save conds and latents

This is the node that goes at the end of whatever's running on your remote GPU. It works like the standard Save Image node - take an image, write it to disk with a filename prefix - but it can also bake the render's latent and both conditioning channels straight into that same file, base64-encoded, riding alongside the pixels. No separate serialization step required: wire the raw tensors in, and this node handles turning them into something the primary instance can pull back out later.

Why it exists

The pack's core trick is running text encoding (T5 + CLIP-L) on a second GPU while your primary card handles UNet and VAE. For that split to actually save you anything, the conditioning produced remotely needs to make it back to the machine doing the sampling - and this is the node that packages it up on the remote end. Save the render here, with the conditioning (and optionally the latent) wired in, and your primary instance can retrieve all of it in one shot: automatically via FetchRemoteWithExtras-Nux's outputs, or manually later via ExtractBase64FromImage-Nux / ExtractBase64FromImageUpload-Nux if you're working from the saved file directly.

Inputs and outputs

Required: images (the IMAGE to save) and filename_prefix (default "ComfyUI", same convention as core ComfyUI).

Optional: workflowName (a plain label for your own bookkeeping), latent (wire in your KSampler's output if you want the latent embedded too), and positive_conditioning / negative_conditioning (wire in CONDITIONING directly - this node does the base64 encoding for you, no need to run it through ConditioningToBase64-Nux first).

No outputs - this is a terminal node (is_output_node: true), same as core Save Image. Its only job is writing the file.

Installing it

ComfyUI Manager: search ComfyUI_NetDist_Plus. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/nux1111/ComfyUI_NetDist_Plus
pip install requests

Restart ComfyUI. No model downloads for this node.

Where people get burned

The optional sockets are genuinely optional - leave latent, positive_conditioning, and negative_conditioning unwired and you just get a normal saved image with no extra payload. That's fine if you only care about the pixels; don't expect FetchRemoteWithExtras-Nux or the extract nodes to hand back data you never wired in here in the first place.

Since it's a terminal node with no output socket, nothing downstream in this graph reads from it - the receiving side of the round trip happens on the other ComfyUI instance, through FetchRemoteWithExtras-Nux or the manual extract nodes reading the file this one produced.

Worth flagging the same caveat that runs through this whole corner of the pack: the README's own Issues list says the author never got a chance to fully test the remote-latent path. The conditioning sockets are the well-worn, demonstrated route - it's literally the pack's headline example workflow. The latent socket is the newer, less-verified addition; if you're wiring it in, treat any weirdness on the receiving end as unexplored territory rather than something you did wrong.