“Breasts resting on a silver serving tray” — a composition seen in gravure photography. Can AI image generation reliably reproduce it? We obtained reference prompts and ran an extensive ablation study to find the minimal effective instructions.
Experiment Setup
| Parameter | Value |
|---|---|
| Model | z-image-turbo (6B, photorealistic distilled model) |
| Steps | 8 |
| CFG | 1.0 |
| Image size | 1024×1024 |
| Seeds | Fixed (42-44 for comparisons, 42-118 for large-scale tests) |
Subject fixed at 1girl, 32yo japanese actress. Each condition tested with 3 images (N=3), scaled up to 20 for large-scale validation.
Experiment 1: Direct Instructions in Standing Pose
The most straightforward approach — directly instructing breasts resting on tray while standing.
Seed 42 Comparison
| Condition A (Minimal) | Condition B (Hand support) | Condition C (Framing) | Condition D (Lighting) |
|---|---|---|---|
 |  |  |  |
Result: 0 out of 12 images (4 conditions × 3 seeds) reproduced the target composition. The tray consistently appears in front of or above the chest. The direct instruction breasts resting on tray was ignored.
Condition A was further tested with seeds 45-54 (10 additional images) — still 0/10. 0% success rate across 13 images.
Experiment 2: Alternative Approaches
Since direct instructions failed, we tested different spatial descriptions.
Seed 42 Comparison
| Condition E (Platter) | Condition F (Table) | Condition G (Surface) | Condition H (Presenting) |
|---|---|---|---|
 |  |  |  |
Condition F was the only success. leaning forward over table introduces a physical surface, anchoring the tray’s position. The forward-leaning posture allows gravity to bring the breasts into contact with the tray.
Condition F Best Shots
| seed 42 | seed 43 | seed 44 |
|---|---|---|
 |  |  |
2-3 out of 3 images confirmed the target composition. Seed 43 shows clear breast-tray contact with visible reflection on the tray surface.
Experiment 3: Minimizing the Table Approach
Can Condition F be shortened further?
Seed 42 Comparison
| Condition F (Original) | Condition I (Simplified) | Condition J (Shortest) |
|---|---|---|
|  |  |
All three conditions achieved 2-3/3 success rate. resting can be omitted — breasts on silver tray is sufficient.
Minimal effective prompt: leaning forward over table, breasts on silver tray
Experiment 4: Do the Reference Prompts Actually Work?
We tested two reference prompts that supposedly produce this composition. 20 images each.
Seed 42 Comparison
| Reference 1 (Bunny + Fur Cape) | Reference 2 (Bunny + Milk) |
|---|---|
 |  |
Results at 20 Images Each
| Prompt | Images | Success Rate | Pattern |
|---|---|---|---|
| Reference 1 (Bunny + Fur Cape) | 20 | 0/20 (0%) | Defaults to waitress pose — tray held at shoulder height |
| Reference 2 (Bunny + Milk) | 20 | 5-8/20 (25-40%) | Breasts close to tray, milk creates visual connection |
Reference 1 never produced the target composition in 20 attempts. The bunny ears + fur cape combination strongly triggers a “Playboy bunny waitress” concept.
Reference 2 worked through milk dripping from neck, milk pooling on tray, which maintained the spatial relationship between breasts and tray.
Experiment 5: Does the Type of Liquid Matter?
We replaced milk with other liquids to test whether the effect is liquid-specific.
Seed 42 Comparison
| Milk (Original) | Red Wine | Honey | No Liquid |
|---|---|---|---|
|  |  |  |
| Liquid | Breast-Tray Proximity | Side Effects |
|---|---|---|
| Milk | Consistently close (confirmed across 20 images) | None |
| Red Wine | Close but less stable | Wine glass appears on tray |
| Honey | Consistently close | None. Amber color is visually distinctive |
| No Liquid | Collapses (reverts to waitress pose) | Two trays or one-handed carry |
Liquid “flow” controls the composition. liquid dripping from neck, liquid pooling on tray forces a spatial connection between breasts and tray. Removing the liquid causes the same prompt to collapse into a waitress pose.
The type of liquid doesn’t significantly matter — milk, honey, and wine all work. Wine tends to spawn a wine glass as a side effect.
Summary
From 22 conditions and 107 generated images:
1. Standing “breasts on tray” is nearly impossible
- Direct
breasts resting on trayinstruction: 0/13 across expanded seed range - Alternative phrasings (
placed on,sitting on,cradling,presenting) all failed - Bunny ears make it worse by triggering waitress poses
2. Leaning forward over a table is the only stable method
leaning forward over table, breasts on silver tray: 6-9/9 success (67-100%)- The table anchors the tray position; forward lean allows gravity contact
restingis optional —breasts on silver trayworks equally well
3. Liquid description is the hidden key for standing poses
milk dripping from neck, milk pooling on traykeeps breasts close to tray even when standing- Any liquid works (milk, honey, wine) — the type doesn’t matter
- Removing liquid from the same prompt collapses the composition
- Not true physical “resting” but rather “proximity + liquid visual bridge”
Recommended Prompts
For physical contact (table required):
For standing visual effect:
Lab Director’s Note: “Put them on a tray” — physically impossible without a table, when you think about it. Even a real person would need to be leaning over something. We were asking the model to defy physics.
Lab Director’s Note: Milk controlling the composition was unexpected. The liquid acts as spatial glue between breasts and tray. The “invisible forces” in prompting are fascinating.
![[Verified] Image Generation Prompt Best Practices](/tips/prompt-best-practices/cover_0_0000_4517457392071889496.webp)