Face Shape & Bone Structure Prompts — Do oval face, high cheekbones, and round face Actually Work?

When you want to craft the “ideal face” in AI image generation, do bone structure keywords like oval face or high cheekbones actually work? We tested 5 conditions with facial contour and bone structure prompts to see if they produce visible changes in facial features.

Experiment Setup

Using the standard subject 1girl, 32yo japanese actress, all images were generated under identical close-up portrait conditions.

Keywords added per condition:

3 images per condition with fixed seeds (1001, 1111, 1234), 15 images total.

Control (Base Only)

First, let’s check the output with just the base prompt.

seed 1001	seed 1111	seed 1234

Seed 1001 produced a collage-style multi-shot composition, and seed 1234 had foliage overlapping the face — unstable outputs despite the close-up portrait specification. Seed 1111 delivered a standard front-facing portrait with a fairly typical, slightly rounded contour.

Experiment 1: oval face, elongated face shape

seed 1001	seed 1111	seed 1234

Seed 1001 had a veil over the face making contour assessment difficult, though the face itself retained roundness. Seed 1111 showed no significant change compared to the control. Seed 1234, with an updo exposing the jawline, had a slightly elongated impression in 1 of 3 images. However, this was not a definitive “oval transformation.”

Experiment 2: high cheekbones, angular face

seed 1001	seed 1111	seed 1234

This condition showed the most visible change across all 5 conditions. In 3 of 3 images, shadows around the cheekbone area were more pronounced, giving the face a more defined, sculpted look. Seed 1111 in particular showed noticeably more prominent cheeks compared to the control, and seed 1234 had clearly visible cheekbone shadows. The angular face keyword likely contributed to the overall sharper appearance.

Lab Director: Interesting that cheekbones had the clearest effect. Probably because cheekbones are easy to represent through shadows. Turns out the model is better at “faking it with lighting” than actually reshaping the contour.

Experiment 3: round face, soft jawline

seed 1001	seed 1111	seed 1234

In 2 of 3 images (seed 1001, seed 1234), the jawline appeared more rounded and the cheeks slightly fuller. Seed 1111 also showed a hint of added roundness, but the difference from the control was small. round face does tend to manifest as contour roundness, but the change remained subtle rather than dramatic.

Experiment 4: narrow jaw, V-shaped chin, small face

seed 1001	seed 1111	seed 1234

In 2 of 3 images (seed 1001, seed 1234), the chin appeared slightly narrower and the overall face shape leaned toward an elongated look. Seed 1111 showed only a marginally thinner chin, with limited difference from the control. Rather than making the face area smaller (as small face might suggest), the effect seemed to tighten the jawline.

Lab Director: Three whole keywords for “narrow jaw, V-shaped chin, small face” and we barely get a slightly thinner chin. The token-to-effect ratio on bone structure prompts is honestly painful.

Same-Seed Side-by-Side Comparison

All 5 conditions at seed 1111:

Control	oval face	high cheekbones	round face	narrow jaw

Note that even with the same seed, subject identity is not preserved — each condition generates a different person. The accurate interpretation is that the keyword “pulls toward” a different person who fits that description, rather than reshaping the same face.

With that caveat, high cheekbones produced the most visually identifiable change. Round face and narrow jaw showed subtle differences, while oval face was nearly indistinguishable from the control.

Summary

The most effective bone structure prompt was high cheekbones, angular face. Cheekbones, being representable through shadows and highlights, appear to be more reliably reflected than contour-shaping keywords (oval, round, narrow jaw).

Meanwhile, shape-based descriptors like oval face and round face had minimal effect, and even the 3-keyword combination of narrow jaw, V-shaped chin, small face produced limited results relative to token cost. If you want to change the impression of a face, makeup and expression keywords may be more efficient than bone structure prompts.

Lab Director: So yeah, controlling facial bone structure through text is seriously hard. Cheekbones is the one keeper here — everything else is basically token waste.

Related Articles

[Verified] Image Generation Prompt Best Practices

Generating a 'High Nose' — Comparing high nose bridge / aquiline nose and 4 Other Keywords

[Experiment] Prompts for Blush, Contouring, and Highlight | 14 Conditions, 42 Images