Video
Video edit
Video edit changes an existing clip with a text prompt: add or remove an object, swap a character, change the
background, put a garment on a person, or restyle the whole thing. The motion of the source stays; the clip you get back has
the same length. It is the same endpoint as generation, POST /v1/videos,
with one extra field: input_video_url.
There are 37 edit models across Decart, Kling, ByteDance Seedance, Alibaba Wan, Google Gemini Omni, xAI Grok Imagine, Runway, Pika and FLUX, hosted on several providers. Editing a live camera stream instead of a recorded file is live video.
Quickstart
Submit the clip and a prompt. You get a job id back straight away; the edit runs asynchronously.
curl https://videorouter.sh/api/v1/videos \
-H "Authorization: Bearer llmr_sk_live_..." \
-H "Content-Type: application/json" \
-d '{
"model": "decart/lucy-2.6-edit",
"prompt": "Give the person red sunglasses",
"input_video_url": "https://example.com/source-clip.mp4"
}'Poll the job until its status is completed (or failed). The finished clip is a VideoRouter-hosted URL in data[0].url.
curl https://videorouter.sh/api/v1/videos/decart:3f9c1c2e-... \
-H "Authorization: Bearer llmr_sk_live_..."
# {"id": "decart:3f9c...", "status": "completed", "data": [{"url": "https://..."}], "cost_usd": 0.246, ...}The same flow end to end in Python:
import time, requests
API = "https://videorouter.sh/api/v1"
HEADERS = {"Authorization": "Bearer llmr_sk_live_..."}
job = requests.post(f"{API}/videos", headers=HEADERS, json={
"model": "wavespeed/wan-3.0-edit",
"prompt": "Change the flower to a bright red flower",
"input_video_url": "https://example.com/source-clip.mp4",
"resolution": "720p",
}).json()
while True:
status = requests.get(f"{API}/videos/{job['id']}", headers=HEADERS).json()
if status["status"] in ("completed", "failed"):
break
time.sleep(5)
print(status["data"][0]["url"])Statuses are queued, in_progress, completed and failed. A failed edit is refunded. See Async jobs for how video jobs are polled.
How edit differs from generation
There is one video endpoint, and the fields you send decide the mode. A prompt on its own is text-to-video; add
input_video_url and an edit model and it becomes an edit. Edit models are marked Video Edit on the
models page, and a model that is not an edit model refuses
input_video_url with a 400 that lists the ones that are.
promptText-to-videoe.g. kling-v3 — a prompt and nothing else.start_image_urlImage-to-videoThe image is the first frame.input_referencesReference-to-videoImages, clips or audio that guide a new generation.input_video_url + promptVideo edit (this page)The clip is modified; the output keeps its length and motion.Edit cannot be mixed with input_references, input_video_references, input_audio_url or end_image_url in the same call.
Parameters
modelrequiredAn edit model from the table below (the same ids GET /v1/videos/models lists).promptrequiredWhat to change. An empty prompt is a 400. Describe only the change; leave the rest of the scene out.input_video_urlrequiredA public HTTPS URL of the source clip (MP4). One clip, a single URL, not an array.start_image_urloptionalA reference image that guides the edit (a garment, an object, a look). Only models with “yes” in the Ref. column accept it; the others return 400.resolutionoptionalThe output tier, for models with more than one (e.g. 480p / 720p / 1080p). The default is the model’s own.
duration_secs, aspect_ratio and seed do not apply to edits: the output follows the source clip.
Editing with a reference image
Models marked yes in the table below take one reference image in start_image_url. Use it to show the model exactly what to add or wear.
{
"model": "decart/lucy-vton-3.6-edit",
"prompt": "Put the jacket from the reference image on the person",
"input_video_url": "https://example.com/source-clip.mp4",
"start_image_url": "https://example.com/jacket.png"
}Models and prices
Prices are dollars per second of your source clip, at the provider’s own rate and before our platform fee. Where a model has more than one resolution, each tier is shown. A model hosted by several providers appears once per host and can fail over between them (see Model fallbacks). The same ids are on /models.
decart/lucy-2.6-editNewest Lucy. Add, replace or remove objects, swap characters, change backgrounds. Optional reference image.Decart$0.04120 syesdecart/lucy-2.5-editPrevious Lucy generation, same price.Decart$0.04120 syesdecart/lucy-vton-3.6-editVirtual try-on: put a garment on the person. Send the garment as start_image_url.Decart$0.04120 syesdecart/lucy-vton-3.5-editPrevious try-on generation.Decart$0.04120 syesdecart/lucy-restyle-2-editRestyles the whole clip. The returned clip can run longer than the source; you are billed on the source.Decart$0.01120 syeswavespeed/kling-v4.0-editKling 4.0 edit. Rate covers the source plus the output (about 2x the per-output rate).WaveSpeed720p $0.188 · 1080p $0.28230 snoatlascloud/kling-v4.0-editAtlas Cloud720p $0.171 · 1080p $0.25730 snofal/kling-v4.0-editFal720p $0.188 · 1080p $0.28230 snorunware/kling-v4.0-editRunware720p $0.188 · 1080p $0.28230 snosandbase/kling-v4.0-editSandBase720p $0.16 · 1080p $0.2430 snowavespeed/kling-v4.0-flash-editKling 4.0 Flash edit, 720p.WaveSpeed720p $0.12620 snoatlascloud/kling-v4.0-flash-editAtlas Cloud720p $0.11420 snorunware/kling-v4.0-flash-editRunware720p $0.12620 snowavespeed/kling-o3-std-editKling O3 (the 3.0 omni generation). Source at least 3 s and 700 px tall. Optional reference images.WaveSpeed$0.12610 syeswavespeed/kling-o3-pro-editWaveSpeed$0.16810 syesatlascloud/kling-o3-std-editAtlas Cloud$0.10710 syesatlascloud/kling-o3-pro-editAtlas Cloud$0.14310 syesreplicate/kling-v3-omni-videoKling 3.0 Omni. Sending input_video_url switches it from reference mode to true editing. Source up to 10 s.Replicatestandard $0.168 · pro $0.224 · 4k $0.4210 snowavespeed/seedance-2.5-editSeedance 2.5. The output is padded to at least 4 s.WaveSpeed480p $0.198 · 720p $0.396 · 1080p $0.9915 syeswavespeed/seedance-2.5-flash-editSeedance 2.5 Flash, 480p / 720p. Source up to 30 s.WaveSpeed480p $0.18 · 720p $0.3630 snowavespeed/seedance-2.0-editSeedance 2.0. At least 4 s of output.WaveSpeed480p $0.135 · 720p $0.27 · 1080p $0.67515 syeswavespeed/seedance-2.0-fast-editWaveSpeed480p $0.117 · 720p $0.23415 syeswavespeed/seedance-2.0-mini-editWaveSpeed480p $0.0675 · 720p $0.13515 syeswavespeed/wan-3.0-editWan 3.0. Optional reference images.WaveSpeed480p $0.1 · 720p $0.2 · 1080p $0.415 syeswavespeed/wan-3.0-prime-editWan 3.0 Prime.WaveSpeed480p $0.15 · 720p $0.3 · 1080p $0.615 syesatlascloud/wan-2.7-editWan 2.7.Atlas Cloud720p $0.2 · 1080p $0.310 syesatlascloud/happyhorse-1.0-editHappyHorse 1.0. Billed on the source plus the output.Atlas Cloud720p $0.28 · 1080p $0.4810 syeswavespeed/gemini-omni-1.2-flash-editGemini Omni 1.2 Flash. At least 3 s is billed. Optional reference images.WaveSpeed360p $0.048 · 720p $0.104 · 1080p $0.14610 syeswavespeed/grok-imagine-video-editGrok Imagine. The lowest-priced host of this model.WaveSpeed$0.0658.7 snoatlascloud/grok-imagine-video-editSame model on Atlas Cloud (failover).Atlas Cloud$0.088.7 snorunway/aleph-2Runway Aleph 2.0. Source up to 30 s.Runway$0.2830 snopika/pikadditionsPika Pikadditions: insert an object or character. Flat $0.03 per request, whatever the length.Pika$0.03 / request—noflux-3-edit-videoFLUX 3 Edit Video from Black Forest Labs, direct.Black Forest Labs$0.0315 snofal/flux-3-edit-videoFal$0.0315 snoatlascloud/flux-3-edit-videoAtlas Cloud$0.03315 snowavespeed/flux-3-edit-videoWaveSpeed$0.0315 snowavespeed/pruna-p-video-editPruna P-Video edit. Optional reference images.WaveSpeed$0.04510 syesBilling and limits
- You pay for the length of your source clip. We measure it when you submit. The output keeps that length, so there is no duration to choose.
- Whole seconds. Most hosts round the clip up to the next whole second. Some bill a minimum: Seedance 2.0 and 2.5 pad the output to at least 4 s, and Gemini Omni and Kling O3 bill at least 3 s.
- Exact charge. For WaveSpeed and Atlas Cloud models the final charge is the host’s own cost for your exact clip, so it can differ slightly from the table. Some models also count the output seconds (Kling 4.0, HappyHorse); their rate already includes both.
- Long clips. The Max clip column is the longest source we bill. A longer clip may be trimmed or refused by the host.
- A hold, then a charge. We hold the estimated cost when you submit and settle it when the job finishes. A failed job releases the hold in full.
- Clip quality. Some hosts reject low-resolution or very short sources. Kling O3 needs at least 3 s and 700 px of height.
Errors
model ‘X’ does not support input_video_url (video editing/upscaling)The model is not an edit model. The message lists the ones that are.requires a non-empty prompt describing the editEdit models need a prompt.cannot combine input_video_url (video editing) with start_image_urlThat model takes no reference image.cannot combine input_video_url (video editing) with input_references/…Edit is its own mode. Use one input style per request.insufficient available credits to reserve $…Your wallet cannot cover the hold for this clip. Add credits, or send a shorter clip.all video providers failed, with input_video_url could not be downloaded in the attemptDecart models fetch the clip through our gateway: it must be a public MP4 of at most 60 MB. Nothing is charged.failed with the host’s reason in errorThe host rejected the clip after accepting it, for example under 3 s or under 700 px tall on Kling O3. The charge is refunded.General error shapes and rate limits are in Errors & rate limits.