Recreating the Sumo 'Hakkeiyoi' Stance with Prompts | 8 Conditions, 24 Images

Key Takeaways

• crouching, wide stance, hands on thighs produced the closest hakkeiyoi pose. In 2 out of 3 images, the wide-legged low crouch with hands on knees and forward lean matched the shikiri stance
• deep squat, legs apart, hands on knees also scored high. All 3 images produced a wide-legged squat with hands near the knees
• Words containing sumo caused the body type to shift larger and more muscular. sumo wrestling stance caused a noticeable physique change in 3 out of 3 images, and a wrestling mat appeared in the background
• shiko dachi stance (karate horse stance) didn’t work as expected. It was interpreted as a fighting/flexing pose rather than a low crouch
• [Deep dive] Adding fists on ground produced a shikiri stance with fists touching the floor. 1 out of 3 images achieved a perfect recreation, with another close match
• Element decomposition without sumo vocabulary is the optimal approach for maintaining body type while achieving the hakkeiyoi pose

What You’ll Learn

• How to express the hakkeiyoi (shikiri) stance in English prompts
• Side effects of sumo-related words on body type and backgrounds
• The effectiveness of decomposing a pose into component elements
• Whether the karate term shiko dachi works in AI image generation

Experiment Setup

Base Prompt

Only the {POSE} portion changes per condition. Clothing (tank top, shorts), location (indoor gym), and framing (full body) are all fixed. The gym background was chosen to prevent sumo-related words from pulling in a dohyo or traditional Japanese room.

Results by Condition

A00: Control — standing, facing camera

seed=42	seed=123	seed=789

Observations

All 3 images show a woman standing upright in a gym, facing forward. Arms are naturally at the sides. Tank top and shorts are maintained. No anatomical issues detected. Functions well as a baseline.

A01: sumo wrestling stance

seed=42	seed=123	seed=789

Observations

All 3 images produced a wide-legged power pose with bicep flexing rather than the hakkeiyoi crouch. The squat depth is shallow, and there’s no forward lean. The biggest issue is body type inflation: all 3 images show noticeably broader shoulders and a more muscular build. Seeds 42 and 789 also generated a wrestling mat on the floor despite the indoor gym specification. Hands go upward rather than toward the knees.

A02: sumo crouch

seed=42	seed=123	seed=789

Observations

A major improvement over A01 in hakkeiyoi accuracy. All 3 images achieved a wide-legged low crouch. Seed 42 has hands on the knees, seed 789 shows a forward lean with hands on knees — close to the shikiri stance. However, the body type trends slightly larger compared to A00, with thicker thighs visible in seeds 42 and 123. The gym background was maintained. The best result among conditions containing sumo.

A03: wide squat stance, leaning forward

seed=42	seed=123	seed=789

Observations

wide squat was interpreted as a side lunge/stretch rather than a symmetrical squat. Seeds 42 and 789 show a lateral lunge with one leg extended sideways. Seed 123 shows sitting on a bench while leaning forward. Body type matches A00 with no inflation. wide squat stance was consistently interpreted as a side lunge rather than a symmetrical wide-stance squat.

A04: deep squat, legs apart, hands on knees

seed=42	seed=123	seed=789

Observations

All 3 images produced a deep wide-legged squat. Seed 42 shows a deep frontal squat with hands on knees. Seed 123 has hands on a bench — hand placement missed, but the squat posture was maintained. Seed 789 has hands together in front but with a sufficiently deep squat. Hands landed near the knees in 2 out of 3 images. Body type matches A00 with no inflation. The element decomposition approach (deep squat + legs apart + hands on knees) reliably produces a shikiri-adjacent pose without the sumo keyword. At 7 tokens it’s slightly longer, but the consistency is high.

A05: shiko dachi stance

seed=42	seed=123	seed=789

Observations

The karate horse stance (shiko dachi) was interpreted as a fighting/power pose. All 3 images show a wide stance, but the squat depth is shallow — closer to standing than crouching. Arms display bicep flexing (seeds 42, 123) or a fighting stance (seed 789), with hands never reaching the knees. Body type shows slightly more muscular arms compared to A00, though less dramatic than A01. shiko dachi did not function as the expected horse stance, and was consistently interpreted as a generic martial arts pose.

A06: crouching, wide stance, hands on thighs

seed=42	seed=123	seed=789

Observations

2 out of 3 images produced the closest match to hakkeiyoi. Seed 42 shows a frontal wide-legged crouch with hands on the upper thighs — essentially the shikiri stance. Seed 789 shows a forward-leaning crouch with hands on knees and an intense expression, delivering the strongest hakkeiyoi feel of all conditions. Seed 123 landed on a bench but maintained legs-apart with hands on knees. Body type matches A00 with no inflation. crouching lowers the center of gravity, and hands on thighs stabilizes hand placement.

Deep Dive: Fists on the Ground — fists on ground

Building on A06 (the top performer), the hand specification was changed to fists on ground to see if we could recreate the moment when wrestlers place their fists on the dohyo.

seed=42	seed=123	seed=789

Observations

Seed 123 produced the closest match to hakkeiyoi in the entire experiment. Wide-legged deep forward lean with both fists touching the floor — the shikiri stance itself. Standard body type with no inflation. Seed 42 shows a forward lean with fists near the ground but asymmetric leg spread. Seed 789 leans toward a side lunge with one hand on the floor.

Compared to hands on thighs (A06), fists on ground pushes hands lower and increases the forward lean. Use this when you want to recreate the specific moment of placing fists on the ground. Stability is slightly lower than A06 (1-2 out of 3 vs A06’s 2 out of 3), with some risk of collapsing into a lunge.

Lab Director: Seed 123 is just straight-up hakkeiyoi. Tank top shikiri — honestly could be a new sport.

Body Type Inflation Analysis

Comparing conditions with sumo (A01, A02) against those without (A04, A06):

Words containing sumo carry a risk of shifting the body type toward a larger, more muscular build. sumo wrestling stance caused dramatic physique changes in all 3 images. sumo crouch had good pose accuracy but confirmed body type side effects. For maintaining body type while achieving hakkeiyoi, the element decomposition approach (A04, A06) is recommended.

Lab Director: The word sumo making the build bigger is honestly kind of on-brand. Points for semantic accuracy, I guess.

Summary Comparison

Recommended Prompts

Two recommended approaches for recreating the hakkeiyoi (shikiri) stance:

Top Pick: crouching-based (5 tokens)

crouching lowers the center of gravity, wide stance opens the legs, and hands on thighs stabilizes hand placement. No body type side effects. 2 out of 3 images achieved high accuracy.

Runner-up: deep squat-based (7 tokens)

Uses more tokens but delivered a stable wide-legged squat in 3 out of 3 images. hands on knees also helps prevent hand artifacts. Choose this when consistency matters more than token efficiency.

Words to Avoid

• sumo wrestling stance — Inflates body type to muscular/large. Wrestling mat bleeds into the background
• shiko dachi stance — Doesn’t function as a horse stance; gets interpreted as a generic fighting pose
• wide squat stance — Tends to be interpreted as a side lunge rather than a symmetrical squat

Lab Director: Turns out poses work better when you describe where body parts are instead of what move you’re doing. crouching + wide stance + hands on thighs — that three-piece combo is worth bookmarking.

Related Articles

[Experiment] Dynamic Pose Comparison — How Far Can Running, Jumping, and Looking Back Go Without Distortion?

Gravure Signature Pose Test: Comparing Leopard Pose, S-Curve, and Pin-Up Pose Reproducibility

[Verified] Image Generation Prompt Best Practices