Concurrent Submit | Zhenzhen_vidu_start-end2video

Most image-to-video nodes animate from one image and let the model decide where things end up. Vidu's start-end mode is stricter and often more useful: you give it a start frame and an end frame, and it has to bridge the two. The shot lands where you told it to land. That's the tool you want for a punch-in that must end on a specific composition, or a camera move between two storyboard frames you already locked. This node is the concurrent wrapper around it - submit to the shared video pool (ten workers), get a task back, collect in Concurrent Collect Videos. If you're testing several ways to bridge the same two frames, this is exactly the batch pattern for it.

Inputs that matter

• start_image - required. The first frame.
• end_image - required. The last frame. These two are the whole point; Vidu interpolates the motion and physics between them.
• model - viduq2-pro (default), viduq2-turbo, viduq1, viduq1-classic, vidu2.0, vidu1.5.
• prompt - optional motion description to steer the middle.
• duration - 1 to 8 seconds, default 5. (Start-end clips cap at 8, a second shorter than plain img2video.)
• resolution - 360p, 540p, 720p (default), 1080p.
• movement_amplitude - auto/small/medium/large.
• is_rec, bgm, off_peak, watermark, wm_position - the usual Vidu extras.
• api_key, seed, skip_error - standard.

Output: a single COMFLY_VIDEO_FUTURE task.

How it works

Same machinery as every Submit node in the pack: ComflyConcurrent.py validates, pushes the original Comfly_vidu_start_end2video call onto the bounded video executor (ten workers), and returns the future immediately. The collector waits on every connected task, keeps slot order, and reports per-slot status. failure_mode: placeholder swaps a failed slot for a blank clip rather than killing the whole batch - handy when you're running several bridges and only keeping the good ones.

Setup

Install the pack once - ComfyUI Manager (search "Comfyui-zhenzhen"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

No model downloads, no VRAM. You need a paid Zhenzhen key from ai.t8star.org (overseas) or api.seedance.nz (domestic). Both frames get billed as inputs, so the cost is a little higher than a single-image i2v call.

Gotchas

The hard truth of start-end: the model honors the two endpoints, but the middle is where it can get creative in ways you didn't ask for - the bridge between a wide shot and a close-up can morph uncomfortably if the two frames aren't visually related. Matching lighting and camera angle between the frames is the single highest-leverage thing you can do; wildly mismatched frames produce warping, not cinema. And the usual pack caveats apply: timeouts aren't guaranteed failures (check your async task page), and the ten-slot video pool is shared across all of the pack's video nodes.