Bytedance Seedance 2.5 Text-to-Video

Text-to-video these days isn't just a prompt in a box. Seedance 2.5 Text-to-Video takes the full modern form: a text prompt as the core, but up to four reference images, four reference videos, and four reference audios to steer style, character, motion, and sound. It's the flagship of the Seedance 2.5 family in this pack - 4–30 second clips, up to 4k output, native synchronized audio - and it's the node to reach for when you want a long, high-res shot built from a description, optionally anchored to reference material so it doesn't drift into generic-land.

Same family story as the other Seedance 2.5 nodes: ByteDance's closed video line, hosted through WaveSpeed's API, nothing running on your machine. Same billing reality too - the resolution tooltip doesn't mince words: roughly $0.18/s at 480p climbing to $1.80/s at 4k. 4k is the expensive toy here; 720p is the sensible default for most work.

How it works

The model dropdown gives you Seedance 2.5 and Seedance 2.5 Turbo (faster, no Spicy tier - that one is i2v-only). The node sends your prompt, references, duration, aspect ratio, resolution, and audio flag to WaveSpeed, then polls every 10 seconds with a 30-minute timeout before returning a URL.

References come in two flavors of input, same as the pack's other video nodes:

• reference_images - up to 4, single URL or list, for style/character/composition guidance. Or connect reference_images_tensor (an IMAGE batch, each slice one reference, capped at 4) and the tensor wins.
• reference_videos - up to 4, total combined duration up to 30s, for motion and pacing guidance. Read the billing note in the tooltip: runs that use reference videos are billed on the combined reference and output duration.
• reference_audios - up to 4, total 30s, for soundtrack guidance.

Inputs that matter

• prompt - required, the video description.
• model - Seedance 2.5 or Turbo.
• duration - 4–30s.
• aspect_ratio - six options: 16:9, 9:16, 4:3, 3:4, 1:1, 21:9. (The i2v sibling doesn't offer this, because the image sets the canvas - here you choose it yourself.)
• resolution - 480p / 720p / 1080p / 4k.
• generate_audio - default on; native synchronized audio. Flip off if you're compositing sound elsewhere.
• seed - cache control only, never sent to the API. Fixed seed reuses the video you already paid for; -1 generates fresh.

Output is video_url (STRING) → Preview Anything for preview/download.

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). No downloads, no GPU needed - the server generates, the node polls and hands back a link.

Common issues

• Cost creep with reference videos - reference video runs bill on the combined reference + output duration, so a reference-heavy run is pricier than it looks. Check the tooltip before queueing.
• References ignored - unlike MiniMax H3's bracket-tag system, these references are positional, not cited. If they're not shaping the result, they're still being sent - but steering the prompt to reference "the subject from the reference image" helps the model use them.
• The 30-minute wait - long clips at high resolution legitimately take several minutes. That's polling working as intended, not a hang.