Which way the subject faces
How precisely do the tags that turn a subject actually turn them?
Horizontal angle separates more cleanly than any of the other axes. Unlike height or framing, the outcome is discrete: facing you, turned aside, or turned away.
The catch is spelling. Natural-looking English like side_view or rear_view drew no response on a booru-trained checkpoint. The four below are the spellings that worked.
Every cut below shares one seed, one base prompt and one sampler setting, and differs only by the tag marked in amber.
What each cut shows
straight-on- Face on, as if in conversation. Most prompts drift here anyway, but naming it locks the composition far more reliably.
from side- The whole body turns. Not just the face: the shoulder line and the sweep of the coat rotate with it, so the silhouette changes a great deal.
from behind- Her back is to the camera, but she turns her head and her profile shows. The back of the coat and the hair carry the frame; for a fully hidden face this tag alone was not enough.
profile- Close to
from side, but the emphasis lands on the facial outline. The body turns less and the head commits more.
How these were made
Shared prompt
masterpiece, best quality, amazing quality, very awa, 1girl, solo, mature female, adult woman, short black hair, brown eyes, white blouse, beige trench coat, black trousers, standing, city street, daytime, clear sky, <tag slot>
Tap a tag to carry it over to the builder
Negative prompt
school uniform, pleated skirt, blazer, serafuku, child, loli, teenage, bad quality, worst quality, worst detail, sketch, censor, jpeg artifacts, watermark, signature, text, extra limbs, bad hands, bad anatomy
Model
rinFlanimeIllustrious_v30 · euler_ancestral/normal · steps 28 · CFG 5 · 832×1216 · seed 777001
Takeaway
All four work. Subject orientation is the most reliably controllable axis in the prompt, so if your compositions keep drifting, pin this one first.
How far this result reaches
The two side-facing tags overlap. Using both together tends to produce a hedge rather than a stronger turn.
Measured on a single Stable Diffusion family checkpoint. The same tags can behave differently on NovelAI, Midjourney or another checkpoint. This is a reference point, not a verdict.