MD: Load Conditioning 📂

Load Conditioning is the reading half of the pack's conditioning cache. Its sibling, MD_SaveConditioning, dumps a CONDITIONING stream to a .bin file in ComfyUI/models/conditionings/; this node reads those files back and hands you a ready-to-use CONDITIONING stream, no text encoder involved. If you've saved caches, this is how you use them.

On its own it's almost nothing - one dropdown, one debug toggle, one output. Its whole purpose in life is to skip re-encoding, which is why it exists as a pair with the Save node and why it shines in workflows where the same prompts get reused across many runs.

The inputs

• conditioning_file - a dropdown listing every .bin file in the conditionings folder (the source registers with ComfyUI's get_filename_list). Anything MD_SaveConditioning wrote shows up here automatically. It's populated at load time, so if you saved a new cache and it's not in the list, refresh/restart the UI.
• debug_mode - Silent or Info, for console confirmation of what loaded.

Output: a single conditioning stream, ready to wire into any node that takes CONDITIONING - your KSampler's positive or negative, or an ACE-Step graph.

The one rule that makes or breaks it

Conditioning is architecture-specific, full stop. The file contains embeddings produced by your model's text encoder, and they are not portable. Loading an SD1.5-cached negative into an SDXL or ACE-Step workflow will give you wrong results, not a friendly error - the tooltip's warning ("Ensure the model architecture matches the cache") is the entire troubleshooting section. This is the same reason the Save node's tips tell you to name files like acestep_neg; the name is the contract.

Why you'd reach for it

Three scenarios. First, batch A/B testing where the negative is constant: encode once, load a thousand times. Second, audio workflows where the text encoder is expensive and the prompt is long - the exact case this pack was built around. Third, any workflow you want to make reproducible without re-deriving the encoding: a .bin is a static artifact, so a cached workflow runs identically.

The honest caveat: the cache file itself is a black box - no inspection, no metadata beyond the filename, no way to know what model made it after the fact. That's fine if you name carefully and remember that Save/Load are a matched pair.

Installing

It ships in MD Nodes, so you install the whole pack:

cd path/to/ComfyUI/custom_nodes
git clone https://github.com/MDMAchine/ComfyUI_MD_Nodes.git
cd ComfyUI_MD_Nodes && pip install -r requirements.txt

Or via ComfyUI Manager - search MD Nodes - then restart. The pack's requirements are broad (audio libs, matplotlib, client SDKs), so you pay the full install even for a two-node cache utility.

Bottom line: this node is 5% of the story and the Save node is the other 95%. If you're caching, use them as a pair, label by model architecture, and enjoy never re-encoding that static negative again.