Video to Video

The Video to Video node transforms an existing video from a text prompt. It keeps the motion of the source clip and changes what you describe: the style, the light, the season or a detail. Use it for style transfer, for variations of existing footage and, with Seedance 2.5 Edit, for edits written as instructions such as make it black and white.

When to use it

• Recorded footage should look painterly, anime, claymation or cinematic.
• A scene should change its conditions, for example from day to night or from summer to winter.
• You want several variations of the same clip.
• You want to edit a clip by instruction and bring in an outfit, an object or a place from reference images. Use Seedance 2.5 Edit.

To change only a few seconds of a clip, use Retake Video. To relight a subject or swap a background from the clip's own pixels, use Relight & Switch.

Quick start

Inputs

The two reference inputs are always visible. On every other model they are dimmed and report that the model does not support them, and a connection to them has no effect.

The output, Video, is the transformed clip.

Settings

Each model adds its own settings:

Models

Video to Video offers six models. Click a model for its settings and credit prices.

The node also offers Luma Modify, which is strong at style transfer and artistic changes.

Which model to choose

The model table uses the catalog names: Wan 2.6 Flash is listed as Wan Flash V2V, and Seedance 2.5 Edit as Seedance 2.5.

• Wan 2.6 Flash is the cheapest and the fastest. Use it to test an idea.
• Wan 2.6 is a general-purpose model with reliable results.
• Wan 2.7 VideoEdit adds control over the sound and can expand a short prompt with AI.
• Luma Modify is strong at style transfer and artistic changes.
• HappyHorse Edit accepts source clips up to 60 seconds long.
• Seedance 2.5 Edit edits the clip by instruction, keeps its length and aspect ratio, and accepts image and audio references.

Edit by instruction with Seedance 2.5 Edit

Seedance edits a video that it receives as a reference when the prompt reads as an edit instruction. The node prepares the run for you:

• The source clip becomes reference video 1. Your prompt is sent as edit @video_1 as follows: followed by your text. If your prompt already starts with that instruction, it is not added twice. The clip is re-encoded first, with the same resolution, frames, length and sound, so that a clip Seedance made itself is edited instead of coming back as a near-copy. The length and the price do not change.
• The result keeps the source clip's length and aspect ratio. That is why this model has no Duration and no Aspect Ratio setting.
• The source clip must be 4 to 30 seconds long. Shorter or longer clips are refused.
• Seedance is told the run is an edit. When every clip is 4 to 30 seconds long, the run also tells Seedance outright that the clip is the video to edit, a draft included. This is the same choice as The video to edit on Generate Video. A single run, a workflow run and an app run do the same.
• Negative prompts are not sent. Write what to avoid into the instruction itself.

Write the prompt as an instruction about the clip:

make it black and white, heavy film grain

Wire images into Image Refs and cite them by position, in the order you connected them. {image:1} is the first image, {image:2} is the second, and so on. Reference audio works the same way, with {audio:1}.

the woman wears {image:1} and stands in {image:2}

Generate Video has an Edit Video preset that runs the same kind of Seedance 2.5 edit.

Presets

The Restyle Looks folder in the node's preset menu has seven ready-made looks: Anime Restyle, Claymation, Cyberpunk Neon, Oil Painting, 3D Animated, Watercolor and Phone Footage Look. The first six select Wan 2.6, a negative prompt that suppresses flicker and distortion, and a restyle instruction that keeps the original motion. Phone Footage Look gives b-roll an everyday handheld look on Wan 2.7 VideoEdit at 720p, and keeps the clip's own sound and length. It renders faces again, so do not use it on clips of people talking. The instruction is stored as pre text, so the prompt stays free for your own additions. See Presets.

Credits

• Wan 2.6, Wan 2.6 Flash, Wan 2.7 VideoEdit, Luma Modify and HappyHorse Edit are priced per model. The model table above lists the prices, and the node shows the price before you run it.
• Seedance 2.5 Edit is billed like a Seedance run with a reference video in Generate Video: the per-second rate of the chosen resolution, times the length of the source clip plus the length of the result. Because the result length is automatic, an edit renders the source clip's own length. Once the source clip is measured, the run reserves credits for that length rounded up to a whole second, plus one second, and never more than 30 seconds. It then refunds the difference down to the length actually delivered. A clip that cannot be measured is reserved at the worst case: a 30-second result, and the longest source clip the model accepts. Your balance must cover the reservation for the run to start.
• The node shows that exact reservation. For Seedance 2.5 Edit, the node's price and its Run button come from the server, which measures the source clip and prices it the way the run does. While the source clip has not been made yet, the node shows a range, from the shortest clip Seedance 2.5 accepts to the longest, and the Run button shows the top of the range. No figure is shown while the clip is being measured.
• A Seedance 2.5 Edit draft is billed exactly like the 480p edit it is. To finalize it at 1080p, wire it into Upscale Video set to Seedance 2.5 HD (from draft). The finalize is priced from the source clip's seconds plus the draft's seconds, and Upscale Video shows the exact price once the draft is wired in. See Seedance 2.5 Draft mode.

Tips

• Change the style, not the structure. Colors, textures and light change reliably. Adding or removing objects works less well. For structural changes, edit keyframes with Generate Image and animate them with Generate Video.
• Describe only the change. change to anime style works better than a description of the whole scene.
• Test cheap, finish sharp. Try the idea on Wan 2.6 Flash, then run the final version on Luma Modify.
• Iterate Seedance at 480p. The reservation grows with the resolution, so switch to 1080p only for the final run. With Draft (480p) on, you can later finalize the take you like at 1080p as the same take.
• Expand short prompts. With Wan 2.7 VideoEdit, turn on Expand prompt with AI when your prompt is very brief.
• Keep prompts consistent. When you restyle several clips of one project, use the same prompt or the same preset for all of them.

From the API

Every model of this node runs through POST /v1/video-to-video. Seedance 2.5 Edit takes this body:

{
  "videoUrl": "https://…/clip.mp4",
  "prompt": "make it black and white, heavy film grain",
  "provider": "seedance-2-5",
  "resolution": "720p",
  "draft": false,
  "audio": true,
  "seed": 42,
  "referenceImageUrl": "https://…/look.png"
}

• Required fields. videoUrl, prompt and provider are required. resolution, draft, audio, seed, referenceImageUrl, referenceImageUrls, referenceVideoUrls and referenceAudioUrls are optional.
• What the server does. The server builds the edit instruction itself, as the node does. audio turns on the clip's native sound.
• Same rules as the node. The price, the 4-to-30-second clip check and the refund down to the delivered length are the ones described in Credits.
• Seedance 2.5 Edit only. draft: true, resolution: "480p" and the three reference lists are accepted for Seedance 2.5 Edit only. Any other model answers 400.

References. Refer to each reference in the prompt with a token: {video:N}, {image:N} or {audio:N}.

• Limits. Seedance 2.5 takes 10 clips in all, the source included, 30 images and 10 audio clips. A body over a limit, or a reference audio clip that the model would refuse because it is too long, answers 400 before anything is reserved. POST /v1/quote answers the same way.
• A token past the end of its list is dropped.
• Extra clips are billed as input. Every clip, the source and the extras, is measured and billed as input seconds. When an extra clip is longer than the source, the output is reserved for that clip, because the model may edit it instead. The edit usually has the source clip's length and is refunded down to it.
• The server names the edit. When every clip, the source and the extras, is 4 to 30 seconds long, the server tells Seedance outright that the run is an edit, a draft included.
• Price it first. POST /v1/quote with "type": "video-to-video" and the same body as params returns what the run will reserve. Nothing is started or charged. For authentication and errors, see the API reference.