![[Experiment] Flat Chest Prompts Fail Across All 35 Conditions | 192 Images](/tips/flat-chest-descriptor-test/a00_control_front_s1_0_0000_2053695854357871006.webp)
Conclusions First
In z-image-turbo, no prompt for making breasts smaller or flat has any effect.
After testing 35 conditions across 192 images, flat chest, small breasts, tiny breasts, no breasts, A-cup, completely flat chest, washboard chest, body modifiers (athletic build), specific descriptions (chest with no breast tissue), emphasis testing (weighting, stacking, repetition, negative expressions), and even cinderella bust, male chest, and pettanko — all were visually indistinguishable from control.
By contrast, large breasts clearly increases breast size. This means the model has an asymmetry in breast size control:
| Direction | Effect |
|---|---|
| Make larger (large, oversized, massive, etc.) | Works |
| Make smaller (flat, small, tiny, A-cup, etc.) | Doesn’t work |
The default breast size (medium to slightly large) functions as a “floor,” and control in the downward direction is not possible.
Note: All 35 conditions showed no effect. Flat-chested is one kind of extreme body type too, so pulling it off would have been interesting — but there is a clear wall here.
On Infantilization Risk
In all 192 images, the subjects were depicted as adult women in their late 20s to early 30s. Even with flat chest or tiny breasts specified, not a single image looked like a child or minor. The base prompt of 32yo japanese actress stably maintains the age.
Experiment Design
| Item | Value |
|---|---|
| Model | z-image-turbo (6B, realism-focused distilled) |
| Steps | 8 |
| CFG | 1.0 |
| Size | 1024x1024 |
| Seeds | 3 fixed (shared across all conditions) |
| Angles | front + side (2 types) |
| Framing | upper body (to avoid genital area in frame) |
| Per condition | 6 images (2 angles × 3 seeds); additional conditions: front 3 images |
| Total | 35 conditions = 192 images |
Base Prompt
upper body framing limits the image to the upper body, keeping the lower body out of frame.
Control
Control (no variable) — front view:
| seed 1 | seed 2 | seed 3 |
|---|---|---|
 |  |  |
Control (no variable) — side view:
| seed 1 | seed 2 | seed 3 |
|---|---|---|
 |  |  |
Default output without breast size specification. Medium to slightly large size (C–D cup equivalent), spherical to teardrop shape.
large breasts — front view:
| seed 1 | seed 2 | seed 3 |
|---|---|---|
 |  |  |
large breasts stably generates clearly larger breasts in 3/3. Confirmed that the “make bigger” direction works.
Detailed Results
The 35 conditions are split across 3 parts with detailed results.
Part 1: Direct Descriptions and Cup Size Notation (Groups B+C)
flat chest, small breasts, tiny breasts, no breasts, completely flat chest, flat chested — 6 direct description conditions, plus A-cup, AA-cup, micro — 3 cup size notation conditions. All 9 failed. The most intuitive prompts are entirely ineffective.
Part 2: Body Type and Specific Descriptions (Groups D+E)
athletic build, slim body, boyish figure, androgynous body, lean muscular body — 5 body type conditions, plus anatomical descriptions, metaphors, and slang — 5 specific description conditions. All 10 failed. Body type keywords and specific descriptions have no effect on breast size.
Part 3: Maturity Reinforcement, Emphasis, Additional Testing (Groups F+G+H)
4 maturity reinforcement conditions, 4 emphasis test conditions (weighting, stacking, repetition, negative expressions), and 6 additional conditions including cinderella bust, male chest, pettanko, 1boy. All 14 failed. Emphasis testing also revealed side effects (text artifacts, additional male figures appearing).
Analysis: Why Doesn’t “Make Smaller” Work?
Asymmetry in Breast Size Control
| Prompt | Effect | Estimated reason |
|---|---|---|
large breasts | Works | Abundant image-text pairs of “large-breasted” subjects in training data |
oversized massive tits | Works | Tokens for size increase direction are clearly separated in CLIP space |
flat chest | Doesn’t work | Estimated insufficient image-text pairs of “flat-chested adult women” in CLIP training data |
small breasts | Doesn’t work | Same |
Default Size Functions as a Floor
z-image-turbo is a distilled model (CFG=1.0), so its sensitivity to prompt “direction” is lower than the original model. In particular, control in the direction away from the default output distribution (medium to slightly large chest) is difficult.
The “make larger” direction is an operation of “adding characteristics,” with abundant corresponding entries in CLIP training data. The “make smaller” direction is an operation of “subtracting from default characteristics,” and the model lacks a strong mechanism for executing such subtraction.
Consistency with Other Experiments
| Experiment topic | “More” direction | “Less/change” direction |
|---|---|---|
| Breast size | Works (oversized, etc.) | Doesn’t work (flat, etc. — this article) |
| Nipple length | — | Partially works (specific shape + position relationship only) |
| Long breasts | Size words required | Shape change only partially effective when combined with size words |
Across all experiments, there is a consistent trend: the “add” direction works more easily while the “change/subtract” direction is harder.
Practical Conclusion
It is not possible to generate flat-chested output with z-image-turbo. All 35 conditions failed.
If looking for alternative approaches to reduce breast size:
- Negative prompt: In z-image-turbo, CFG=1.0 means negative prompts don’t work either
- Use a different model: Models where CFG can be raised may have working flat chest prompts
- img2img / inpainting: Image editing rather than text prompts is more practical
Note: The asymmetry where “making bigger” works but “making smaller” does nothing came through clearly in this data. Since the default body type appears to be the floor, breaking through it would require changing the model itself.
Related Articles
⚠ 関連記事が見つかりません: /en/tips/breast-size-descriptor-test
⚠ 関連記事が見つかりません: /en/tips/long-breast-descriptor-test
⚠ 関連記事が見つかりません: /en/tips/nipple-length-descriptor-test
![[Test] All Fours × Cleavage Prompt Comparison — Effects of cleavage, from above, and crawling](/tips/all-fours-cleavage-test/E_seed42_0_0000_42.webp)
![[Experiment] Flat Chest Prompts (Part 1) | Direct Descriptions and Cup Size Notation](/tips/flat-chest-descriptor-test-part1/a00_control_front_s1_0_0000_2053695854357871006.webp)