Image Batch Ends

Long videos out of image-to-video models are built one clip at a time: generate a clip, take its last frame, feed that frame back in as the seed for the next generation, and stitch. That's the "last-frame chaining" trick people use to push Wan and LTX past their native frame counts. The fiddly part is the bookkeeping - you need the closing frame out of one batch and you need to drop the duplicate frame where the two clips meet so the seam isn't a doubled still. Image Batch Ends is all of that bookkeeping in one node.

It's one of the batch-manipulation helpers in WAS Node Suite (WASasquatch's MIT pack, 457 nodes in v3). Where a node like Image Batch Repeat pads a clip to a length, this one peels the edges off a batch and hands you exactly the pieces a continuation workflow wants.

What comes out

Everything is a batch even when it's one frame, which keeps the socket types clean and means you can wire straight into another node that wants an IMAGE. The seven outputs are:

• first and last - frame 0 on its own, and the closing frame on its own. last is the one you feed to an I2V model to carry on from where the clip stopped.
• first_n and last_n - the opening and closing N frames, where N is the count input (default 1). Want the last eight frames as context? Set count to 8.
• without_first and without_last - the whole batch minus frame 0, or minus the final frame. This is the seam fix: a continuation model repeats the frame you seeded it with, so without_first on the returned clip drops that repeat before you join.
• batch_size - how many frames came in, numbered 0 to batch_size−1. Wire it to anything that needs the clip length told to it.

The count input caps out high enough that it doesn't matter - more than the batch holds just gives you the whole batch. An empty batch is refused.

The workflow this exists for

Say your first clip is 81 frames. You want 162. You take last from the batch, feed it to the I2V sampler as the start frame, and the model hands back an 81-frame continuation whose frame 0 is a copy of that seed. Stitch the two batches and the same frame appears at positions 81 and 82 - a visible hitch. So you join the first clip with the continuation's without_first, and the join is seamless.

From the other side, if your continuation model consumes a context tail, last_n gives you the frames to feed, and without_last is the mirror for trimming the incoming end. The node was clearly written by someone who'd done this dance by hand and was tired of slicing tensors with index nodes.

Batch gotchas

Two things trip people up. First, a batch of one frame: first and last are the same frame, and without_first/without_last have nothing to drop - they just answer that single frame. Nothing breaks, but don't read "without first = 0 frames" and build logic on it. Second, every output is a separate tensor, so if you wire several outputs and only change count, re-run everything downstream - there's no caching trick saving you.

Install is the standard WAS Node Suite one: ComfyUI Manager, search "WAS Node Suite v3", or clone https://github.com/WASasquatch/was-node-suite-comfyui into ComfyUI/custom_nodes and restart (ComfyUI 0.14.0+, Python 3.10+). v3 needs no pip step and downloads nothing - a real relief after v2, whose pinned dependencies were a running source of import failures when ComfyUI updated.

If this node looks like half of what you need, the other half is its sibling Image Batch Split, which cuts a batch at an arbitrary frame. Ends peels the ends; Split cuts anywhere. Between the two, the chaining toolkit is complete.