Earngenix Logo
Skip to main content

Glossary · ComfyUI Basics

What Is CLIP in ComfyUI?

By Earngenix Team ·

⚡ Quick Answer

CLIP is a text encoder model that reads your prompt and converts it into a numerical representation called conditioning — the format the sampler actually understands. Without CLIP, your typed words never reach the image-generation process at all.

Stable Diffusion and Flux-family models can't read English directly. CLIP (Contrastive Language-Image Pretraining) was trained to map text and images into the same mathematical space, so a phrase like "a red bicycle" produces a vector the checkpoint was trained to associate with red bicycles.

Where You'll See It

On SD1.5 and SDXL checkpoints, Load Checkpoint already outputs a CLIP connection alongside MODEL and VAE — nothing extra to load. On Flux and other split-architecture models, a DualCLIPLoader or CLIPLoader node loads the CLIP files separately, and its output feeds directly into CLIP Text Encode (Prompt), which is where you actually type your prompt text.

Quick Example

A Flux workflow commonly loads clip_l.safetensors together with t5xxl_fp8_e4m3fn.safetensors through a DualCLIPLoader node set to type flux. CLIP-L handles short, style-driven phrasing, while the much larger T5XXL model handles longer, more literal instructions in the same prompt.

If a workflow you downloaded shows a red "Load Checkpoint" or "CLIPLoader" node, it's almost always a missing CLIP file — check the exact filename in the node and download it into ComfyUI/models/clip.

Common Mix-Up

CLIP vs. CLIP Vision: CLIP (used here) encodes text. CLIP Vision is a related but separate model that encodes images instead, and is what nodes like IP-Adapter use to read a reference photo — the two are not interchangeable.

Frequently Asked Questions

For SD1.5, SDXL, and most single-file checkpoints, no — the Load Checkpoint node already outputs CLIP alongside MODEL and VAE. For Flux and other newer architectures, CLIP ships separately and you load it with a DualCLIPLoader or CLIPLoader node.

The prompt still gets encoded, but the resulting conditioning won’t match what the checkpoint was trained to expect. You typically get images that ignore parts of the prompt, look garbled, or lose fine detail and text accuracy.

Not exactly — Flux uses two text encoders together: a CLIP-L model and a much larger T5XXL model. CLIP handles shorter, style-heavy cues while T5XXL handles longer, more literal instructions.

See It In Action

Ready to see CLIP at work?

Our interface overview walks through every core node, including where CLIP sits in the graph.

Published: 2026-09-04 · Last updated: 2026-09-04

Discussion

Join the discussion

Sign in to leave a comment or reply

💬

No comments yet

Be the first to share your thoughts!