Just because a model can produce 4K does not mean that you can produce good pictures with just a few words. The higher the resolution, the more obvious the areas that are not explained clearly in the prompt words will be exposed. The following are ideas for writing three common requirements for MiniMax H3 Image.
Multi-picture reference: Give each picture a good job
H3 Image can eat up to 16 reference images. There are many pictures. The most common problem is that the model does not know which picture controls what, and mixes the colors of picture A into the products of picture B.
The writing method should be named and numbered, for example: "Pictures 1 to 3 are for the product, refer to the shape, color and material; picture 4 is for the model, refer only to facial features and hairstyle; picture 5 is the scene, refer only to light and tone, no objects in it should appear."
The last sentence "Don't show the objects in it" is very important. Clutter in the scene reference pictures is often moved into the final picture.
Small print: write in full, write in short, write in position
The text on posters and detail pages is where 4K images are most likely to overturn. A few experiences:
- Enclose the text to appear in quotation marks and write it word for word.
- Each block of text should be as short as possible, and there should not be too many blocks of text in one picture.
- Indicate the location and size, such as "big title in the center at the top" and "small text in the lower right corner"
- After drawing the picture, zoom in and check word by word. If you make a mistake, you can use local editing to correct it individually.
Material: use specific nouns and less adjectives
For words such as "high-end" and "good texture", it is difficult for the model to fall on the pixels. Change to specific materials: frosted glass, brushed aluminum, top-grain cowhide, cashmere knit, matte ceramic. Plus light: side backlighting, softbox front lighting, and natural light from the window. The material texture is visible in 4K. The more specific you write it, the less plastic it will look like.
High speed first, standard later
H3 Image has a Highspeed version, which officials say is about 3 to 4 times faster than mainstream tools. When adjusting the prompt words, use the high-speed version to quickly test. Once the writing method is stable, switch to the standard mode to produce a 4K finished product, which can save a lot of time.
These writing methods can be tried directly on PixPix, the official partner site of MiniMax, and PixTV. PixPix prefers product pictures, and PixTV prefers short drama storyboards and videos. The site also contains entries for MiniMax.
If you want to make an AI short drama by yourself, you can first go to TVLab to look through the library of short drama routines and opening hooks. The topic selection will be much faster.