Seedance 2.0 Fast · Text to Video

Text to video at 720p with sound, from ByteDance's faster 2.0 tier. It's the node you use when you have a shot in your head and no reference to give it.

And it has a checkbox almost nobody expects: enable_web_search. Turn it on and the model searches the web for information related to your prompt before rendering - which sounds like a curiosity until you want something vaguely current or specific ("a 1970s rally car on a gravel stage") and don't want to trust the training set. The tooltip is also honest that requests with it on are billed at the web-search rate, and in the shipped price table that's a jump of roughly half again over the standard per-second rate. So: off unless you need it.

How it works

The node assembles a request from your widgets, quotes it, submits it under an idempotency key so a retry can't double-charge, polls with live status printed on the node, and downloads the MP4 as a ComfyUI VIDEO. Fields left at defaults are omitted so the model picks. Nothing runs locally; the model itself is closed, so calling it is the only route, and the trade is per-call metering instead of GPU time.

The inputs that matter

prompt - subject, action, camera, light, sound and constraints. Because generate_audio is on by default, include the sound. Models produce a generic wind-and-hum ambient if you don't say otherwise.

duration_seconds - 4 to 15 seconds, or -1 for smart duration where the model decides. The billed quantity, so it's also your main cost control.

resolution - 480p or 720p. Explore at 480p; the money is better spent on takes than pixels at this tier.

aspect_ratio - adaptive by default, plus 16:9, 9:16, 4:3, 3:4, 1:1, 21:9. With no source image, adaptive effectively means the model chooses, so set it explicitly if you care about the frame.

generate_audio - request synchronised output audio. For 2.0 Fast, unlike 1.5 Pro, the per-second rate in the shipped table doesn't change with this switch, so there's no reason to turn it off unless you hate the track.

enable_web_search - described above. Use it for specificity, not as a default.

return_last_frame - returns the clip's last frame as an IMAGE on the last frame output. With it off, that socket blocks anything downstream rather than emitting a blank.

bitrate_mode - standard or high. Encoding only; high is worth it for anything you'll grade or deliver.

watermark - off by default. Leave it.

seed - fixed by default. Same graph, same clip, no second bill. Randomise when you mean it.

Outputs

video, last frame (needs return_last_frame), and task id. The last-frame output is what makes multi-shot sequences workable in one graph: shot A's final frame becomes shot B's opening image.

Installing it

ComfyUI Manager → search PopNow → Install → restart. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Then a key from popnow.ai/api/console/keys, entered under Settings > PopNow or exported as POPNOW_API_KEY, and a second restart so your account's catalog becomes the node list. No model downloads; the pack installs nothing beyond what ComfyUI already ships.

Where people get burned

Web search pricing is the sneaky one: leave the toggle on out of curiosity and every run after that is quietly more expensive. It's the first thing to check when a quote looks wrong.

Smart duration (-1) plus web search is the combination that produces quotes people don't expect - both make the run less predictable, and only the quote stage protects you. Set Settings > PopNow > Max credits per run.

Text-only video also fails in a specific way worth knowing: with no image anchoring the look, the model's house style shows through harder, and text or logos in frame come out as decorative nonsense. If a shot needs exact branding, generate the still first and use Seedance 2.0 Fast · Image to Video - it starts from your pixels instead of guessing.