CLIP Text Encode (Lumina 2)

Lumina 2 is one of the models that moved prompt encoding from "tag soup through CLIP" to "an actual conversation with a language model," and this node is where that shows up in the UI. Instead of one text box you get a system-prompt dropdown and a user prompt. It's the closest thing ComfyUI has to a chat-format text encoder, and once you understand why the dropdown exists, the node stops being weird and becomes the obvious way to talk to the model.

How it works

Under the hood the node does something deceptively simple: it takes your chosen system prompt, glues on a <Prompt Start> marker, then appends your user prompt and feeds the whole string through Lumina 2's text encoder. That's it. The system_prompt combo box isn't offering you free text - it's a choice between two officially tuned presets, and the wording is a good summary of what each one buys you:

• Superior - "generate superior images with the superior degree of image-text alignment." The general-purpose preset, and the one most people will leave on.
• Alignment - "generate high-quality images with the highest degree of image-text alignment." The stricter follower, tuned for cases where sticking exactly to what you wrote matters more than creative interpretation.

So the dropdown is really a style-of-obedience dial: pick Alignment when you need literal adherence, Superior for everything else. The user_prompt is your actual request, and clip is the Lumina 2 text encoder loaded via a CLIPLoader with the lumina2 type.

The output and the model files

The single CONDITIONING output carries the encoded prompt into your sampler's positive slot - the standard setup, no negative slot on the built-in path. The node ships with ComfyUI core (no install), and the real download is the Lumina 2 stack: the checkpoint plus its text encoder and VAE in the standard models/ folders. The ComfyUI example workflows for Lumina 2 show the exact files.

Where people get confused

Two things usually trip people up. First, because the prompt format is LLM-style, you should write it that way - full sentences, natural instructions, not comma-separated tag spam. The encoder reads your text like a chat turn, and prompt-engineering habits from SDXL actively hurt here. Second, the system prompt is a choice, not a slot to paste your own instructions into - people instinctively try to type in the combo box, and it doesn't take free text. If you want to experiment with your own system behavior, you'd write the system-style instruction into the user prompt itself rather than the dropdown. And if your text encoder isn't loading, check that the CLIPLoader type is set to lumina2 - grabbing a wrong-type encoder is the usual culprit and it fails in confusing ways.