Building a prompt up
What locks down as you add one clause at a time?
Starting out, it is not obvious what to write first. Below, appearance, then setting, then light and framing are added to a subject. The seed is fixed, so the prompt is the only thing that changed.
Clothing is specified from step one. Left unspecified, this checkpoint defaults somewhere unpublishable — an empty prompt is not neutral, it hands the frame to the model's bias. That is part of the lesson.
The cuts below share a seed and a sampler, but the prompt differs from cut to cut. Open any cut to see the full prompt it received.
What each cut shows
step 1- Subject and clothing only. With no setting named, the model picked an indoor doorway on its own, and with no hair colour named the hair drifted to ash brown.
step 2- Adding appearance and expression pinned the hair colour and eyes. The setting was still unwritten, yet the indoor doorway scene the model chose in step one carried over.
step 3- Pose and setting bring the park path in, placing the figure inside a scene.
step 4- Light and framing lock a full-body straight-on shot with sunlight through the leaves. From here, rerolls stay close.
How these were made
Shared prompt
Differs per cut — each cut page has it in full.
Negative prompt
nsfw, nude, topless, bottomless, bikini, swimsuit, lingerie, underwear, panties, bra, cleavage, ass focus, breast focus, thigh focus, wet skin, suggestive, 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, missing fingers, extra digits, deformed, mutated, blurry, lowres
Model
rinFlanimeIllustrious_v30 · euler_ancestral/normal · steps 28 · CFG 5 · 896×1152 · seed 777005
Takeaway
Whatever the prompt omits, the model fills in. If your results keep moving, look for what you did not write — usually the setting and the framing.
How far this result reaches
Which clause locks what varies by checkpoint. This model makes unusually strong assumptions about the setting.
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.