Text encoders: CLIP, T5 and others

The part of a model that reads the prompt, and why some models have two or three.

#In plain words

The drawing model does not read letters. A text encoder turns your prompt into numbers first, and the drawing model follows those numbers. In ComfyUI the text encoder travels along links of type CLIP, whatever kind it really is.

There are three generations of text encoders:

• CLIP (the text half of OpenAI's CLIP, in sizes clip_l and clip_g). It was trained on short image captions, reads up to 75 tokens at a time and understands words and short phrases better than grammar. SD 1.5 uses clip_l; SDXL uses both clip_l and clip_g.
• T5 (Google's T5-XXL, and umt5_xxl, its many-language version). A language model that reads whole sentences, a few hundred tokens long. Flux.1 and SD3 use it next to CLIP; Wan uses umt5_xxl alone; Chroma uses T5 without CLIP.
• Large language models: Qwen2.5-VL for Qwen-Image, Qwen3 4B for Z-Image, a Mistral model for Flux.2 dev, Llama-based encoders for HunyuanVideo. They read long, detailed prompts and follow who does what and where.

Why two? A model trained with two encoders was taught to use both views of the prompt at once: CLIP for style and look, T5 for meaning and layout. Both must be loaded. SD3 even uses three (clip_l, clip_g, t5xxl).

#Where you meet it in a graph

• Load Checkpoint gives CLIP for SD 1.5 and SDXL.
• Load CLIP (CLIPLoader) loads one encoder; DualCLIPLoader loads two; TripleCLIPLoader loads three. Each has a type field naming the family.
• CLIP Text Encode (Prompt) takes CLIP and your text and gives CONDITIONING. Some families have their own text nodes (for example CLIPTextEncodeFlux with separate fields for CLIP and T5, or TextEncodeQwenImageEdit for editing).
• CLIP Set Last Layer makes CLIP stop one or two layers early ("clip skip"), which some SD 1.5 anime models ask for.

#Values and what they change

• type must match the model family.
• Text encoder files come in precisions: fp16 is the reference; fp8 versions (t5xxl_fp8_e4m3fn, umt5_xxl_fp8_e4m3fn_scaled) take half the memory with a small loss.
• stop_at_clip_layer: -1 is normal; -2 is "clip skip 2".

#Typical mistakes

• Only one of two encoders loaded: an error, or a model that ignores much of the prompt.
• The wrong type: nonsense pictures.
• Writing tag lists for a sentence reader (or long prose for CLIP): see How models read the prompt.
• Running out of memory on T5-XXL: use the fp8 file.

#How the app helps

Check tells when a prompt is longer than CLIP's 75-token chunk or too long for T5. The card of a model in the Models tab shows its family, the prompt lesson explains the two ways of reading, and Which loader for which file sorts the loaders. In Properties, a field with a built-in note has a button What this field does next to its name; it opens the note under the field: what the field does and what to put in it.

#See also

• The prompt
• Prompt syntax and weights
• Conditioning
• Separate loaders
• Model families