Bytedance Seedance 2.5 Video Edit

Video-to-video editing by prompt is the workflow that made Bernini and VACE local legends - take an existing clip, describe the change, get a rewritten clip back. Seedance 2.5 Video Edit is the hosted version of that idea, and it's the pack's cleanest building block: it takes the source clip as a URL, so it chains directly off the video_url output of any video node in this pack. Generate a T2V clip, wire its output straight into this node, describe the change, and you've got an edit without ever touching a frame encoder.

The service prepends "Edit the input video." to your prompt - that's how it frames the job. Your prompt is the "what changes": "make it dusk and add rain", "change the character's jacket to red", "turn this into a slow-motion close-up." Output duration and aspect ratio follow the input clip, so you're not re-deciding the canvas on every edit.

How it works

The video_url input is the heart of it. Clips over 30 seconds get trimmed; clips under 4 seconds get padded - the model has a working window and normalizes you into it. From there the node sends your prompt, the video, optional references, resolution, and audio flag to WaveSpeed, then polls every 10 seconds with a 30-minute timeout before returning a URL.

The model dropdown offers Seedance 2.5 and Seedance 2.5 Turbo. Reference support mirrors the T2V node: reference_images (up to 4, or an IMAGE batch via reference_images_tensor) for style/identity/appearance, and reference_audios (up to 4) for soundtrack or voice guidance.

Inputs that matter

• prompt - required, what to change.
• video_url - the source clip. This is where the chaining magic lives: drag the video_url output from a T2V, I2V, or Extend node in here.
• model - standard or Turbo.
• resolution - 480p / 720p / 1080p / 4k. Read the billing note: it counts input plus output duration, roughly $0.11/s at 480p up to $1.10/s at 4k. Editing a clip costs you the clip you feed it and the clip it makes.
• generate_audio - default on; synchronized audio for the edited video.
• seed - cache control only. Fixed seed reuses the already-paid edit; -1 reruns.

Output is video_url (STRING) → Preview Anything for preview/download. And because it's a URL output, the edited clip can feed straight into another node - including Extend.

Installing and configuring

Part of the ERPK Collection pack. ComfyUI Manager: search erpk, install ERPK Custom Nodes, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

WaveSpeed API key from wavespeed.ai into Settings > ERPK > API Keys (right-click canvas > ERPK Settings). Server-side generation, nothing to download, no GPU required.

Common issues

• "A source video URL is required" - you queued without wiring video_url. Every video node in the pack emits one; connect it.
• The edit is way shorter/longer than my clip - clips are trimmed to 30s and padded past 4s before the model sees them. Long source material gets windowed; that's by design.
• The bill is steeper than a generation - edit billing counts input plus output duration. Editing a 20s clip at 1080p costs more than generating a 20s clip; budget accordingly.
• Nothing previews - video_url is a link, not a tensor. Preview Image won't show it; use Preview Anything.