MakeBatchFromFloatList

Take MakeBatchFromIntList, swap every integer for a float, and you have this node. It exists for exactly the same reason: the InfiniteTalkMultiImage node returns per-segment values as lists - including real_start_time_list, the actual start times the schedule settled on, which are floats - and the workflow's for-loop needs a single batch object it can pull one item out of per iteration. MakeBatchFromFloatList is the "collect them all into a batch" half of that pattern.

Mechanically it's a two-liner: an float_list input that accepts multiple connected FLOAT values (it's declared INPUT_IS_LIST, so ComfyUI gathers them into a Python list), and a single float_batch output. One value in, one value out, no conversion. Multiple values in, and they're packed into a torch.float32 tensor. That tensor is what the rest of the workflow indexes.

In the pack's workflow you'll see it feeding the real_start_time_list output from the scheduling node into the loop, where GetFloatByIndex pulls batch[index] to tell the current segment which start time it belongs to. Build with this node, read with GetFloatByIndex - they're a matched pair.

Inputs: float_list (FLOAT, multiple). Output: float_batch (FLOAT). Zero knobs. Category InfiniteTalk, but like its integer sibling it's a generic utility - any workflow that needs to collapse a fan-out of floats into one batch gets the same benefit.

Installation is the pack's standard story:

cd ComfyUI/custom_nodes
git clone https://github.com/xuhongming251/ComfyUI-InfiniteTalk-MultiImage

or ComfyUI Manager → search ComfyUI-InfiniteTalk-MultiImage → Install → restart. The pack has no Python dependencies of its own and downloads no models.

There's genuinely nothing that goes wrong with this node on its own. The one thing to keep straight is that the tensor it produces has a fixed length set at build time - if your upstream scheduling node returns a different number of values than you expected, the batch silently has that many entries, and the index node downstream will clamp instead of complaining. When your start times look slightly off in the final video, count the entries in the batch first.