Smart Overlapping Batcher (Warper)

Chunking a long video into sampler-sized batches solves your VRAM problem but creates a worse one: the seams. Process a 600-frame clip as independent 81-frame chunks and the boundaries between chunks shimmer, jump, and generally betray that nothing remembers the frames across the cut. Smart Overlapping Batcher (Warper) attacks exactly that - instead of clean cuts, each batch overlaps its neighbor by a number of frames you choose, so the model always sees continuity across the boundary. The README calls it out plainly: "to help maintain temporal consistency."

It's the fancier sibling of Smart Video Batcher (Warper) in the ComfyUI Warper Nodes pack from workflow author AIWarper. If your chunks are independent enough that seams don't matter, use the plain batcher. If you're doing video inpainting, frame interpolation, or anything where temporal coherence is the whole game, this is the one you want.

How it works

Three inputs:

• image - your full frame stack.
• batch_size - target chunk size (default 81, minimum 2).
• overlap_size - how many frames each chunk shares with its neighbor (default 16).

Mechanically it's a sliding window: the first chunk is frames 0–batch_size, then it advances by step = batch_size - overlap_size, and so on. Two design choices are worth knowing. First, if a window would run past the end, the node grabs the last batch_size frames as the final chunk instead of emitting a runt - so every chunk is full-size and the final chunk naturally overlaps the previous one. Second, overlap_size must be smaller than batch_size; if it isn't, you get a warning and the original batch back unchanged. Output is image_batches - the pack's IMAGE_BATCHES list type, which feeds Get Batch By Index (Warper) to pull chunks out one at a time.

Installing it

Standard Warper install: ComfyUI Manager → search "ComfyUI-WarperNodes" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/AIWarper/ComfyUI-WarperNodes
# restart ComfyUI

Torch-only, no extra deps, no model downloads. The README's RAFT checkpoint is for Flow Visualizer only.

Common issues

The overlap math is where people misjudge. Your overlap_size is frames shared, so with batch 81 and overlap 16 you're feeding 65 new frames per chunk - that means more chunks than you'd expect, and each run is slightly longer. Budget for it. The "final chunk = last batch_size frames" behavior means the last chunk's frames were already seen in the previous chunk, so outputting every chunk and stitching them naively will duplicate content at the tail; you typically only keep each chunk's new region when reassembling. And if you set overlap_size at or above batch_size, the guard returns the whole clip as one batch - which silently skips the chunking you asked for. Watch the console warnings; they're the node telling you it bailed.