Load Conditioning (Mago)

MagoLoadConditioning is the other half of the conditioning cache that ships in ComfyUI-MagoNodes. Where Save Conditioning (Mago) bakes a CONDITIONING to disk, this node pulls it back. Alone it does almost nothing; wired after a save it's the thing that lets a production LTX workflow run with the text encoder never loading into VRAM.

The pitch is simple: if your prompt is fixed - the README's example is an LTX HDR-conversion prompt that never changes - you encode it once, cache it, then run your graph forever with CLIPTextEncode deleted. For LTX-2 and 2.3 that means not loading a Gemma 3 12B text encoder at all, which was the single biggest VRAM hog (and OOM source) on those models' launch day. Only the diffusion model and VAE sit in memory.

How it works

There's one input, filename, and it's a dropdown built from whatever .pth files live in output/conditioning/. Pick a cache, get a CONDITIONING out the other end, ready to wire into LTXVConditioning or straight into a guider:

Load Conditioning (hdr_pos) ┐
Load Conditioning (hdr_neg) ┴→ LTXVConditioning → CFGGuider → SamplerCustomAdvanced

A couple of honest caveats before you run this in production. First, the dropdown is populated when ComfyUI builds the node list - save something new and you need to refresh the node list (or re-add the node) before it shows up. Second, and this is the one to actually read: the node loads with torch.load(..., weights_only=False) because conditioning dicts carry non-tensor metadata, and the author is explicit that you should only load your own cache files. weights_only=False is a footgun if you ever point it at a stranger's file. Don't.

If you run it without any cache, the dropdown shows a placeholder, <no cached conditioning - run Save first>, and loading it raises a clean FileNotFoundError - that's the author telling you to do the bake pass first, not a crash. The node also hashes on the cache file's mtime, so if you re-save a conditioning under the same name, Comfy re-executes downstream instead of serving stale output.

Installing

Part of the MagoStudio/ComfyUI-MagoNodes pack - via ComfyUI Manager (search "ComfyUI-MagoNodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/MagoStudio/ComfyUI-MagoNodes

Restart ComfyUI. No extra dependencies, no model downloads. This node and its save sibling are the only things in the pack under Mago Nodes/Conditioning, and they're about as low-risk as custom nodes get.