1 · Word order changes the picture
The model weights early tokens more heavily than late ones. Whatever appears in the first third of your prompt is what it commits to; the rest nudges. So put what the picture is before where it happens.
The generator already does this for you when you use the chips — subject, outfit, pose, scene, extras, then your free text last. Your words land at the end deliberately, so they refine the scene rather than fighting it.
Weaker
in a sunlit kitchen with plants everywhere and morning light coming through the window, a woman making coffee
Four location clauses arrive before the subject. The model spends its confidence on the room and hands you a beautiful kitchen with a vague person in it.
Stronger
a woman making coffee, cream sweater, hair tied back, sunlit kitchen, plants on the sill, morning light
Subject first, then what she looks like, then the room. Same information, ordered so the person is the picture.
2 · Say how many people are in the shot
This is the single most useful habit and almost nobody does it. If a prompt implies contact or a second party without stating a count, the model guesses — and it guesses badly, often adding a person, sometimes adding a limb.
Write solo when it is one person. Write the count explicitly when it is more.
Weaker
leaning on someone's shoulder, laughing
"Someone" is not a person the model can count. Expect two heads sharing a torso, or a third arm.
Stronger
two women, one leaning on the other's shoulder, both laughing, park bench, autumn
Stated count, stated relationship between them, then the setting. The anatomy problems mostly disappear on their own.
Why this matters more on anime. Anime checkpoints are trained on tagged images where the character count is always written down. Leave it out and the model fills the gap from what it saw most often, which is rarely what you meant.
3 · Anime and realistic want different sentences
These are two different models on this site, not one with a style switch, and they read prompts differently.
Realistic: write sentences
Photographic checkpoints understand descriptive language. Camera words work: shallow depth of field, backlit, 35mm, overcast. Describe it the way you would describe a photograph to someone.
Anime: write tags
Anime checkpoints are tag-trained. Short comma-separated descriptors beat flowing prose, and the quality prefix these models expect is added for you automatically.
Realistic
a woman reading by a window, soft overcast light, shallow depth of field, knitted cardigan, cup of tea on the sill
Anime, same idea
1girl, solo, reading a book, window light, cardigan, teacup, indoors, soft shading
Both are good prompts. Swapping them produces worse results in each direction — a tag list gives the photographic model very little to work with, and a paragraph gives the anime model something it was never trained to parse.
4 · Specific beats long
A prompt is not a wish list. Past roughly forty words the model starts averaging your requests together instead of honouring all of them, and you get a picture that is vaguely all of it and precisely none of it.
Cut adjectives that do not change pixels. "Beautiful", "amazing", "stunning" and "perfect" describe your opinion of the result, not the result. "Freckles", "wet hair", "chipped nail polish" describe things a model can draw.
Weaker
a beautiful gorgeous stunning woman, amazing quality, perfect face, best photo ever, ultra detailed masterpiece, incredible lighting
Nine words of praise and one noun. There is nothing here to draw.
Stronger
a woman just out of the sea, wet hair pushed back, salt on her shoulders, late afternoon sun, squinting slightly
Every phrase changes something visible. Shorter, and far more directive.
5 · Negatives, and where they stop working
Image generations use a negative prompt, and it is already filled in with what these checkpoints need. You do not have to write one.
Two things worth knowing about negatives generally:
- Never put a colour in a style negative. Excluding a colour drains it out of everything, including skin. This cost us weeks of washed-out anime output on a sister product before anyone traced it.
- A negative cannot add anything. It only suppresses. If you want something in the picture, it goes in the positive — there is no way to ask for a result by ruling out its opposite.
Video ignores negatives completely. Clips render at a guidance setting where the model skips the unconditional pass entirely, so a negative prompt is not weak there — it is not read at all. Anything you want in a clip has to be in the positive prompt.
6 · Video is decided by the source image
Clips start from an image you already made, and that image does more work than the motion prompt does.
- Framing sets how much can move. A tight head-and-shoulders crop has nowhere to go, so you get breathing and a small head turn. A full-body shot leaves room for walking, swaying, reaching.
- If a clip feels static, change the picture, not the settings. This is the most common wrong turn. People tune steps and schedulers for an hour when the answer was a wider source shot.
- Describe one motion, not three. Five seconds fits one idea. "She turns her head and laughs and pushes her hair back and looks away" gets you a blur of all four.
Weaker
she moves around a lot, very dynamic, lots of action, cinematic
No specific motion, and "cinematic" means nothing to a motion model.
Stronger
she turns her head slowly toward the camera, hair sweeping across her shoulder
One motion, a direction, and a secondary detail that sells it. This is roughly what the built-in motion presets look like.
7 · Five things that quietly waste credits
- Fighting a chip with your text. If you selected a scene chip and then describe a different room in your prompt, the model receives both and blends them. Deselect the chip instead.
- Stacking every extra. Extras are modifiers; six at once average each other out. Two is usually the ceiling before they start cancelling.
- Re-rolling instead of editing. A different seed on the same prompt gives you a different version of the same mistake. Change one phrase and re-run.
- Asking for a wardrobe change in a video prompt. The clothing is already decided by the source image. Demanding something different is what makes outfits flicker mid-clip.
- Describing the mood instead of the frame. "Intimate", "sensual", "moody" are things a viewer feels, not things a model draws. Give it the light, the pose and the distance, and the mood follows.
Start with something small
Pick a scene chip, an outfit chip and one pose, write six words of your own, and generate. Then change exactly one thing and generate again. Two rounds of that teaches more about how these models read than any list can, including this one.