Conclusions
coherent anatomy, correct hands and fingers has no visible effect in z-image-turbo.
In a 24-image comparison (2 scenes × with/without × 6 images each), no significant difference was confirmed in finger distortion rate, body balance, or overall quality.
Why It Has No Effect
The reason lies in how z-image-turbo works.
- z-image-turbo is a distilled model operating at CFG=1.0 — Prompt adherence is inherently limited, and subtle nuance instructions like
coherentandcorrecttend not to work coherent anatomyis an abstract concept — CLIP doesn’t understand “coherent (consistent) anatomy” as a specific image feature. Concrete pose specifications likehands on hipsare far more effective- Short 8-step inference — With few steps, small differences in conditioning are unlikely to be reflected in results
What Actually Works
What’s effective for preventing finger distortion is not abstract instructions like coherent anatomy, but stabilizing hands with specific poses:
| Technique | Example | Effect |
|---|---|---|
| Have them hold something | holding coffee cup, cotton candy in hand | Hand/finger shape fixed by the object |
| Place hand on body | hands on hips, chin resting on hand | Hand position becomes defined |
| Hide the hand | hands in pockets, arms behind back | Avoids the depiction altogether |
| Specific pose names | peace sign, waving | Defined poses with plenty of training data |
These have been demonstrated in the morning bed test with chin resting on hands and the summer festival test with cotton candy in hand.
Is It Safe to Delete from Prompts?
Safe to delete. The 5 words (~7 tokens) of coherent anatomy, correct hands and fingers are just wasting the 75-token limit. It’s more effective to use those tokens for specific pose or environment descriptions.
coherent anatomy, correct hands and fingers appears frequently in AI image generation prompts, intended to “maintain body consistency” and “correctly render hands and fingers.” But does it actually work in z-image-turbo?
Tested with 24 images.
Experiment Design
Two scenes, generating 6 images each with only the presence/absence of coherent anatomy, correct hands and fingers changed.
- Scene A: Hands on hips at poolside — a pose where hands are visible but relatively stable
- Scene B: Waving with fingers spread in a park — a distortion-prone pose with spread fingers
With 6 images per condition, we check whether there is a trend difference rather than a lucky one-time success or failure.
Scene A: Poolside × Hands on Hips
Without (6 images)
 |  |  |
 |  |  |
With (6 images)
 |  |  |
 |  |  |
Scene A Results
| Metric | Without (6 images) | With (6 images) |
|---|---|---|
| Finger distortion | 0–1 images | 0–1 images |
| Body balance | Generally good | Generally good |
| Overall quality | No difference | No difference |
No difference visible. Since hands on hips is inherently a stable pose, hand and finger rendering is stable regardless of coherent anatomy being present.
Scene B: Park × Waving (fingers spread)
Testing with a “waving with fingers spread” pose that is more prone to finger distortion.
Without (6 images)
 |  |  |
 |  |  |
With (6 images)
 |  |  |
 |  |  |
Scene B Results
| Metric | Without (6 images) | With (6 images) |
|---|---|---|
| Fingers appear as 5 | 4–5 images | 4–5 images |
| Finger fusion/disappearance | 1–2 images | 1–2 images |
| Body balance | Good | Good |
Again, no clear difference visible. Even with the waving with spread fingers pose, no significant difference in distortion rate was confirmed with or without coherent anatomy.
Summary
| Item | Conclusion |
|---|---|
| Effect in z-image-turbo | None (no significant difference in 24 images) |
| Token waste | ~7 tokens wasted |
| Recommendation | Delete and replace with specific pose specifications |
