Contents

ComfyUI Prompting Tutorial: Positive and Negative Prompts and Weight Syntax

Prompts determine what appears in an image. This section covers how to write effective prompts and the special syntax used in ComfyUI.

Write in English

Most image models are trained primarily on English data, so English prompts work best. If you're not comfortable writing in English, first describe the image you want in Chinese, then ask a translation tool or AI assistant to translate it into English. Some newer models, such as Qwen-Image, understand Chinese better; check the model documentation for details.

Prompt structure

A good prompt usually includes these parts:

Part	Example
Subject	a young woman reading a book
Details	wearing a knitted sweater, long black hair
Setting	in a quiet library, wooden shelves
Style	watercolor illustration / cinematic photo / anime style
Lighting	soft morning light, golden hour, neon lights
Composition and lens	close-up, wide shot, from above, 35mm lens, shallow depth of field
Quality	highly detailed, sharp focus

Combined:

a young woman reading a book in a quiet library, wearing a knitted sweater, long black hair,
wooden shelves, soft morning light through the window, cinematic photo, 35mm lens,
shallow depth of field, highly detailed

Prompting styles for different models

Model	Prompting style
SD 1.5 / SDXL	Focus on keywords, separated by commas. Earlier words have more influence.
Anime-style SDXL fine-tunes	Often use tags such as `1girl, solo, smile, school uniform`; see the examples on the model page.
Newer models such as Flux and Qwen-Image	Understand complete natural-language sentences. You can write a paragraph as if describing a photo, and they can also render text more accurately.

With newer models, a clear description is usually more effective than a stack of keywords:

A photo of a small coffee shop at night. A wooden sign above the door reads "Good Coffee".
Warm yellow light comes from inside, and rain is reflecting the light on the street.

Negative prompts

Tell the model what you don't want. Common choices for SD 1.5 / SDXL include:

blurry, low quality, worst quality, jpeg artifacts, watermark, text, signature,
deformed, bad anatomy, extra fingers, cropped

Include only problems that actually occur in your negative prompt; you don't need to stuff it with a long list.
Many newer models (and cases where CFG is set to 1) don't use negative prompts. Follow the settings in the workflow template.

Weight syntax

In ComfyUI, you can use parentheses to adjust the importance of a word:

Syntax	Effect
`(word)`	Weight 1.1
`(word:1.3)`	Weight 1.3 (stronger)
`(word:0.7)`	Weight 0.7 (weaker)
`(red hat, blue scarf:1.2)`	Adjust the weight of multiple words together

After selecting text, press Ctrl + ↑ / Ctrl + ↓ to adjust its weight quickly.

Weights are generally between 0.5 and 1.5. Values that are too high can distort the image.

ComfyUI calculates weights a little differently from A1111 WebUI, so prompts copied from WebUI may not produce exactly the same results.

Useful keywords

Photography

photo, DSLR, 35mm, 85mm portrait, bokeh, shallow depth of field, film grain, long exposure

Lighting

soft light, golden hour, backlit, rim light, studio lighting, volumetric light, neon

Composition

close-up, medium shot, full body, wide angle, from above, from below, symmetrical, rule of thirds

Style

watercolor, oil painting, pencil sketch, anime, pixel art, 3d render, isometric, flat illustration, ukiyo-e

An iterative approach

Start with a short prompt and confirm that the subject is correct.
Fix the seed, then add or change one part at a time and observe the results.
If an element keeps missing, move it earlier, increase its weight, or describe it more specifically.
If an element is too strong, lower its weight or remove the related description from the prompt.

Changes to text encoder nodes

In addition to the basic CLIP Text Encode, some models use dedicated text encoder nodes (such as the advanced SDXL node or Flux nodes) that let you set resolution or guidance. When using an official workflow template, use the nodes provided in that template.

Updated October 12, 2026