Key Takeaways
- Stairs function as “physical support” for the pose. Without location prompts, the body tilt remained small to moderate, but adding stair prompts clearly increased the backward lean (P1: 3/3, P2: 3/3)
leaning back, gazing upward dreamilyproduces a natural tilt. A dreamy atmosphere and gentle backward lean coexist, with upward gaze confirmed in 3/3 imagesarched back, looking up, head tilted backproduces a more dramatic arch. The neck and collarbone lines are emphasized, with a strong head tilt confirmed in 3/3 images- Simple
concrete outdoor stairsvs detailedold school building ... weathered surfaceshows no difference in pose reproduction. The difference lies in architectural detail (weathered surfaces, railings, building walls) - Without location prompts, the difference from P0 (no pose) tends to shrink. Without context for “what to lean against,” the model defaults toward a standard sitting pose
Experiment Design
We cross two axes — “leaning back and looking up” pose variations and outdoor stair location descriptions — in a matrix to observe independent effects and interactions.
Model & Parameters
| Parameter | Value |
|---|---|
| Model | z-image-turbo |
| Steps | 8 |
| CFG | 1.0 |
| Image Size | 1024×1024 |
| Seeds | 42, 123, 789 (3 fixed seeds) |
Base Prompt
{POSE} and {LOCATION} are swapped per condition.
Pose Axis
| ID | Condition | Prompt Diff |
|---|---|---|
| P0 | Control (sitting) | sitting |
| P1 | Lean back + gaze up | sitting, leaning back, gazing upward dreamily |
| P2 | Arched back emphasis | sitting, arched back, looking up, head tilted back |
Location Axis
| ID | Condition | Prompt Diff |
|---|---|---|
| L0 | Control (no location) | (removed) |
| L1 | Simple stair prompt | concrete outdoor stairs |
| L2 | Detailed stair prompt | old school building concrete outdoor stairs, weathered surface |
Total Images
3 poses × 3 locations × 3 seeds = 27 images
Results by Pose
P0: Control (sitting)
Pose prompt is sitting only. Observing how location differences affect the sitting pose.
P0 × L0: No Location
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: All 3 show an upright sitting pose. Backgrounds vary — s1 near a window, s2-s3 against concrete walls. No upward gaze in any image.
P0 × L1: Concrete Outdoor Stairs
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: 3/3 show concrete stairs with the subject sitting on them. Posture remains upright, similar to P0L0. The location prompt provides a “place to sit” but doesn’t alter the pose itself.
P0 × L2: Old School Building Stairs
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: 3/3 show old building stone stairs. Compared to L1, railings and building walls appear in the background, with visible weathered surfaces. Pose remains upright sitting, same as L1.
P1: Lean Back + Gaze Up (leaning back, gazing upward dreamily)
P1 × L0: No Location
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: 3/3 show upward gaze — a clear difference from P0. However, body tilt is small to moderate, with the subject typically leaning against a wall. gazing upward dreamily has a strong effect on gaze direction.
P1 × L1: Concrete Outdoor Stairs
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: 3/3 show stairs + upward gaze. Compared to P1L0, the body tilt is noticeably larger. The stair steps provide a “surface to lean against,” naturally amplifying the leaning back prompt. Hands placed on steps to support the body appear in multiple images.
Lab Director’s Take: The stairs giving the pose “physical justification” is genuinely interesting. You don’t even need to specify what she’s leaning against — the location fills in the gap automatically.
P1 × L2: Old School Building Stairs
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: 3/3 show old building stairs + upward gaze. Pose reproduction is comparable to L1. The difference is additional architectural detail (weathered surfaces, railings, building walls), which adds atmosphere. s1 shows a particularly strong body tilt, closest to the original prompt’s intent.
P2: Arched Back Emphasis (arched back, looking up, head tilted back)
P2 × L0: No Location
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: 3/3 show a strong upward gaze. Head tilt is clearly larger than P1L0, with emphasis on the neck line. head tilted back has a pronounced effect. However, backgrounds are highly inconsistent — ground, grass, bench — showing the instability of having no location prompt.
P2 × L1: Concrete Outdoor Stairs
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: 3/3 show stairs + strong backward lean. Body tilt is more consistently large compared to P2L0. The stair angle physically supports the back arch, producing even more dramatic compositions than P1L1.
P2 × L2: Old School Building Stairs
| seed 42 | seed 123 | seed 789 |
|---|---|---|
![]() | ![]() | ![]() |
Observation: 3/3 produce the most dramatic results of all conditions. The head tilts far back, maximally emphasizing the neck and collarbone lines. The old building stairs function as a “wide surface to lean against,” maximizing the back arch. s2 has a particularly extreme neck angle with strong photographic impact.
Lab Director’s Take: P2L2 s2 — that neck angle is wild. When
arched backhits this hard combined with old school stairs and an off-shoulder top, it’s just… chef’s kiss.
Interaction Summary
Pose Reproduction Matrix
Evaluation of backward lean + upward gaze reproduction by condition.
| L0 (no location) | L1 (concrete stairs) | L2 (old school stairs) | |
|---|---|---|---|
| P0 (sitting) | △ Upright sitting | △ Upright sitting | △ Upright sitting |
| P1 (leaning back, dreamily) | ○ Gaze up. Small-medium tilt | ◎ Gaze up + large tilt | ◎ Gaze up + large tilt |
| P2 (arched back, head tilted) | ○ Strong gaze up. Unstable BG | ◎ Strong arch + stairs | ◎ Maximum arch + atmosphere |
Key Interactions
1. Stairs provide “physical support” for the pose
Without location prompts (L0), the model auto-selects walls or ground for leaning back. With stair prompts, the steps consistently function as a “surface to lean against,” increasing body tilt. This is a positive interaction between pose and location prompts.
2. Without pose prompts, stairs only fix the sitting location
With P0 (sitting only), adding L0→L1→L2 didn’t change the upright posture. Stairs don’t “induce” a pose — they “amplify” an existing pose prompt.
3. Simple (L1) vs detailed (L2) shows no pose difference
concrete outdoor stairs (3 tokens) vs old school building concrete outdoor stairs, weathered surface (8 tokens) showed no notable difference in tilt angle or gaze direction. The extra 5 tokens in L2 affect architectural detail (weathering, railings, building walls) and atmosphere quality, but not pose reproduction.
Practical Recommendations
Recommended Prompts
For reproducing the “leaning back and looking up” composition:
Usage Guide
- Natural atmosphere → P1 (
leaning back, gazing upward dreamily) + L1 (concrete outdoor stairs). Token-efficient too - Dramatic composition → P2 (
arched back, looking up, head tilted back) + L2 (old school building ... weathered surface). Maximum neck/collarbone emphasis - Always include a location prompt → Without context for “what to lean against,” backward lean reproduction drops
Lab Director’s Take: “Lean back and look up” needs both the prompt AND a place to lean against. Stairs, bench, wall — anything works, but having that physical justification stabilizes the model’s output.



























![[Experiment] 6 Camera Angle Comparison — How Does a Poolside Scene Change with Low Angle, Bird's Eye, and Dutch Angle?](/tips/camera-angle-comparison/cover.webp)
![[Hands-On] Reproducing School Nostalgia × Cinematic Lighting](/tips/school-nostalgia-cinematic-prompt-test/E_seed42.webp)
![[Verified] Image Generation Prompt Best Practices](/tips/prompt-best-practices/cover_0_0000_4517457392071889496.webp)