Save Latent MXD

Long WAN 2.2 image-to-video runs are brutal on GPU time. The workflow Maxed-Out designed around this node splits the job in two: Save Latent I2V MXD stores your high-noise latent mid-generation, so you can come back later and finish the low-noise pass - refining, upscaling, or re-sampling without re-running the expensive first half. The tooltip says it plainly: this saves the "high-noise latent ... for later low-noise finishing."

It's the companion to the LoadLatent_I2V_MXD / LoadLatents_FromFolder_I2V_MXD nodes in the same pack, and together they form the two-stage "latent handoff" pattern that's become standard in serious WAN 2.2 workflows.

What it saves and how

You give it four things:

• samples - the latent tensor (the one with most of the noise still in it)
• positive and negative - the WAN conditioning from your image→video encode
• filename_prefix - base name for the saved files (default I2V)

The node writes a .latent safetensors file that embeds the current workflow graph and sampler metadata, plus a .cond.pt sidecar holding the conditioning. That's the clever part: when the loaders in this pack read the file back, they reconstruct not just the latent but also the steps, CFG, sampler, scheduler and conditioning that were in play when you saved. You pick up exactly where you left off instead of hand-restoring settings.

It's an output node with no outputs - it exists to write files. Save output goes to ComfyUI/input/latents (that's where the loader looks).

Where it fits

The typical arc: run your WAN 2.2 I2V first stage, hit this node before the final denoise steps, queue, then load the latent in a second workflow tuned for finishing - maybe at higher resolution or with a different upscale pass. Because the workflow is embedded in the file, loading the latent and dragging the resulting video back into ComfyUI can even bring both stages onto one canvas at once.

Gotchas

• Keep the .latent and its .cond.pt sidecar together. The loader reads both; orphan one and you'll lose your conditioning and settings.
• This node is built around WAN 2.2's latent format. It'll happily save anything you hand it, but it's meant for the I2V/VACE 2.2 pipeline the pack ships.

Install

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

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

No requirements.txt, no models to download - the pack is all Python logic. If you see the node in the menu under MXD/Latents (I2V), you're good to go.