Extracts base64 encoded latent and conditioning data from an image tensor's metadata.

NetDist_Plus can smuggle more than pixels inside a saved PNG - its SaveImageWithBase64-Nux node can bake the latent tensor and both conditioning channels from a render into that same file, base64-encoded alongside the image. This node is one of two ways to pull that payload back out. Where its sibling, ExtractBase64FromImageUpload-Nux, works off a file picker, this one takes an IMAGE that's already flowing through your graph - useful when the picture with the embedded payload is already wired into your pipeline and adding a separate upload step would just be extra clicking.

Inputs and outputs

One required input: image, typed as IMAGE - a live tensor socket, so it accepts whatever's already coming out of a Load Image node or an earlier point in your workflow, rather than a dropdown of files.

Two outputs: latent_base64 and conditioning_base64, the same pair its upload-based sibling produces. From here they typically feed LoadLatentFromBase64-Nux and ConditioningFromBase64-Nux respectively, to turn the strings back into a real LATENT and CONDITIONING.

The README doesn't spell out exactly how extraction works on a decoded in-graph tensor versus a file on disk - but functionally the two nodes are siblings with the same job, just different plumbing on the way in. If your image is already sitting in your graph as an IMAGE, use this one; if you need to browse for a file, use the Upload variant.

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 required for this node.

Where people get burned

Same core trap as its upload sibling: this only returns something useful on an image that had the payload embedded by SaveImageWithBase64-Nux in the first place. Feed it a plain image with no embedded latent/conditioning data and you'll get empty outputs rather than an error - that's expected behavior, not a bug.

Because this node takes a decoded tensor rather than reading a file straight off disk, keep an eye on what sits upstream of it in your graph. As long as the image traces back cleanly to the file SaveImageWithBase64-Nux produced, you should be fine - but if anything in between re-loads or re-encodes the image through a lossy path, you risk losing the very metadata this node is trying to read.

And the honest caveat that applies across this whole corner of the pack: the author's own README says he never got a proper chance to test the remote-latent round trip. Conditioning is the well-worn path (it's the pack's actual demonstrated use case - running text encoding on a second GPU); latent extraction is the part to treat with a bit more skepticism if the numbers coming out look off.