Skip to content

Best Practices for Writing Prompts

1. Controlling Image Content

  • The AI will try to display whatever is written in the prompt;howeverif something is not mentioned in the prompt, it may not appear in the image.
  • Ifthe prompt is written in a confusing manner, the image quality will likely be poor.
  • Ifthe content of the image is beyond conventional imagination, the AI may not be able to depict it perfectly. For example, toothbrushes, in the shape of bumblebeeswill balance betweentoothbrushesand bumblebees：
Leaning towards toothbrushes Leaning towards bumblebees Compromise
alt text alt text alt text
  • More prompts are not always better，too many can make the image cluttered, anda few prompts can still yield good results. For example, using only1 car , photography, high-rescan produce the following image:

alt text

2. Controlling Image Style

  • Generally, you can control the style by using common terms such asanime、comic foranime illustration style，realistic、photograpyforrealistic style
  • Refer to here, for modifying image styles.
  • Forambiance、design sense、high-end feeletc., it's recommended to usedescriptive terms for the content of the imageto achieve better results.
  • Forearly large models，you might need to switch modelsto change styles effectively. For recent large models like FLUX or SD3.5, if a specific image style is needed, it’s recommended to load a matching lora model for better results.
  • Due to model training, some prompts are implicitly linked with styles. For example, 1 girl tends to trigger anime illustration style, while 1 woman is more realistic. Adjust prompts carefully based on the desired outcome.
1 girl, bike, spring 1 woman, bike, spring
alt text alt text

3. Controlling Composition

  • Describe the structure of the image and its content to help the AI better express the composition, such as at the top of the image、at the bottom of the image、center section,etc..
  • Provide camera information to control the viewpoint, such asclosely view、looking at viewer、top view,etc,.

4. Removing Image Size Information

  • In Stable Diffusion, size information is set separately, so there's no need to include size information in the prompt (width, height, aspect ratio).

5. Image Aspect Ratios

  • Recommended ratios are1:1、2:3、3:2、3:4、4:3、16:9、9:16. Ratios like 1:2, 1:3, 2:1, 3:1 (long or wide) are less effective compared to conventional sizes.
  • SD 3.5系列 is suitable for 1 million pixel resolution professional use cases. For the FLUX series, Pro, Dev, and Schnell versions can output images up to 2 million pixel resolution, while Pro Ultra can output up to 4 million pixel resolution. Dev and Schnell can be used locally, while Pro and Pro Ultra require official API usage.

6. Adjusting Prompt Weights

  • To emphasize or de-emphasize certain prompts, use () or [] for weight adjustments. Each layer of parentheses adjusts the weight by 1.1x or 0.9x, e.g.,((cow)) [1.21x emphasis], (boy)[1.1x emphasis],[moon] [0.9x de-emphasis].
  • You can also use (cow:1.5)to directly adjust the weight of a specific prompt.

7. Common Issues

  • AI still struggles with accurately understanding human anatomy; try to avoid or partially redraw problematic areas. Common issues include hand deformities, body proportion distortions, and awkward lying poses.
  • The greasy look issue often occurs when depicting realistic scenes, characterized by overly slick textures and lack of grain. Solutions include:
  • Ensuring the prompt does not lean towards realistic scenes.
  • Adding terms like photograph, realistic style, studio photography to the prompt to emphasize realism.
  • Incorporating Lora models specialized for realistic scenes in the workflow. The advantage is better image quality, but the disadvantage is reduced flexibility in switching styles.