Lantent Compose Muti

"Lantent Compose Muti" - yes, that's the actual display name, and no, we don't know what the author had against the letters a, t, and l. Typo aside, this is the pack's cleverest node, because it's the one built to consume the pack's own strangest output.

Where LatentCompose blends exactly two latents, LatentComposeMuti treats a batch of latents as a sequence and walks through it with a single ratio. The latents input is a LATENT tensor with more than one item in the batch dimension - how many is up to you. ratio is a position along that sequence: 0.0 gives you the first latent, 1.0 gives you the last, and anything in between smoothly slerps between whichever two neighbors bracket that point. It's a scrub bar for a stack of latents.

Why it exists: the UnsamplerCustom pairing

This is the piece the README's "use with unsample nodes" is pointing at. UnsamplerCustom in this same pack returns the entire unsampling trajectory - every intermediate latent, from near-clean all the way to full noise - stacked into one batch. That batch is exactly what LatentComposeMuti expects. Feed it the ladder, set ratio, and you get a latent at any noise depth you want, without re-running anything.

Want to resample your image from a gentle 20% of the way back to noise? ratio ≈ 0.2. Need it right up against pure noise for a heavy re-imagination? Crank it toward 1. Because the batch order from UnsamplerCustom is fixed (index 0 = closest to your original, last = noisiest), the ratio becomes a clean, predictable dial instead of a guess.

The mechanism itself is simple under the hood: it finds the position s = (batch_size - 1) * ratio, grabs the two neighboring latents, and slerps between them with the local fraction. Same spherical-interpolation math as the pack's other nodes - constant magnitude, no blurry lerp valley.

Inputs and output

• latents - a LATENT with a batch of 2 or more
• ratio - position along the sequence, 0 to 1, default 0.55

One LATENT out, ready for a KSampler or a VAE Decode. You can build the batch yourself too - stack a couple of latents with something like LatentBatch - as long as you keep the order meaningful.

Installing it

Part of ComfyUI-LantentCompose: ComfyUI Manager → search "LantentCompose" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Apache0ne/ComfyUI-LantentCompose

Then restart. No dependencies, no downloads - the README's "None" is the whole story.

Common issues

• Batch order is everything. The node assumes your batch is an ordered sequence and interpolates between neighbors. Scramble the order and your scrub path zigzags through unrelated latents. With UnsamplerCustom's cleanest→noisiest ordering this is a non-issue; with hand-built batches, check what you stacked.
• A batch of one is a silent no-op. If B == 1, the code returns the input unchanged - no error, no interpolation. If your "scrub" isn't doing anything, check you actually fed it a stack.
• Same shape and same latent space rules apply, as with the rest of the pack: mixing latents from different models' VAEs blends meaningfully only if they share a latent space.

It's a thin node - one input, one ratio - but it's the missing dial that turns an unsampler's raw output into a tool you can actually steer.