Вид сверху - пристрелка промптов

Изучаю работу с перспективой в рисовальных нейронках. Конкретно в Fooocus (Stable Diffusion, а точнее SDXL). Нужно как-то на английском сказать - вот, рисуй с видом сверху. GPT 4 дал мне примеры, я их пробую.
Он понимает смысл перспективы, но ужасно косячит в деталях. Я предполагаю, потому что нейронку учили в основном не на таких фото. Это не стандартные рисунки, не массовые перспективы, поэтому нейронка рисует сама, а когда она рисует сама, то получается много галлюцинаций. Ибо опыта жизни и анатомии у нее нет.
======
Собственно сбивается даже край стола.
Top-down view of a cozy kitchen, as if the viewer is suspended from the ceiling. A family of four is preparing dinner together. There's a father, mother, a teenage son, and a young daughter. The kitchen is filled with modern appliances, and there's a large, colorful fruit bowl on the counter.


=============================
В этом промпте что-то есть, тут можно поиграть со стилем, с атмосферой, поменять все на аниме
An aerial view of a room, as if the viewer is looking straight down from the ceiling. In the center of the room, there's a table with a map spread out on it. Around the table, precisely four female characters are seated: a magician, a warrior, a priestess, and a thief. Each character is distinct in their attire and appearance, reflecting their roles. The magician is cloaked in mystical robes, the warrior wears armor, the priestess is in ceremonial garb, and the thief is in stealthy attire. The room is dimly lit, with shadows casting across the walls, creating a mysterious and ancient atmosphere.


=============================
Aerial perspective of a bustling city street at night, as seen from above, like observing from a high rooftop. Bright neon signs, diverse groups of people walking, taxis and buses moving along the street, and small shops lining both sides of the road.


=============================
View from above a classroom, like peering down from the ceiling. Twenty students, a mix of boys and girls, are focused on their exams. The teacher, a middle-aged woman, is monitoring them closely. The classroom is well-lit, with student artwork on the walls.


=============================
Overhead view of a public library reading room, as if looking down from a high balcony. People of various ages are scattered around, reading, researching on laptops, and browsing shelves. The room is spacious with large windows and rows of bookshelves.


=============================
Bird's-eye view of a beach scene, as if observed from a high cliff. Families are picnicking, kids are building sandcastles, and surfers are riding waves. The beach is bordered by a lively boardwalk with small shops and food stalls.


=============================
Looking down on a crowded concert hall from an upper balcony. A symphony orchestra is in the midst of a performance, with a diverse audience in rapt attention. The hall is grand, with ornate decorations and a large, shimmering chandelier.


=============================
Aerial view of a farmer's market, as if seen from a drone above. Stalls are filled with fresh produce, artisanal goods, and flowers. Shoppers of various ages and backgrounds are browsing and interacting with vendors. The atmosphere is lively and colorful.


=============================
Похоже, у этой нейронки не было в обучающих данных фотографий людей с самолетами - и она не понимает их взаимные пропорции
Overhead perspective of a busy airport terminal, like observing from a high vantage point. Travelers are checking in, waiting, and walking with luggage. The terminal is modern with large information screens, and outside, planes are visible on the tarmac.


=============================
View from above an open-plan office space, as if looking down from the ceiling. Workers are at their desks, in meetings, and taking breaks in communal areas. The office is modern and bright, with plants, comfortable seating, and collaborative workspaces.


=============================
Top-down view of a sprawling garden party, as if viewed from a tall tree. Guests in elegant attire are mingling, a string quartet is playing, and there's a buffet table set with an array of dishes. The garden is lush, with flower beds, pathways, and twinkling lights.


Все тестировалось на 1 seed, на одних настройках в Fooocus (дефолтные стили)
Выводы
- искажения страшные почти всегда
- на перспективу с видом сверху работают фразы виды
[Top-down | Aerial] [view of | View from above] [ xxxxxx ], as if [viewed|looking] from a [ drone | ceiling | tall tree | an upper balcony ].
В квадратных скобках даны варианты.
Полагаю, может работать во всех рисовальных нейронках. И это не все варианты. Английский богаче.
|
</> |