Protoface Seedance 2.0 Fast · Text to Video

What this node actually is

You can't download Seedance. ByteDance's video line is closed in both directions - every Seedance and Seedream model stays on an API while the company gives away the components around them (PuLID, Depth Anything, Bernini) for free. So a node called "Seedance 2.0 Fast · Text to Video" is not a sampler. It's an HTTP client with your billing attached, and it lives in the Protoface pack, which is a third-party aggregator wrapping a few dozen closed models behind one key.

Fast is the middle rung of three Seedance 2.0 tiers. Not confusing at all. Mini is cheapest, vanilla 2.0 is the flagship, Fast sits between them and - this is the part the marketing copy glosses over - it only goes up to 720p. No 1080p, no 4K. If you need either, this is the wrong node and the regular Seedance 2.0 node is the right one.

How it works, mechanically

The node collects a prompt, a handful of settings, and any reference media you attach, uploads the media to Protoface, POSTs a job to their API, and polls until a video comes back. Then it downloads the result and hands ComfyUI a native VIDEO object as if a local model had produced it. There's no checkpoint, no VRAM cost, no node graph under the hood.

Under the pack's surface there's one shared implementation across all ~44 models - the per-model controls are generated from the provider's public request schema. That's why the tooltips read like API docs and why new models mostly arrive through a catalog refresh rather than a pack update.

The controls you'll actually touch

prompt is the big one: describe scene, action, camera movement, mood. Treat it like a shot brief rather than a tag soup - these models read full sentences.

resolution gives you 480p or 720p. aspect_ratio covers 16:9, 9:16, 4:3, 3:4, 1:1 and 21:9. duration is 4–15 seconds, integer.

generate_audio defaults to true, and that's worth noticing before your first queue. These nodes generate synchronous native audio - you're paying for a soundtrack whether or not you wanted one.

reference_images, reference_videos and reference_audios are the interesting ones, and they're why the "text to video" label undersells the node. They're dynamic list inputs: connect native ComfyUI Image/Video/Audio nodes (or existing asset ids and public URLs) and the list grows as you add them. Character consistency from a subject photo, style transfer from a reference clip, an audio bed to lip-read against - all doable from this node, no separate "reference to video" variant needed.

Then the two settings that aren't model parameters at all. generation_id is the paid-job handle: keep it fixed and requeue, and you resume the same run instead of buying a new one; increment it for a genuinely new generation. It is not a seed, and there is no seed field on these video nodes - reproducibility comes from the resume machinery, not from a number you can reproduce later. request_scope is auto-assigned per node and only matters if you're queueing from a script.

Outputs: video (a list, even when it's one clip), files (every original asset downloaded, which is what you want if the provider hands back something odd), and run id - a string worth keeping, because it's what the Protoface Resume Run node takes when a poll times out.

Install

ComfyUI 0.41.0+ and Python 3.10+. Manager route: Extensions → search Protoface → install protoface-nodes → restart and reload the browser. Manual route:

cd ComfyUI
git clone --branch v0.1.0 --depth 1 https://github.com/protoface-ai/protoface-comfyui.git custom_nodes/protoface
python -m pip install -r custom_nodes/protoface/requirements.txt

Use ComfyUI's own interpreter for that pip line - python_embeded/python.exe on the Windows portable build. The requirements are light (httpx, jsonschema, Pillow, PyAV), so nothing here fights with torch. Then click Protoface in the lower-right corner and paste an API key from your Protoface account. For headless boxes, set PROTOFACE_API_KEY (or PROTOFACE_API_KEY_FILE) instead.

What bites people

The 720p ceiling. People queue a draft here, love it, then have to switch node to get 1080p - which means a new node, a new generation_id, and a fresh charge. Decide the resolution tier before you iterate.

Costs are per output second, and reference video costs extra. At 720p this tier runs about 12 credits per second, so a 5-second clip is ~60 credits - and if you attached a reference_videos clip, its input seconds are metered on top, capped at 15 seconds total. That's the sneaky part of reference-driven generation: the reference is not free.

Worth it?

For draft-to-final iteration inside one graph, yes - 720p Fast is the sane middle for "does this shot work?" money. For anything destined for 1080p, skip the tier and go straight to Seedance 2.0 or 2.5. And watch generation_id: on a paid node, that integer is the difference between refining a shot and rebuying it.