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

Conclusions First

• Static poses (standing, crouching, walking) showed zero distortion across all 3 images. Intended postures were reproduced stably.
• Among dynamic poses, jumping is most prone to distortion. 2 out of 3 images for jumping showed foot shape anomalies or shoe disappearance. Running had 0 distortion in 3/3 images.
• Looking back over shoulder reproduced well with no distortion confirmed.
• Running hand through hair had lower reproducibility. In 2 out of 3 images, the pose became “holding hair while running” rather than a stationary hair-tuck.
• Dynamic poses tend to shift clothing from the specified prompt (tank top, shorts) toward running-wear.
• [Deep Dive] Jumping distortion can be avoided by changing the wording. “leaping,” “hopping on one foot,” and “jumping rope” all produced zero limb distortion. Leaping in particular is recommended as a practical upgrade over jumping.

What You’ll Learn

• Whether static vs. dynamic poses differ in distortion rate
• How accurately prompts for intense movement (running, jumping) are reproduced
• Which conditions are more likely to cause limb shape anomalies or unnatural postures
• Stability differences by pose type

Experiment Design

Base Prompt

Only the {POSE} part is swapped per condition. Clothing, location, and framing are all fixed.

Evaluation Criteria

• Pose reproducibility: Is the specified movement depicted as intended?
• Limb distortion: Abnormal finger count, unnatural joint bending, limb fusion
• Postural naturalness: Is the center of gravity/balance physically plausible?
• Overall image quality: Any negative impact on elements other than pose (face, background, clothing)?

Results by Condition

A00: Control — standing

Using a stationary standing pose as the baseline.

seed=42	seed=123	seed=789

Observations

All 3 images show an upright frontal stance in a park setting. Both feet together, arms naturally at the sides. Full body is in frame, with no finger count anomalies or joint distortion in any of the 3. seed=42 has slightly unclear footwear, but no shape anomaly. All 3/3 successfully reproduced the “stationary standing pose,” functioning as a valid control.

A01: Running

seed=42	seed=123	seed=789

Observations

All 3 images show a sideways running pose with arms swinging front-to-back and legs in a wide stride. Hair flows behind, expressing movement. No limb distortion in 3/3. Joint bending is natural, with no unnatural weight distribution.

However, clothing shifted from the specified “tank top, shorts” toward running-wear (sports tank top + running shorts) in all 3 images. The “running” pose specification appears to influence clothing as well.

A02: Jumping — jumping in the air

seed=42	seed=123	seed=789

Observations

All 3 images show airborne jumping poses with arms spread upward or to the sides. The intent to jump was reproduced in 3/3.

However, this condition also had the most distortion. In seed=42, one leg is bent unnaturally large, with the knee joint angled in a physically impossible direction. In seed=789, the feet appear bare, with shoes gone. seed=123 had a relatively natural posture with no clear distortion. 2 out of 3 images had some form of shape anomaly — the highest distortion rate among all 7 conditions.

A03: Looking Back — looking back over shoulder

seed=42	seed=123	seed=789

Observations

All 3 images show the subject facing away from camera and looking back over the shoulder. The intended pose was accurately reproduced in 3/3. Upper body twist is natural, and no limb distortion was confirmed.

seed=42 shows a back-facing composition with the face visible over the shoulder. seed=123 shows a slightly side-angled turn. seed=789 has what appear to be bare feet, but there is no pose distortion. Overall, high reproducibility and a stable condition.

A04: Running Hand Through Hair

seed=42	seed=123	seed=789

Observations

All 3 images show a hand raised near the head, but instead of a stationary hair-tuck gesture, the composition reads as holding hair while moving. In seed=42 and seed=123, the pose resembles walking or running, suggesting the “running” in “running hand through hair” was interpreted as locomotion. seed=789 is even closer to a running pose.

No finger distortion was confirmed in 3/3. Even with the hand near the head, no shape anomalies appeared in the fingers. However, the intended “stationary hair-tuck” was only reproduced in approximately 1/3 images; 2 or more of the 3 became movement-while-moving poses.

Lab Director comment: The “running” in “running hand through hair” is probably getting read as an actual running action, right? “brushing hair back with hand” would probably give you a stationary pose instead. Kind of surprised the fingers came out clean though.

A05: Crouching

seed=42	seed=123	seed=789

Observations

A crouching pose was reproduced in all 3 images. Knees bent into a low position, with the crouching intent reflected in 3/3. No limb distortion confirmed.

seed=42 has the upper body slightly larger-framed with the feet close to the edge. seed=123 shows a full-body frontal crouch. seed=789 uses a lower angle with what appear to be bare feet. Overall a stable condition with good pose reproducibility and low distortion.

A06: Walking — walking towards camera

seed=42	seed=123	seed=789

Observations

All 3 images show the subject walking toward camera. One foot is forward with weight mid-transfer. No limb distortion in 3/3.

The difference from standing is subtle, but while standing shows both feet together in a stationary position, walking shows one foot forward with a visible stride. Clothing remained the specified tank top and shorts in all 3 — no clothing shift like what occurred in the running condition.

Cross-Condition Comparison (Same Seed)

All 7 conditions compared side by side at seed=42.

Condition	seed=42
A00: standing
A01: running
A02: jumping
A03: looking back
A04: hand through hair
A05: crouching
A06: walking

Summary

Distortion Rate Comparison

Key Findings

1. Jumping is the most distortion-prone dynamic pose. Possibly because determining full-body balance in mid-air is complex, 2 out of 3 images had foot shape anomalies or shoe disappearance. Running had 0 distortion in 3/3, showing a clear difference even within dynamic poses.
2. Static poses (standing, crouching, walking) are stable. All 9 images across the 3 conditions showed no limb distortion.
3. Looking back over shoulder has both high reproducibility and stability. The complex compound pose of turning back over the shoulder while facing away was accurately reproduced in 3/3.
4. “Running hand through hair” gets pulled toward locomotion. The word “running” is interpreted as a running action, and stationary hair-tuck reproducibility was around 1 out of 3. An alternative expression (e.g., “brushing hair back with hand”) is worth considering.
5. Dynamic poses tend to deviate from specified clothing. In the running condition, clothing shifted to running-wear in 3/3. The meaning of the pose influences clothing as well.
6. Even with the hand near the face, no finger distortion occurred. In the hand-through-hair condition, 3/3 images had no finger shape anomalies — proximity of the hand to the face was not a direct cause of distortion.

Deep Dive: Jump Wording Comparison

In the first-stage test, “jumping in the air” produced limb distortion in 2 out of 3 images — the highest distortion rate among all 7 conditions. Based on this, we ran an additional test to see if changing the jump wording could avoid distortion.

Additional Experiment Design

Base prompt and seeds (42, 123, 789) are identical to the first stage. Only the pose section was replaced with the following 4 conditions.

D00: leaping

seed=42	seed=123	seed=789

Observations

All 3 images show a dynamic leaping pose with limbs spread wide in the air. Like the jumping condition (A02), the subject is airborne, but with arms extended upward in a more dramatic composition. No unnatural joint bending or limb shape anomalies were confirmed in 3/3. seed=123 appears barefoot, but there is no shape anomaly. The knee joint breakage and shoe disappearance that occurred in A02 (jumping) were absent in leaping — 0 out of 3 images.

D01: hopping on one foot

seed=42	seed=123	seed=789

Observations

All 3 images show one foot raised with the other hopping. The single-foot jump intent was reproduced in 3/3. One knee is bent and lifted, clearly conveying the pose’s characteristic. No limb distortion in 3/3. A barefoot tendency was present but no shape anomalies. With one foot grounded, the balance may be easier to stabilize — the results show less distortion than jumping.

D02: mid-air, both feet off ground

seed=42	seed=123	seed=789

Observations

All 3 images show an airborne composition, but the sense of leap varies. seed=42 is close to the ground with weak floatiness. seed=123 shows a wide-legged airborne stance. seed=789 is a dramatic wide-armed leap. seed=42 has somewhat ambiguous foot depiction, though it can’t be definitively called a shape anomaly. 0 to 1 images may have mild anomalies — an improvement over A02 (jumping) at 2/3, but not fully stable either.

D03: jumping rope

seed=42	seed=123	seed=789

Observations

All 3 images show a jump-rope composition with both hands holding the rope and the subject lightly airborne. The jumping rope intent was accurately reproduced in 3/3. The rope is visible, and the pose works as a prop-inclusive scene. No limb distortion in 3/3. Shoes (sneakers) were maintained in 3/3 — the shoe disappearance seen in A02 (jumping) did not occur. The modest jump height and feet-together stance may contribute to the stability.

Deep Dive Cross-Condition Comparison (seed=42)

Condition	seed=42
A02: jumping in the air (first stage)
D00: leaping
D01: hopping on one foot
D02: mid-air, both feet off ground
D03: jumping rope

Deep Dive Summary

Findings

1. “Jumping in the air” distortion can be avoided by changing the wording. leaping, hopping on one foot, and jumping rope all had 0 limb distortion in 3 images.
2. “Jumping rope” is the most stable. Zero distortion, shoes maintained, and pose reproduction all good. Specifying a prop (the rope) constrains the posture, resulting in a more stable output.
3. “Leaping” is a practical upgrade over jumping. The same airborne action but dramatically lower distortion rate. Leaping is recommended as the first choice for jump-type poses.
4. “Mid-air” has variable floatiness. Specifying a state rather than an action results in varying degrees of jump intensity per seed.
5. Bare-foot tendency is common across all jump variants. Multiple barefoot images were found in D00–D02. Only jumping rope reliably maintained shoes.

Lab Director comment: Wait, leaping being 3/3 with no distortion when it means basically the same thing as jumping — honestly didn’t expect the output to change that much. And jumping rope being 3/3 stable too, probably because holding the rope anchors the posture. Looks like leaping should be the go-to for jump poses.

Related Articles

[Experiment] 6 Camera Angle Comparison — How Does a Poolside Scene Change with Low Angle, Bird's Eye, and Dutch Angle?

[Experiment] Scene Setting Prompts for Falling | 8 Conditions, 24 Images

[Verified] Image Generation Prompt Best Practices