BytePlus Seedance 2.5 First-Last-Frame to Video

Text-to-video is a slot machine. Image-to-video is a slot machine that starts where you told it. First-last-frame is the one that lets you direct: here's frame one, here's frame thirty, fill in the motion between them. For anything with a choreographed move - a product turning, a character walking from A to B, a door closing - it's the difference between praying and planning.

This node is the 2.5/2.0 family version, which means the same aggressive pricing and the same native audio generation as its text-to-video sibling, plus both ends of your shot. It also has an escape hatch the 1.0 nodes don't: instead of feeding it images, you can name a BytePlus asset for either frame.

How it works

Under the model picker you get prompt, resolution, duration (4–30 s on 2.5, 4–15 on 2.0), generate_audio and - on 2.5 - output_format. Then, outside the picker, first_frame and last_frame image inputs, first_frame_asset_id and last_frame_asset_id string inputs, seed, watermark, generation_count and non_blocking.

The frame handling is where this node earns its keep. Seedance 2.5 takes the first frame's own aspect ratio and asks for ratio="adaptive", so the frames aren't resized behind your back. On the 2.0 models the pack pre-sizes your local frames to the exact output pixel dimensions before submitting - that's a deliberate fix for a 1080p stretch the core nodes had, and it means you don't have to resize anything yourself.

The two frame sources are mutually exclusive per end: give it a first_frame image or a first_frame_asset_id, not both. And you must give it one of them - a first frame is required, a last frame is optional. If you animate from an asset, first_frame_asset_id also accepts a bare asset://<id> URI or a public https:// image link.

Outputs

VIDEO is a list - with generation_count above 1 you get every generation, each from seed + N, and the next node runs once per video. last_frame hands you the final frame of each video as one image batch, which is exactly what you want for chaining shots: the output of one generation becomes the first_frame of the next. response is the task JSON, and draft_task_id is there when you're running a Draft model.

Wire VIDEO into a Save Video node, or into an upscaler, or into Concatenate Video if you're stitching a sequence.

Install and key

cd ComfyUI/custom_nodes
git clone https://github.com/byteplus-sa/ComfyUI-BytePlus-ModelArk
pip install -r ComfyUI-BytePlus-ModelArk/requirements.txt

Restart (ComfyUI 0.31.0+), or install through Manager by searching BytePlus ModelArk. Save the ModelArk key and region under Settings → BytePlus, or as BYTEPLUS_API_KEY / BYTEPLUS_REGION in user/.env, and enable the model in the ModelArk console. Connected frame images are sent inline - no Comfy.org login needed here, which is why this node is a decent first video test even if you never touch the reference-video features.

Where people get burned

Frame mismatch is the classic. If your two frames are different sizes or wildly different aspect ratios you'll get either an error or a video that lurches at the seam. Match them before you start; the node handles sizing for the 2.0 models but it can't invent a coherent transition between a 16:9 portrait and a 9:16 landscape.

The second trap is believing the seed. It doesn't make results reproducible - it only makes the node re-run. If you liked a take, keep the video, because the same seed will hand you something new.

And the third is the one nobody warns you about until it's billed: generation_count is parallel billed generations, and on video that's the most expensive dial in the pack. Combined with audio generation on by default, a "quick test" at 30 seconds and three variations is a genuinely expensive mistake. Use Draft mode for the exploring, this for the take.