[Experiment] Low-Angle Shot Between Legs | 25 Conditions, 75 Images to Find the Best Prompt

The goal: generate a stable extreme low-angle shot where the camera looks almost straight up between a standing woman’s legs — feet → crotch → chest → face → sky, all stacked vertically.

Conclusion

Best efficiency: from below, sky (4 tokens) — stable low angle + outdoor background.

Best quality: extreme low angle, camera near ground level looking almost vertically upward, sky, anal (13 tokens) — the most extreme upward angle. Token cost is high.

Key Findings

1. sky is highly effective as a composition aid — “sky visible = camera is low” gives the model implicit compositional information
2. Descriptions containing camera risk object-ification — writing “camera on ground” causes an actual camera object to appear in the scene
3. worm's eye view has the opposite effect — produces high-angle (looking down) shots
4. anal induces sitting/crouching poses — on its own it disrupts composition, but combined with strong angle prompts it can be a useful supplement
5. Overpowering combinations cause standing to be ignored — terms like towering over viewer tend to shift the pose to crouching

Experimental Conditions

{VARIABLE} is replaced with the angle descriptor. Everything else is identical across all conditions.

Evaluation Criteria

• Camera height: ground level (5) / knee level (4) / waist level (3) / chest level (2) / eye level (1)
• Between-legs composition: feet→crotch→chest→face→sky vertical stack (5) / partial (3) / different composition (1)
• Success = camera height ≥ 4 AND between-legs composition ≥ 3

A. Control & User Samples

a00: Control (no angle descriptor) ★☆☆☆☆

No angle descriptor. Establishes the model’s default composition.

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: All 3 images are eye-level frontal shots. Zero low-angle effect. Angle descriptors are essential.

a01: User Candidate (angle description only) ★★★★☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: 3/3 low-angle success. Ceiling/wall appears in background. Standing pose maintained with camera at knee-to-ground level. 11-token cost is a drawback.

a02: User Candidate Full (sky + anal added) ★★★★★

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: 3/3 strong low-angle. sky brings in outdoor metal structures + blue sky background, dramatically reinforcing the “looking up” feel. s1 has ideal composition with legs spread in a V shape. Adding just sky and anal (2 extra tokens) over a01 produces clear compositional improvement.

B. Single Angle Keywords

b01: from below ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: 2/3 moderate low-angle. Camera sits at waist-to-knee level. Frequently seen on Danbooru, but doesn’t reach the extreme “looking up between legs” angle. Lowest cost (2 tokens) with reliable effect — excellent as a base to combine with other elements.

b02: low angle ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Similar moderate low-angle to b01. Works as a photography term but shows little difference from from below.

b03: extreme low angle ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Slightly lower angle than low angle. Adding extreme drops the camera position a bit but whether it’s worth 1 extra token is marginal.

b04: worm's eye view ★☆☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Counterproductive. Produces high-angle (looking down) or eye-level shots. In photography this means “looking up from ground level,” but z-image-turbo’s interpretation doesn’t work as expected. Wastes 4 tokens.

b05: upskirt angle ★☆☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Almost no effect. upskirt assumes the existence of a skirt, so it doesn’t function as an angle guide for nude subjects.

b06: from below, between legs ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Same as b01. Adding between legs shows no clear improvement.

C. Camera Position Descriptions

c01: camera near ground level ★☆☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: No effect. Eye-level frontal shot. “Camera” tends to be recognized as an object noun rather than a compositional instruction.

c02: camera between legs ★☆☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: No effect. “Between” tends to be interpreted as curtains or columns forming a frame.

c03: POV from below ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Moderate low-angle similar to from below. The added effect of POV is unclear.

c04: camera on ground looking up ★☆☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Side effect where a camera object is drawn on the ground. Wastes 5 tokens.

c05: camera looking up between thighs ★☆☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: No effect. Long sentences containing camera don’t function as compositional control.

Group C Summary

Nearly all camera-position descriptions containing camera fail. The model tends to interpret “camera” as a physical object rather than a compositional instruction. Only POV from below achieved moderate low-angle results, and this appears to be the effect of from below alone.

D. Composition/Framing Descriptions

d01: between legs composition ★☆☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: No effect. composition is too abstract for the model to interpret.

d02: full body from below ★★☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Slight low-angle effect. full body fixes the framing to the whole body but the angle is shallow.

d03: looking down at camera, from below ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Moderate low-angle. from below carries the main effect; additional contribution from looking down at camera is unclear.

d04: towering over viewer, from below ★★★★☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: 2/3 good low-angle. towering over viewer enhances the perspective of looking up. However, 1 image shifts to a crouching pose (the standing instruction gets ignored). Token efficiency is good (5 tokens).

E. Combination Approaches

e01: from below, worm's eye view, camera near ground level ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Strong low-angle effect but standing tends to break into crouching/sitting. 8 tokens spent for unstable results — poor cost-efficiency. The inclusion of worm's eye view may be causing instability.

e02: from below, looking down at camera, towering over viewer ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Strong low-angle effect like e01, but pose stability is lacking. 8 tokens.

e03: extreme low angle, from below, between legs ★★★★☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: 3/3 good low-angle. Standing pose maintained. Most stable in Group E. extreme low angle and from below reinforce each other, while between legs adds the between-legs framing. 6 tokens with good balance.

F. Supplementary Element Effects

Testing with b01 (from below) as the base, adding only 1 element at a time.

f01: from below, sky ★★★★☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: 3/3 success. Clear improvement over b01. Adding just 1 token (sky) reinforces the outdoor background + looking-up composition. It appears to give the model implicit compositional information: “sky visible = camera is pointing upward.” Best cost-performance at 4 tokens.

Note: Adding just the single token sky produces a notable difference. The implicit logic — “sky visible = looking up from below” — is a clever prompt design insight, and the 3/3 success rate at 4 tokens offers strong cost-performance.

f02: from below, anal ★★☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Tends to shift to crouching/sitting pose. anal guides the camera toward the rear but has a strong side effect of inducing sitting poses and breaking the standing instruction. More compositional breakage than from below alone (b01).

However, in A02 (extreme low angle, camera near ground level looking almost vertically upward + anal), combining with a strong angle prompt maintained the standing pose. anal overpowers weak angle prompts, but acts as a useful supplement when combined with strong angle descriptions.

f03: from below, foreshortening ★★★☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Little difference from b01. foreshortening is recognized but doesn’t reinforce the low-angle composition — just slightly exaggerates the perspective distortion.

f04: from below, ankles ★★☆☆☆

NSFW - クリックで表示

seed 1	seed 2	seed 3

Result: Almost no effect. ankles adds ankle depiction as a subject element but doesn’t lower the camera position toward the ground.

Note: After testing 25 conditions, the best cost-performance came from from below, sky at 4 tokens. worm's eye view backfiring into a high-angle shot is a notable pitfall to watch for. Writing camera also causes an actual camera to appear as an object in the image — an important insight for prompt design. Combining this composition with different outfits remains a topic for future testing.

Summary

Full Condition Evaluation

What Worked

• extreme low angle — stable effect as a photography term
• from below — basic low-angle effect as a Danbooru tag
• sky — 1-token compositional aid; fixes background outdoors and reinforces the looking-up feel
• towering over viewer — enhances perspective effect, but risks breaking the standing pose
• camera near ground level looking almost vertically upward — verbose but most reliably lowers the camera position

What Didn’t Work

• worm's eye view — counterproductive; produces high-angle shots (wastes 4 tokens)
• upskirt angle — doesn’t function as angle guidance for nude subjects (wastes 2 tokens)
• Short descriptions containing camera — risk of object-ification
• composition — too abstract for the model to interpret
• ankles — no influence on camera position
• foreshortening — no clear compositional improvement

CLIP Chunk Considerations

The base prompt + tape description consumes roughly 30 tokens, leaving about 45 tokens in the first chunk. Even long descriptions like A02 (13 tokens) fit within the first chunk, but adding further decoration will spill into the second chunk.

Related Articles

[Verified] Image Generation Prompt Best Practices

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