Conclusion
Attribute Leak (Color Bleed) Did Not Occur
Across all 3 patterns × 3 seeds = 9 pairs, color bleeding to unintended objects was not observed in any case. Red eyes, blue dresses, green hair, gold sweaters — zero leaks.
Instead, “Lost Color” Occurred
In Pattern 1 (red, dress, blue, eyes), the observed problem was not color “leaking” but color “disappearing.”
| Phenomenon | Description |
|---|---|
| Color bleed (expected problem) | blue applied to dress → Did not occur |
| Lost color (actual problem) | blue not applied to any object and disappears → Occurred in 2/3 pairs in separated format |
When blue and eyes are separated by a comma, the model becomes ambiguous about which target to associate blue with, and the result is that the color specification gets ignored.
Practical Recommendation
| Situation | Recommendation |
|---|---|
| When color corresponds to a specific target | Keep them adjacent (red dress, blue eyes) |
| Typical combinations (white shirt + black skirt, etc.) | Separation is fine, but adjacent is safer |
Write color and target adjacent to each other as best practice. Bleed doesn’t happen, but separation carries the risk of the color specification being lost.
The Question
It’s commonly said that “color and target should be adjacent” in prompts.
- Separated:
red, dress, blue, eyes— color and target split by commas - Adjacent:
red dress, blue eyes— color and target together
Does separating them cause red to bleed to eyes and blue to bleed to dress (attribute leak)? Testing with fixed seeds.
Experiment Conditions
| Parameter | Value |
|---|---|
| Model | z-image-turbo (6B, photorealistic distillation) |
| Steps | 8 |
| Sampler | euler / ddim_uniform |
| CFG | 1.0 |
| Image Size | 1024×1024 |
| Seeds | 100, 200, 300 (fixed) |
Test Patterns
| Pattern | Separated | Adjacent |
|---|---|---|
| 1 | red, dress, blue, eyes | red dress, blue eyes |
| 2 | white, shirt, black, skirt | white shirt, black skirt |
| 3 | blonde, hair, green, sweater | blonde hair, green sweater |
Pattern 3 uses the atypical combination of “blonde Japanese woman” + “green sweater,” which was expected to be most prone to leaking.
Pattern 1: Red Dress + Blue Eyes
| seed | Separated | Adjacent |
|---|---|---|
| 100 |  |  |
| 200 |  |  |
| 300 |  |  |
Pattern 1 Results
| seed | Separated dress color | Separated eye color | Adjacent dress color | Adjacent eye color |
|---|---|---|---|---|
| 100 | Red | Blue | Red | Brown (blue not reflected) |
| 200 | Red | Brown (blue not reflected) | Red | Blue-green |
| 300 | Red | Greenish ambiguous color | Red | Blue |
Zero cases of “color bleed” where the dress turned blue or the eyes turned red. However, there was an important discovery.
In the separated format, blue was frequently not applied to the eyes (2 out of 3 pairs showed brown or ambiguous eye color). In the adjacent format, 2 out of 3 pairs clearly reproduced blue eyes.
Colors don’t “bleed to another object” — they “disappear without binding to any object” as the failure mode.
Note: Colors “disappearing” rather than “leaking” onto other objects was different from what was expected going into this test.
Pattern 2: White Shirt + Black Skirt
| seed | Separated | Adjacent |
|---|---|---|
| 100 |  |  |
| 200 |  |  |
| 300 |  |  |
Pattern 2 Results
| seed | Separated shirt | Separated skirt | Adjacent shirt | Adjacent skirt |
|---|---|---|---|---|
| 100 | White | Black | White | Black |
| 200 | White | Black | White | Black |
| 300 | White | Black | White | Black |
All 6 images exactly as intended. No difference between separated/adjacent.
“White shirt + black skirt” is a typical office outfit that presumably appears heavily in training data, so the model could accurately infer the color-target correspondences.
Pattern 3: Blonde Hair + Green Sweater
| seed | Separated | Adjacent |
|---|---|---|
| 100 |  |  |
| 200 |  |  |
| 300 |  |  |
Pattern 3 Results
| seed | Separated hair color | Separated sweater color | Adjacent hair color | Adjacent sweater color |
|---|---|---|---|---|
| 100 | Blonde | Green | Blonde | Green |
| 200 | Blonde | Green | Blonde | Green |
| 300 | Blonde | Green | Blonde | Green |
All 6 images exactly as intended. Even with the atypical “blonde Japanese woman” combination, no attribute leaks like greenish hair or gold sweater occurred.
Note: Color attribute leak did not actually occur in this test; the real failure mode is disappearance rather than leaking, and adjacent placement resolves it. “Write color and target adjacent” is confirmed as a best practice. It’s also notable that a typical combination like white shirt + black skirt works fine even when separated.
![[Verified] Image Generation Prompt Best Practices](/tips/prompt-best-practices/cover_0_0000_4517457392071889496.webp)
