Prompt to Conditioning

Most of ComfyUI's built-in text encoders make you type the prompt directly into a text box on the node. That's fine until the same prompt needs to feed three different places, or you want to build it up from other nodes first. LC Prompt to Conditioning is the fix: it takes a string from a socket and encodes it into CONDITIONING with whatever CLIP you hand it. No text box, no retyping - the prompt lives somewhere else, and this node is just the encoder.

The name is the whole job. It's the "Convert widget to input" idea from the plumbing layer applied to text encoding: break the value out of a widget so the graph has one authoritative source for it. You wire a string node, or the pack's own Positive/Negative prompt boxes, into the string socket, drop a CLIP in from your checkpoint loader, and out pops conditioning that plugs straight into a sampler.

How it works

Under the hood it's the same tokenize-then-encode path as core CLIP Text Encode. The one structural quirk worth knowing: string is socket-only (forceInput), which is exactly what keeps you honest about where the prompt comes from. And if the socket arrives empty, it encodes an empty string rather than erroring - so an unwired prompt quietly gives you blank guidance instead of a loud failure. Wire it up, don't leave it dangling.

The inputs that matter

• clip - the CLIP model from a checkpoint or CLIP loader. Every architecture's conditioning comes from here, so if your CFG feels wrong, it's usually the scale, not this node.
• string - the prompt text, wired from an LC Positive/Negative box, an LC Show Text, or any STRING source.

Output is a single conditioning socket. That's it. Feed it to a KSampler, or into conditioning-manipulation nodes if you're doing regional or dual-pass work.

Install

It ships in the LC123 pack, which you get from ComfyUI Manager (search "LC123") or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC123_nodes

Then restart ComfyUI. The pack needs nothing extra - just ComfyUI's own torch and numpy - and it's MIT licensed. On a cloud-hosted ComfyUI (RunningHub and similar), LC123 isn't preinstalled, so you'll get "node not found" errors on shared workflows until you request the pack or run local.

Gotchas

Because this node has no widget, it's easy to forget the string is wired from upstream - if you edit a prompt in the node you think is here but is actually a text box, nothing changes. If your prompt is going nowhere, follow the wire back to the actual source node. And remember: with the modern guidance-distilled models, conditioning comes from the CLIP but the steering comes from CFG - set that on the sampler side, not here.