Bytedance Seedance 2.0 Text-to-Video

Seedance 2.0 Text-to-Video is ByteDance's video model served through WaveSpeed, and it's the most feature-dense text-to-video node in this pack. Prompt in, video URL out - but between those two points you get a four-tier model dropdown, up to four reference images and videos and audios, native synchronized audio, web search grounding, and 4–15 second clips at up to 1080p. That's a lot of dials, and most of them are genuinely useful.

The ByteDance context matters here: Seedance is an API-only product line - the KB's ByteDance panel is explicit that the entire Seedream and Seedance lines are hosted products with no open weights. So this node isn't a wrapper around something you could run locally; it's a paid API call dressed up as a graph node. What you get for that money is a strong commercial video model with native audio-visual generation, which is a capability tier most local setups can't touch.

The inputs that matter

• model - the dropdown: Seedance 2.0 (standard), Turbo (faster, 720p/1080p only), Fast (cheaper), Fast Turbo (cheap + turbo, 720p/1080p only). These route to distinct WaveSpeed endpoints with real price differences - Fast is the budget pick for iteration, standard for final results
• prompt - the video description
• reference_images - up to 4 URLs for style/character/composition guidance
• reference_videos / reference_audios - up to 4 each, total duration under 15s
• duration - 4 to 15 seconds
• aspect_ratio - 16:9, 9:16, 4:3, 3:4, 1:1, 21:9
• resolution - 480p / 720p / 1080p
• generate_audio - native synchronized audio, on by default
• enable_web_search - pulls real-time info into generation (useful for "a news-style clip about X")
• seed - reproducibility, -1 for random
• reference_images_tensor - an IMAGE batch alternative to the URL list; each batch slice becomes one reference (capped at 4), sent as base64, and it takes precedence over the URL field when connected

The output is video_url (STRING), destined for Preview Anything.

Why the references matter

This is the differentiator vs. the pack's WAN nodes: Seedance T2V accepts reference images, videos, and audio. Feed it a character shot as reference_images and a clip as reference_videos and the prompt can ask for "this character, this camera move." It's how you keep identity and motion language consistent without training anything. The reference_images_tensor variant is the ComfyUI-native path - load an image with Load Image, wire the tensor in, and the node base64-encodes it itself instead of making you host a URL.

Install and setup

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

Or "ERPK Custom Nodes" from ComfyUI Manager, restart, and set a WaveSpeed API key in ERPK Settings. Generation polls every 10 seconds with a 15-minute timeout - a 15-second 1080p clip is the slow end of that, so queue and walk away.

Troubleshooting

A few things to know before you burn credits. The Turbo variants are 720p/1080p only - select 480p with a Turbo model and you'll hit an endpoint error or a silent fallback. generate_audio on by default means every run bills for audio; turn it off for silent test runs. And the reference caps are real: more than four of anything gets truncated at four, and reference video/audio over 15s total will be rejected. If a run comes back empty, check the [WaveSpeed] log lines in the console - the API error message is usually there.