Protoface Seedance 2.5 · Edit Video

The operation, and why it's the one worth paying for

Restyling footage you already have is the capability the open ecosystem hasn't nailed. You can run Wan locally and get motion; getting this specific clip, with the performance intact, wearing a different world is a different problem. Seedance 2.5 Edit does it at up to 4K, and this is the standard (non-turbo) version - the one to use when the output matters.

What you'll notice first: the schema got simpler

The 2.0 edit node had six model-ish controls. This one has three. duration is gone - no more "auto-detected from the input" ambiguity to reason about, and no way to force a shorter cut. aspect_ratio is gone too. Output length and shape follow the source clip, which is what you wanted anyway, and the loss of rope is arguably a feature.

prompt still describes the edit, and the tooltip carries an implementation detail worth knowing: the prefix "Edit the input video." is added automatically. So you write the instruction itself - "swap the street for a snowy alley, keep the walk and the coat" - and don't waste words framing it as an edit. Write it twice and you get a slightly confused instruction.

video is the clip: native ComfyUI Video in, or a URL/data URI/asset id, ideally from a previous Protoface job via the asset-reuse path so it isn't re-uploaded.

resolution - 480p, 720p, 1080p, 4K - is your only real dial.

generate_audio defaults true, and the tooltip is the important one: set it false and the input video's original audio track is preserved on the output. For anything with speech or a licensed track, false is the correct answer and the default is wrong for you.

reference_images and reference_audios are optional dynamic lists to steer the edit with a look or a voice reference. They supplement the source clip; they don't replace the video input.

generation_id keeps the paid job reusable - identical inputs plus a fixed ID resumes rather than re-bills. Increment it to deliberately generate again. request_scope is auto-assigned per node; headless clients must supply a stable unique one.

Outputs: video (list), files (every original download), run id (feed to Resume Run after a timeout).

Install

ComfyUI 0.41.0+ and Python 3.10+.

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

Manager: search Protoface, install protoface-nodes, restart, reload the browser. Pip with ComfyUI's interpreter, not your system Python. Then paste a key via the Protoface button, or export PROTOFACE_API_KEY. Four dependencies, no models.

Where it goes wrong

You can't shorten the edit. No duration field means you edit the whole clip you pass in. Always cut the section you care about first with a local node; it's cheaper and gives you control the API doesn't offer.

Input seconds bill on top of output. The edit operation meters the source video as well as the result, and the catalog caps that input at 30 seconds - four seconds minimum. Long footage is a two-part cost, and 4K on both sides is the expensive corner of the whole pack.

Invented audio over dialogue. Default generate_audio true will re-sound your clip. Flip it.

Assuming this equals the turbo node. Same schema, narrower resolution range on the turbo side, faster turnaround. Drafting on turbo and delivering here is a sensible chain; drafting at 720p and expecting it to be the delivery is not.

Re-pushing after a crash. Don't bump generation_id. Requeue unchanged inputs and the pack resumes the recorded run; that's what the journal is for.

Verdict

The cleanest editing node in the pack: three knobs, a good output, and an automatic prefix that keeps your prompts on task. Draft at 480p, deliver at 1080p, and keep your source clips short.