Compel Prompt Weighting

This guide covers Compel’s prompt weighting syntax and long prompt support for CLIP-based diffusion models.

What Is Compel?

Compel is a prompt encoding library that provides two features for CLIP-based models:

Prompt weighting — control the influence of individual concepts with (word:weight) syntax
Long prompt support — break prompts longer than 77 tokens into chunks, removing CLIP’s token limit

Compel works by converting text prompts into pre-computed embeddings that are passed directly to the diffusion pipeline, bypassing CLIP’s tokenizer limits.

Supported Models

Compel is available for models that use CLIP text encoders:

Model	Architecture	Compel Support
Juggernaut XL v9	SDXL	Yes (dual text encoders)
DreamShaper XL Lightning	SDXL	Yes (dual text encoders)
SDXL Turbo	SDXL	Yes (dual text encoders)
Realistic Vision v5.1	SD 1.5	Yes (single text encoder)
Flux.1-dev	Flux	No (T5 text encoder)
Qwen-Image-2512	Qwen	No (Qwen text encoder)
Z-Image Turbo	ZImage	No (custom text encoder)

Models without Compel support use their native text encoders directly and handle long prompts via max_sequence_length configuration.

Prompt Weighting Syntax

Use (concept:weight) to control how strongly a concept influences the generated image:

Weight > 1.0 — emphasize the concept
Weight < 1.0 — de-emphasize the concept
Weight = 1.0 — default (same as no weighting)

Examples

# Emphasize dramatic lighting and de-emphasize blur
a beautiful landscape with (dramatic lighting:1.5), avoiding (blur:0.5)

# Strong emphasis on a specific style
portrait of a woman, (oil painting style:1.4), (soft focus:0.7)

# Emphasize specific elements in a scene
a cozy cafe with (warm lighting:1.3) and (steam rising from coffee:1.2)

# Balance multiple concepts
(photorealistic:1.3) cityscape at (golden hour:1.5), (cinematic:1.2) composition

Weight guidelines:

1.1–1.3 — subtle emphasis, nudges the model toward the concept
1.3–1.5 — noticeable emphasis, concept becomes a focal point
1.5–2.0 — strong emphasis, may produce exaggerated results
0.5–0.8 — gentle de-emphasis, concept is present but reduced
< 0.5 — strong de-emphasis, concept is largely suppressed

Long Prompt Support

Without Compel, CLIP-based models truncate prompts at 77 tokens (~50–60 words). With Compel, prompts of any length work:

# This long prompt works without truncation on SDXL/SD15 models:
a highly detailed, photorealistic landscape painting of a serene mountain
valley at sunset, with dramatic lighting, golden hour atmosphere, misty
background, lush green meadows in the foreground, snow-capped peaks in
the distance, a winding river reflecting the warm sky colors, scattered
wildflowers in purple and yellow, ancient pine trees framing the scene,
wispy clouds catching the last rays of sunlight

Compel automatically chunks the prompt into 77-token segments and concatenates the resulting embeddings, preserving the full prompt content.

LoRA Integration

When a LoRA is active, its trigger words are appended to the prompt before Compel encoding. Since Compel removes the 77-token limit, trigger words are always included in full — there’s no risk of them being truncated.

You can also apply weighting to LoRA trigger words:

a mountain landscape, (phrsink:1.2), pen and ink illustration

Practical Tips

Start subtle — begin with weights of 1.1–1.3 and adjust from there
Don’t over-weight — values above 1.5 often produce artifacts or distorted results
Combine with negative prompts — for models that support them (Juggernaut XL, Realistic Vision), negative prompts complement weighting by excluding unwanted elements
Test without weighting first — generate a baseline image, then add weighting to steer specific aspects
Weighting doesn’t work on non-CLIP models — Flux, Qwen, and Z-Image ignore the (word:weight) syntax; the parentheses and numbers become part of the literal prompt text