Latent Passer

Latent Passer is the same idea as Eclipse's other passers, but for the heavy one: a latent tensor. Take a latent in, hand the identical latent out, no conversion, no processing. Its only job is to let you route a latent across a crowded canvas cleanly - because in a real workflow, the latent from your sampler often has to travel a long way, and a wire that crosses half the graph is a wire that gets damaged when you rearrange nodes.

The mechanism

One latent input, one latent output, straight pass-through. The type here is ComfyUI's match-type (COMFY_MATCHTYPE_V3), which means it accepts latent-shaped data as ComfyUI defines it - the dictionary of latent tensors plus metadata that samplers and VAE nodes exchange. Because it's typed, ComfyUI will refuse to connect something that isn't latent-shaped, which is the safety you want: you can't accidentally route a conditioning through it and wonder why everything breaks later.

The value of a passer is organizational, not computational. ComfyUI caches aggressively and passers are pure identity functions, so a well-placed passer costs you nothing at runtime - and it saves you from the classic failure where a long wire gets silently rerouted when you move a node. Place one right next to your KSampler or VAE Decode, run the connection in from the far side, and the visible wire stays short. In the jargon of the plumbing layer, it's a legibility node: it exists so a graph too large to read stays one you can still work in.

Where it fits

Latents are the most expensive thing to carry in a ComfyUI graph, and the most common place you need this is in upscale-and-refine or multi-pass workflows where a latent from one pass feeds a node several groups away - or in pipe-based graphs where you occasionally want to tap a latent out of a pipe and run it to a specific consumer. The passer lets you place that tap exactly where it's readable. If your workflow is small, you don't need it; if it sprawls, you'll want a few.

Inputs and output

• latent - the latent to pass through, verbatim.
• latent (output) - the same latent, unchanged.

Install

Part of ComfyUI_Eclipse, so the pack install gets it:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse

Or install ComfyUI_Eclipse from ComfyUI Manager and restart. No models, no dependencies. It's the same "wire with a label" trick as the Int Passer, just for the type that carries your actual image data in compressed form - when the graph gets big, you'll be glad it's there.