Seedance – Image to Video (First + Last Frame)

First-frame-only video generation is a dice roll. You describe the motion, the model starts from your image, and the ending is whatever it felt like - sometimes great, sometimes a completely different shot. This node removes that bet. You give it a first_frame and a last_frame, and Seedance fills in the middle. The start and end are pinned; only the journey is up for grabs.

Same pack, same DNA: it's an API wrapper around ByteDance's closed Seedance 2.0 model (dreamina-seedance-2-0-260128) on the BytePlus ARK endpoint. No local weights, no GPU work, metered per call. The two images travel to the cloud as PNG data-URIs with explicit first_frame / last_frame roles, and the API interpolates motion between them.

Why this node is the chaining workhorse

The trick that makes it sing is last_frame. Every Seedance node in this pack outputs the final frame of the video it just made as an IMAGE tensor - and this node takes a last_frame as input. So the loop you'll actually build:

LoadImage → SeedanceI2VFirstFrame → last_frame ─┐
                                               ↓
                          SeedanceI2VFirstLastFrame (first_frame ← last_frame above)

You chain clips without a cut: each new shot starts exactly where the last one ended. The obvious extension is a seamless loop - feed the same image into both first_frame and last_frame and the model has to find its way back to the start. It's the closest thing this pack gives you to a closed loop out of the box.

What you actually set

• first_frame and last_frame - IMAGE tensors, both required. Connect anything that outputs an IMAGE: LoadImage, a VAE decode, the last_frame or images output of another node.
• prompt - "Smooth camera movement between the two frames." is the default, and it's hard to improve on for a basic transition. Get specific if you want actual choreography.
• ratio - adaptive by default, so the API fits the frames you gave it. If you force a ratio that doesn't match your images, you're asking for a crop.
• duration - 4–15 seconds for Seedance 2.0.
• resolution, generate_audio, watermark, return_last_frame - same story as the rest of the pack.

There are also first_frame_url / last_frame_url overrides if your frames are already hosted, plus optional reference_video / reference_audio sockets you reference as [Video 1] / [Audio 1] in the prompt.

The six outputs are the pack standard: video (native VIDEO), last_frame, frames (all frames as a batch), video_url, video_path, and last_frame_url.

Install

ComfyUI Manager, search comfyui-seedance-nodes. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/theclueless-ai/comfyui-seedance-nodes
pip install -r ComfyUI/custom_nodes/comfyui-seedance-nodes/requirements.txt

Light dependency list - requests, Pillow, numpy, opencv-python - and zero model downloads, because the model isn't yours to download. Set ARK_API_KEY as an environment variable or paste it into the api_key field, then restart ComfyUI. Key comes from the BytePlus console.

Where people get burned

• Treating it like a morph. It's interpolation in the generative sense, not a frame-to-frame morph - the model invents the in-between, and a prompt like "turn the vase into a flower" is a request, not a guarantee.
• The usual API stuff. Empty key → instant error; slow jobs (1080p, long duration) can outlast the 600s default max_wait, so bump it; and every call is metered, so chaining ten clips in one workflow is a real bill, not a hypothetical. Draft at 480p and low duration first.
• Frames that don't match. If your two images have wildly different aspect ratios or styles, the "fill" has more to reconcile - keeping them visually consistent gives the model less room to drift.