arkennemasis fal · Seedance 2.5 Reference to Video · $0.22-1.16/s

Character consistency is the hard problem in generated video, and it's the reason this node exists next to the plain image-to-video one. Instead of one starting frame, you feed Seedance a library: reference images it refers to as @Image1, @Image2, reference videos as @Video1, reference audio as @Audio1. The model locks the character, the set and the palette, and the prompt tells it what to do with them.

ArkFal_bytedance_seedance_2_5_reference_to_video is the multimodal endpoint of Seedance 2.5, still API-only like the rest of ByteDance's flagship line - its open releases are accelerators and adapters, never the generator. Pricing is per second of output: $0.22 at 480p, $0.47 at 720p, $1.16 at 1080p, with a discount when video references are involved.

What you set

• task - the important one, default reference. reference uses your media as guidance. editing modifies a reference video and automatically coerces aspect_ratio and duration to auto. extension continues a reference video and coerces aspect_ratio to auto. That coercion is silent - if you set a duration and then wonder why it was ignored, this is why. Different jobs, same node.
• image_1 … image_6 - up to 30 images total, each under 30MB, JPG/PNG/WebP/BMP/TIFF/GIF/HEIC. Reference them in the prompt as @Image1, @Image2. Each socket can carry a batch, so six sockets is more room than it looks.
• video_1 … video_3 - up to 10 reference videos, MP4/MOV, each 1.8–30.2 seconds and under 200MB, with a combined duration under 30.2 seconds. Referred to as @Video1.
• audio_1 … audio_3 - up to 10 files, MP3/WAV, same 1.8–30.2 second window, 15MB each. @Audio1.
• Total files across all three modalities: 50, which is the real ceiling to plan against.
• prompt - where the references get used. Names plus actions; the model is doing identity matching, not image captioning.
• resolution, draft, duration (auto or 4–30s), aspect_ratio (auto, 16:9, 9:16, 1:1, 21:9), generate_audio (on, free), bitrate_mode, codec, seed (-1 for the model's choice; results can still vary slightly at a fixed seed).

Plus max_cost_usd, reuse_identical_run, extra_json, max_concurrent.

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

The pricing quirk you should understand

With video references in play you're billed for both the input and the output video, at a 0.6 multiplier - fal's own example is roughly $0.2838 per second at 720p when a video input is present. So a 10-second reference video that produces a 10-second output is twenty seconds of billing, discounted. Audio and image references don't add like that, but they do count as inputs to the shot.

In practice: build the character from images and use video references only when the motion itself is the thing being transferred. That's cheaper and usually gives you cleaner identity.

Install

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

Restart ComfyUI, or let ComfyUI Manager install it - search arkennemasis. Then:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

in the .env beside run_nvidia_gpu.bat, or in the ComfyUI folder. There is no key widget on the node, on purpose: workflow JSON stores widget values verbatim, so a typed key travels with every file you share.

Where people get burned

Thirty images and ten videos is an easy way to spend upload time and money on references the model dilutes rather than uses. Curate: three or four clean angles of the character beat twenty random frames. The mixed-modality ceiling (50 files) is a hard wall, not a suggestion.

And the task switch. If you're planning to edit a reference video, know that it wants to control duration and aspect ratio itself - you're not keeping a timeline here.

Everything else is the pack's standard fal machinery: results to output/fal/bytedance_seedance-2.5_reference-to-video/, a request log in output/fal/_requests.jsonl, automatic backoff on rate limits, Cancel that reaches fal, and free recovery of an interrupted job by request id.