns 9:16, a 1:1 product loop returns 1:1. - **Stable, well-lit footage edits cleanest.** Crisp subjects and even lighting give the model a clean signal to preserve. Slow-to-moderate camera moves keep that signal coherent through the runtime. - **Motion blur and low light degrade the signal.** Camera shake adds frame-to-frame drift to the edit. The same prompt on a steady shot and a hand-held shake produces noticeably more wobble on the latter. Stabilize before sending if you have the option. ### [Tips](https://runware.ai/docs/models/runway-aleph-2-0/guides/editing-video#tips) 1. **Lead with a transformation verb.** The verbs Aleph honors cleanly are *change*, *replace*, *swap*, *add*, *remove*, *restyle*, *relight*. A prompt that starts with one of them communicates that the source clip is the baseline. A prompt that reads like an image caption asks the model to redraw from scratch. 2. **Always pin the rest of the scene in place.** End the prompt with a short clause naming what the model must hold ("keep the framing, the lighting, the road"). Without it, the model treats unmentioned elements as fair game and the clip drifts away from your source. 3. **Reach for a frame anchor the moment the prompt feels vague.** Brand assets, exact typography, custom patterns, and any look a designer signed off on belong in an anchor frame, not a prompt string. The prompt names the change, the anchor defines the truth. 4. **Match the anchor framing to its pinned moment.** An anchor pinned to `frame: "first"` should look like the source's opening frame with only the intended edit applied. A wildly different composition is harder for the model to map onto the rest of the clip. 5. **Edit multi-shot reels as one clip when the subject is continuous.** Aleph handles cuts on its own when the subject and setting carry through. Splitting and re-stitching produces inconsistencies that the single-call path avoids. 6. **Stabilize and clean the source before editing.** Aleph preserves what it receives. Bring the shot up to delivery quality first, then apply the edit, rather than relying on the model to fix grade or stabilization at the same time as a content change. 7. **Use a generator when the goal is to invent, not transform.** Aleph edits what is already there. Adding a creature flying through empty sky, generating an entirely new scene, or producing footage of something that was never shot are jobs for [Seedance](https://runware.ai/docs/models/bytedance-seedance-2-0) , [HappyHorse](https://runware.ai/docs/models/alibaba-happyhorse-1-0) , [Veo](https://runware.ai/docs/models/google-veo-3-1-lite) , or another text-to-video model. Pair them with Aleph for the polish pass.