即梦AI&豆包 Jimeng AI & Doubao Image Generation Tips

[baike]Text-to-image techniques[/baike]
Jimeng and Doubao use LLMs (large language models) to control descriptions and understand output, so how you control the image output depends on your description techniques.
Image generation follows the main elements of subject, environment, shot, and style; these three constitute time, space, and character—the three elements of the picture—and then there is detailed control of these three.


Example:
Subject: character
Prompt: a person
This is the most basic phrasing, but if you want precise image control, it obviously will not work. A person includes many kinds of descriptions, such as a man, 18 years old, Chinese, hairstyle, height, weight, etc. If you want to describe this person in detail, there are hundreds of ways to describe him; for example, a 35-degree pointed chin, controlled with a specific shape, will almost never change. Similarly, we can then use height and weight to roughly control the body shape. After the body details are described, there is clothing, so you need to describe the clothing's size, color, material, and features separately.

Subject: action

Prompt: playing football
"Playing football" is likewise the most basic phrasing. On a football field, playing football alone, practicing shooting, no goalkeeper, standing 30 meters from the goal, the football flying into the air, body posture showing XXXX
It is worth noting that you likewise need to describe the football—what kind of football it actually is.
Environment: football field
This phrasing is the same as when describing a person: you likewise need to describe the specific shape and details of this football field, such as 300 meters long, 200 meters wide, green lawn, specific spectator stands, etc.
The space is summer in Beijing, China, 36 degrees, 1:30 p.m., clear weather.
Shot:
Use a Konka XFWE model camera to record (without enabling any extra camera equipment), distance is medium shot (10 meters), shooting the character's back.
Style: realistic real-world style

Getting the basic prompt
The final roughly complete prompt should be: a man [Chinese, 20 years old, 180 cm tall, 80 kg, black buzz cut with the longest hair strands 5 cm, skin color #FC05216.... mouth, nose XXXX] wearing [solid green short-sleeved shirt, solid black shorts, white football socks up to the ankles] practicing shooting alone on a football field; at this moment he is 30 meters from the goal, completing the shooting motion, and the football [traditional black-and-white football] flies into the air at a position between the goal and the person, at a height of 3 meters. It is currently summer, 36 degrees, 1:30 p.m., clear weather.


With such a prompt, multiple generations will also vary only slightly. If you want to precisely control a certain part, just describe it in more detail.
This is the image-generation method for making a single image. If you want continuous image generation, we need to use the image-to-image function.

[baike]Image-to-image techniques[/baike]

Image-to-image likewise needs to use descriptions of character, time, and space, because a picture is always explained by these three things. Using image-to-image is simply that we have already generated an image—for example, we first generated a fixed character—so what we need to do next is have this fixed character perform in other times and spaces.
This is the single-image way of playing. In fact, we can also fix time and space; we can fuse a fixed character with a fixed spatial image.
The prompt is simply describing what Image A does in Image B; in this way, you can easily produce comics.