arkennemasis fal · Seedance 2.5 US Reference to Video · $0.26-1.40/s

The other three Seedance nodes generate video. This one takes video as input - up to ten clips, thirty pictures and ten audio files - and can either use them as guidance, modify one of them, or continue one where it left off. That's a video editor inside your graph, and it's the least understood node in the pack.

It's also the one where reading the schema pays for itself, because task changes the meaning of every other widget on the node.

How it works

The mechanism is fal's reference conditioning: your media is uploaded to fal's CDN, referenced by position in the prompt, and the model conditions on it while generating the new shot. Prompt-side, images are addressed as @Image1, @Image2, videos as @Video1, audio as @Audio1, and the numbering follows the sockets - each socket can carry a whole batch, and the first picture of image_1 is number one, then the rest of that batch, then image_2. Get the order wrong and the model confidently edits the wrong person.

The submission, polling, download and logging path is the same as every fal node in the pack: estimate, upload free, one billed call, result into output/fal/bytedance_seedance-2.5_us_reference-to-video/, request logged to output/fal/_requests.jsonl.

The inputs that matter

task is the one to set first:

• reference uses your media as guidance for a fresh shot.
• editing modifies a reference video. Per the tooltip it also coerces aspect_ratio and duration to auto - so if you set a duration and nothing seems to listen, that's why.
• extension continues a reference video and coerces aspect_ratio to auto to keep the framing continuous.

prompt is where the @Image1 and @Video1 references live; empty is allowed for pure edits. resolution and duration behave as on the other Seedance nodes, and seed defaults to -1 (the tooltip is upfront that the same seed can still drift slightly - treat it as best-effort, not a lock).

The reference sockets are optional individually but they're the whole point: image_1 through image_6, video_1 through video_3, audio_1 through audio_3. Their constraints are real and enforced - images up to 30 MB each, up to 30 of them; videos 1.8–30.2 seconds, ≤200 MB each, MP4/MOV, 300–6000 px per side, 24–60 fps, with combined video duration capped at 30.2 seconds; audio MP3/WAV, ≤15 MB each, 1.8–30.2 seconds. And across everything, 50 files maximum.

Outputs: video, video_path, seed, draft_id, info.

Money, because this node is different

Same per-second rates as the US image-to-video twin - roughly $0.26 / $0.57 / $1.40 per second at 480p / 720p / 1080p - but with a twist worth knowing: when video inputs are supplied, the price is multiplied by 0.6. At 720p with a reference clip that lands near $0.34/s. Handing it a clip makes a run cheaper, not more expensive.

Install

Manager → search comfyui-arkennemasis, or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart, then put FAL_KEY=... in the .env beside run_nvidia_gpu.bat. No key box on the node - that's deliberate so keys never travel inside shared workflows.

Common issues & troubleshooting

"More than 50 files." The total across images, videos and audio is capped. Six full image sockets each carrying a batch of ten is sixty before you've added a clip.

The edit came back the wrong shape or length. Expected under task: editing - duration and aspect ratio get coerced to auto. Extension does the same to aspect ratio only.

@Image2 isn't the picture you meant. Batching reorders the numbering. One item per socket while you're setting a shot up is the boring, reliable way to work.

A reference video is refused. Duration, size, fps and per-side resolution are all checked; a 4K phone clip or something under 1.8 seconds will be rejected, and the node stops before billing rather than paying for a failure.

The default cost cap trips. max_cost_usd is $20 out of the box, and a long 1080p clip clears that easily. 0 disables the cap.