Load Latent Pipe MXD

If you do WAN 2.2 image-to-video work, you've met the round-trip problem. You generate a latent, then later you want to load it back up - with its conditioning, its sampler settings, everything that made it - and run it again, or branch off from it. Normally that means remembering or re-deriving every parameter. Load Latent Pipe MXD is the lazy version: it reads a .latent file plus its .cond.pt sidecar and hands you the whole run as a single MXD_LATENT_PIPE, ready to unpack and feed a sampler.

How it works

This is the pipe-bundling wrapper around the pack's plain Load Latent MXD. It looks in your ComfyUI/input/latents folder for .latent files (with subfolders allowed), and for the .cond.pt sidecar sitting next to the chosen file. That sidecar is where the magic lives: when the Maxed Out save node wrote the latent, it embedded the positive and negative conditioning, steps, cfg, sampler, scheduler, end-at-step, and the model shift. Load the latent and every one of those comes back with it. The single output is an MXD_LATENT_PIPE - the same wire the batch loader produces, so swapping between "load one" and "load a folder" is a one-connection change. To get the values back out as individual sockets, feed the pipe into Unpack Latent Pipe MXD.

Inputs that matter

• latent - dropdown of .latent files under input/latents, grouped by folder so the prev/next arrows walk one folder at a time.
• run_folder - flip this on and hitting Queue Prompt auto-queues every latent in the file's folder, one after another. That's the "process my whole finished batch" button.
• refresh_before_run - when run_folder is on, re-scan the folder for new files right before the queuing loop starts, instead of trusting the dropdown as of whenever you last loaded the workflow. Use this when another workflow is still writing latents into the same folder as you hit queue - the classic "two workflows chasing each other" pipeline.

Wiring it in

The pipe output goes straight into Unpack Latent Pipe MXD, which hands you samples, positive, negative, shift and the sampler settings to feed a KSampler (or WAN KSampler). It's built for the WAN 2.2 I2V latent workflow this pack wraps, but the mechanism is generic - any .latent written by the pack's save nodes loads cleanly.

Installing

ComfyUI Manager → search "Maxed Out" → install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Maxed-Out-99/ComfyUI-MaxedOut.git

No pip dependencies or model files to fetch.

Common issues

The classic gotcha: a latent written by another pack or a bare core Save Latent won't have the .cond.pt sidecar, so the conditioning and settings come back as empty or defaults. This node is a round-trip tool for latents this pack saved - for foreign latents, use the plain latent loaders and set your own parameters. And if your dropdown looks stale after another process wrote new files, that's exactly what refresh_before_run is for.