Faishme Repeat Latent Batch

Faishme Repeat Latent Batch is the latent-space twin of the pack's Repeat Image Batch: it takes a LATENT batch and repeats each sample a set number of times, in place, so a batch of [A, B] with amount 2 becomes [A, A, B, B]. Where it differs is where you have to look for it. Thanks to a typo in the pack's source, this node declares its category as FaishmeDebug - so in ComfyUI's right-click menu it does not sit with its siblings under FaishmeNodes. It hides under FaishmeDebug instead. If you installed the pack and can't find this node where you expected it, that's why.

How it works

Latents in ComfyUI are dictionaries - the samples tensor plus optional extras like noise_mask and batch_index - so repeating a latent batch means repeating every key, not just the tensor. This node does that carefully:

• samples gets the same row-wise repeat as the image version ([A, A, B, B] ordering).
• noise_mask, if present and batched, is repeated to match, so masks stay aligned with their samples.
• batch_index, if present, is extended so each duplicated sample gets a fresh, correct index. That's the detail that keeps downstream nodes from getting confused about which latent belongs to which position.

That care is the whole reason to use this over a generic tensor repeat: it keeps the latent dict internally consistent, which matters once noise_mask and batch_index are in play.

Inputs and outputs

• samples - the LATENT batch to expand.
• amount - copies per sample, 1 to 64 (default 1). Note the cap is 64 here, versus 4096 on the image version - the latent space is where VRAM pressure bites first, so the ceiling is tighter.

Output is a single LATENT batch, amount× the input size.

Why you'd use it

Same motivation as the image version, one stage earlier in the pipeline. If you're generating variations from pre-encoded latents - repeat each latent N times, then run each copy through a sampler with different conditioning - doing it in latent space avoids decode/encode round-trips. In a batch fashion workflow, that's exactly the sort of thing that keeps a long generation run from doubling in wall-clock time.

Installing it

The pack-level install - ComfyUI Manager, search "ComfyUI_faishme", or:

cd ComfyUI/custom_nodes
git clone https://github.com/AkashKarnatak/ComfyUI_faishme

Restart ComfyUI, then look under FaishmeDebug in the menu, not FaishmeNodes.

Where people get burned

The menu location is the #1 confusion - expect it to take a second to find. After that, remember the 64-copy cap and the same VRAM warning as its image sibling: multiplying latents multiplies what the sampler has to hold. And if your latent dict carries a noise_mask, it does get repeated, so a mask designed for inpainting scales correctly - but it scales, which means an unbatched mask can behave differently than you expect. Read the amount, check the menu, and it's a solid batch-shaping utility.