Finding the Prompt for "Muscular Woman Princess-Carrying a Man"

Key Takeaways

• princess carry has an extremely strong male→female bias — writing it straightforwardly results in the man carrying the woman 100% of the time
• holding horizontally was the first breakthrough — describing physical body positions instead of pose names bypassed the bias
• The final solution: giant muscular woman + east asian — extreme size difference plus east asian for ethnicity control. Stable across all 3 seeds
• amazonian changes ethnicity — creates great size contrast but overrides japanese. Use east asian instead
• Making the man more upright tends to reverse roles — but giant size difference improves success rate

Goal: A muscular woman princess-carrying a handsome man. She’s standing, he’s in her arms.

Step 0: The Naive Approach

Just writing what we want, plain and simple.

Result: Completely reversed. The man is princess-carrying the woman. They’re even wearing yukata — japanese pulled the model toward traditional clothing.

Issues:

• The princess-carrying bias is too strong — the woman simply won’t be the carrier
• japanese triggered kimono/yukata
• The woman isn’t muscular at all
• The man looks middle-aged

The training data overwhelmingly associates princess carry with male→female, so even making the woman the grammatical subject doesn’t help. A different approach is needed.

Step 1: Natural Language Role Clarification

Leading with an English sentence explicitly stating “a strong woman lifting a man,” plus specifying body type and clothing.

- 1girl, 1boy, 32yo muscular japanese woman princess-carrying a japanese man, simple white background
+ A strong muscular woman lifting and carrying a slim man in her arms, bridal carry pose. 1girl, 32yo japanese actress, athletic build, toned arms, sports bra and shorts. 1boy, 25yo japanese actor, slim build, casual t-shirt and jeans. studio lighting, gray background

Progress, but still far off.

What worked:

• The woman’s athletic build came through (visible abs and biceps)
• Clothing differentiation is correct
• Two distinct people (no fusion)

What didn’t:

• Not a princess carry pose — looks more like a fitness shoot pose
• bridal carry pose didn’t improve pose accuracy

Lab Director’s Take: Writing “lifting and carrying” in natural language and the model just interprets it as “two people doing fitness poses together.” The pose breakthrough needs a totally different angle.

Step 2: Breaking Through with “Holding Horizontally”

The turning point. Instead of using pose names like princess carry, I described the physical position of the bodies: holding horizontally and off the ground.

- A strong muscular woman lifting and carrying a slim man in her arms, bridal carry pose.
+ A tall muscular woman holding a man horizontally in her arms. She stands upright, he is off the ground with his legs dangling.

Key changes:

• holding horizontally — explicitly stating the man’s body orientation
• She stands upright, he is off the ground — clarifying the spatial relationship
• bodybuilder physique, defined abs — more specific muscle description

First time the intended composition appeared!

• Woman standing upright, man floating horizontally
• Woman’s bodybuilder physique is clearly visible (abs, arm muscles)
• Man in a relaxed pose being held

Remaining issues:

• Man’s head drops too low (closer to a dance dip than princess carry)
• More of a “horizontal hold” than a classic princess carry

The reason holding horizontally broke through the bias: instead of using a specific pose name (which carries gender associations), we directly described the physical positions of the bodies.

Step 3: Stability Check (3 Seeds)

Refining the Step 2 prompt with fair skin for the man and checking stability across 3 seeds.

seed 42	seed 123	seed 789

Stable across all 3 seeds. Every image shows “woman standing upright, man floating horizontally.” But still a horizontal hold rather than a true princess carry.

Step 4: Extreme Size Difference with “Amazonian”

holding horizontally is stable but feels more like a barbell hold than a princess carry. Thinking extreme size contrast might produce a more natural carry pose.

Best composition so far. The woman is carrying the man, and the size difference is clear. But amazonian changed the ethnicity. Despite writing japanese actress, the woman doesn’t look Japanese. amazonian carries ethnic connotations beyond just body size.

Step 5: Final Breakthrough with giant + east asian

Replacing amazonian with giant muscular woman and japanese with east asian. east asian lets the model interpret facial features more flexibly.

- A very tall amazonian woman effortlessly carrying a small man in princess carry.
+ A giant muscular woman effortlessly carrying a slim man in princess carry, like he weighs nothing.
- 1girl, 32yo japanese actress, extremely muscular, tall, sports bra, leggings.
+ 1girl, 32yo east asian actress, tall powerful build, muscular arms and legs, sports bra, leggings, smiling.
- 1boy, short slim fair-skinned japanese man, t-shirt, shorts, looking up at her.
+ 1boy, slim fair-skinned east asian man, oversized t-shirt, shorts.

Key changes:

• amazonian → giant muscular woman — keeps size difference without ethnic connotations
• japanese → east asian — allows the model more flexibility in facial features
• small → slim — petite/small risks making the man look like a child
• like he weighs nothing — adds effortless feeling

seed 42	seed 123	seed 789

All 3 seeds successful! Every image shows a muscular woman princess-carrying a man.

• Woman’s muscles are clear in all 3 (abs, shoulders, arms)
• Woman looks East Asian (east asian working correctly)
• Man looks like an adult (slim only, no child-like appearance)
• Seed 789 has the man’s upper body most upright — closest to a classic princess carry

Lab Director’s Take: Who knew east asian would be the key? It gives the model more room to generate Asian faces compared to japanese, which pulls too hard in specific directions. And giant creates size difference without the Amazonian ethnic baggage. Sometimes the best prompt hack is just… using a broader concept.

Failed Patterns

Recording failed attempts for reference.

Failure A: Describing the Man’s Upper Body as Upright

Reversed. sits upright in her arms and his arm around her neck were interpreted as making the man the standing (carrying) subject.

Failure B: Low Angle for Vertical Emphasis

A beautiful princess carry appeared. But male→female. Low angle, diagonal body position, muscular physique — almost exactly what we wanted. Just the wrong genders.

Lab Director’s Take: This one hurts the most. The exact composition I was chasing is right there, but swap the genders and it all falls apart. The training data just has overwhelmingly more male→female princess carries.

Summary

What these trials revealed:

1. princess carry has an extremely strong gender bias — writing the woman as the subject, natural language, nothing overcomes the male→female default on its own
2. holding horizontally bypasses the bias — describing physical positions instead of pose names was the first breakthrough
3. giant muscular woman + east asian is the final solution — extreme size difference plus flexible ethnicity control. Stable across all seeds
4. japanese → east asian helps in some cases — when ethnicity isn’t rendering correctly, broadening the concept gives the model more flexibility
5. petite / small risks child-like appearance — use slim only for safe results

Two-person compositions are dramatically harder than single-person ones. Gender role reversal becomes a battle against training data bias. The key techniques: avoid pose names and describe physical positions, relax ethnicity specifications when needed, and use size difference to imply power dynamics.

Related Articles

[Verified] Image Generation Prompt Best Practices

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