[Experiment] Prompts for Sanpaku Eyes | 9 Conditions, 27 Images

Conclusions First

• upward gaze was the only effective condition for reproducing sanpaku eyes (where the iris sits high and white is visible below). In all 3 images, the iris moved upward and white sclera was visible beneath it
• sanpaku / sanpaku eyes are direct terms but had almost no effect in z-image-turbo. Iris position was nearly identical to the control
• three whites interpreted “three” as a quantity and caused a critical side effect — multiple faces generated in a grid layout
• lower white caused the framing to shift downward, cutting the eyes out of frame
• cold stare / dead eyes / intense stare affect expression (neutral/emotionless) but did not change iris position

Experiment Design

Base Prompt

upper body + frontal composition was used to allow clear observation of iris position.

A00: control (no keyword)

seed 1	seed 2	seed 3

Observation: In all 3 images, the iris is positioned near the center of the eye — no white visible below. s1 is looking slightly to the right; s2 and s3 have the body angled sideways with the face turned toward camera. Expression is natural and calm.

A01: sanpaku

seed 1	seed 2	seed 3

Observation: No clear change in iris position across all 3 images compared to control. s1 and s2 have nearly identical composition and expression to control, with iris centered. s3 shows a slight smile not seen in the control, but no white sclera below the iris. sanpaku alone does not appear to communicate the concept of sanpaku eyes to z-image-turbo.

A02: sanpaku eyes

seed 1	seed 2	seed 3

Observation: s1 shows a tighter crop of the face with slightly wider open eyes, but iris remains centered. s2 has the body angled sideways looking back at the camera — the iris sits slightly higher than center and a faint lower white is barely visible. s3’s iris position is nearly the same as control. A slight effect was confirmed in 1 out of 3 images, but reproducibility is low.

A03: three whites

seed 1	seed 2	seed 3

Observation: Critical side effects in all 3 images. s1 shows a collage of 6 faces in a grid, s2 similarly shows a 6-panel grid, and s3 generates 2 people of different sizes in one image. “Three” is interpreted as a quantity, producing multiple subjects. The iris position of individual faces is normal — no sanpaku characteristics. Not usable for sanpaku purposes.

A04: lower white

seed 1	seed 2	seed 3

Observation: In all 3 images, the framing shifts dramatically downward, cutting the eyes out of frame. s1 shows only the area from the mouth down, s2 shows the chin and neckline area, and s3 shows only the upper body from the neck down. “Lower” was interpreted as a compositional directive, causing the camera framing itself to drop. With no eyes visible, evaluating iris position is impossible.

A05: upward gaze

seed 1	seed 2	seed 3

Observation: In all 3 images, the iris clearly shifts upward and lower white sclera is visible — sanpaku is reproduced. s1 faces forward but gaze alone is directed upward, with clearly visible lower white. s2 and s3 have the face tilted slightly upward, but in both cases the iris sits high in the eye with lower white exposed. This condition shows the most clear change in iris position compared to control. As a side effect, s3 shows a change in hair length.

Lab Director Comment: Wait, this is actually great. The lower white for sanpaku is right there. It’s kind of interesting that sanpaku doesn’t work but you can pull it off by going the roundabout route with upward gaze — prompt paraphrasing skills matter.

A06: cold stare

seed 1	seed 2	seed 3

Observation: Iris position remains near center in all 3 images — no lower white visible. The difference from control is in expression: all 3 show a neutral, closed-mouth expression without the softness of the slight smile in control s3. s1 is cropped slightly closer, s2 has the body angled sideways with a slightly sharper gaze. No effect on gaze direction or iris position — not usable for sanpaku reproduction.

A07: dead eyes

seed 1	seed 2	seed 3

Observation: Iris centered in all 3 — no lower white. Difference from control is in expression: all 3 show a near-expressionless face. s1 faces forward with a rigid expression, s2 is expressionless facing forward, s3 similarly shows an unreadable expression. Compared to control, the smile and eye movement are gone — expressions trend flat. However, no change in iris position, so this doesn’t contribute to sanpaku reproduction.

A08: intense stare

seed 1	seed 2	seed 3

Observation: Iris centered in all 3 — no lower white. Difference from control is similar to cold stare — expression only. All 3 show a neutral forward-facing expression. s3 has the face tilted slightly upward, but not as pronounced as upward gaze, and lower white is not visible. No clear difference from cold stare was confirmed either.

Summary

Results for all 9 conditions:

upward gaze is the most effective condition for reproducing sanpaku eyes (iris raised with lower white visible). However, it tends to tilt the entire face upward rather than just moving the gaze — additional techniques would be needed if you want sanpaku with a straight-on face.

The direct term sanpaku had no effect in z-image-turbo. cold stare / dead eyes / intense stare affect expression (neutral, emotionless) only — not iris position. They’re useful for tone, but produce a different effect from sanpaku.

Lab Director Comment: So, upward gaze is the only real answer for sanpaku eyes. It’s a shame sanpaku doesn’t land, but the approach of paraphrasing visual features into English seems applicable elsewhere too.