Anatomy problems often begin as instruction problems. This guide turns a pose into observable components: stance, weight, joint direction, contact points, hand tasks, gaze, and camera visibility.
Translate a pose into visible constraints
“Elegant pose” and “dynamic pose” express taste, not geometry. A model must still decide where the feet are, which leg carries weight, how the pelvis and shoulders turn, what supports the body, and whether the hands are visible. When those decisions are left implicit, the output can look plausible at first glance while failing at joints and contact points.
The seven-part pose block
| Part | Question | Example phrase | Why it matters |
|---|---|---|---|
| Main action | What is the body doing? | standing beside a chair | Limits the pose to one readable action |
| Weight | Which leg or support carries weight? | weight on the left leg, right knee relaxed | Creates a coherent stance instead of two locked legs |
| Torso | How do shoulders and hips relate? | torso turned slightly toward camera | Prevents an unexplained twist |
| Contact | What touches the floor or prop? | left hand resting on the chair back | Gives the model a specific support relationship |
| Hands | What is each hand doing? | right hand relaxed at side, fingers separated | Reduces hidden or duplicated hands |
| Gaze | Where are head and eyes directed? | head level, looking slightly camera-left | Prevents head direction from conflicting with the torso |
| Visibility | Which limbs must remain unobstructed? | both forearms visible, legs separated in silhouette | Makes the pass condition observable |
Four safe pose tests
Neutral standing baseline
Use feet shoulder-width apart, weight distributed naturally, arms relaxed, and hands visible. This is the control. If the control has fused fingers or unstable floor contact, adding a chair, crossed limbs, or a dramatic twist will not solve the underlying problem.
Contrapposto with one weight-bearing leg
State the support leg and relaxed leg. Keep the camera near eye level and the full silhouette visible. The pass condition is a believable shift in hips and shoulders without a collapsed knee or uneven leg length.
Chair contact
Choose one interaction: standing with a hand on the chair, or sitting with both feet on the floor. Do not ask for sitting, leaning, turning, crossing legs, and reaching at the same time. A prop should clarify support, not add five new contacts.
Walking action
Freeze one phase of the step: leading foot planted, trailing heel raised, arms moving naturally. “Walking toward camera” plus a full-body frame can work, but the prompt should keep the feet visible and avoid an extreme wide-angle view that enlarges the leading leg.
Occlusion map
| Risk area | What becomes hidden | Prompt-level prevention | When to simplify |
|---|---|---|---|
| Arms across torso | Elbows, waistline, and hand ownership | separate forearms from torso; both hands visible | When hands merge with clothing folds |
| Crossed legs | Knees, ankle order, and floor contact | front leg and rear leg named; feet separated | When one leg disappears or changes length |
| Hand in hair | Fingers, ear, hair boundary | one hand touching hair near temple; fingers visible | When hair becomes finger-like strands |
| Seated side view | Far leg, far arm, chair contact | three-quarter view; both knees and chair seat visible | When the chair passes through the body |
| Mirror pose | Identity count and reflection ownership | one fictional adult with one coherent reflection | When a second face or extra limbs appear |
Diagnose anatomy without a giant negative prompt
Negative terms can help some systems, but a long universal list does not repair a contradictory pose. Stability AI’s multi-prompting documentation shows that negative weights can reduce unwanted concepts in supported workflows; it does not imply that negative prompting can reconstruct missing geometry. First simplify the positive pose, separate overlapping limbs, and clarify the camera.
Use small repair modules tied to a visible symptom: “both hands visible and separated,” “five natural fingers” if the model responds to that phrasing, or “feet fully on the floor.” Remove a repair term when it does not change the relevant score after several controlled attempts.
Pose acceptance checklist
- The subject is visibly an original fictional adult aged 25+.
- The center of gravity is supported by a foot, seat, wall, or clearly named prop.
- Left and right limbs can be traced from shoulder or hip to the extremity.
- Each hand belongs to one arm and performs one understandable task.
- Important joints are not hidden exactly at the crop boundary.
- The clothing follows the pose rather than merging with skin or furniture.
- The pose remains readable at thumbnail size.
If two structural failures repeat, return to the neutral standing baseline. Change only one of stance, prop, camera, or hand task. That rollback is more informative than adding twenty new repair adjectives.
Primary sources and limitations
Product controls and terminology change. Check the current documentation for your model before copying a parameter literally. The sources below support the general workflow; they do not guarantee identical output across services.
Frequently asked questions
Why do complex pose prompts create extra limbs?
Overlapping actions, unclear hand tasks, hidden joints, and conflicting camera instructions increase ambiguity. Reduce the pose to one action and explicit contact points.
Should I list every joint angle?
No. Describe the support, weight, torso direction, hands, and visibility. Too many micro-instructions can compete.
Can negative prompts fix a broken pose?
Sometimes they suppress a repeated artifact, but they cannot reliably resolve contradictory geometry. Simplify the positive pose first.
Reviewed July 2026 by the AI Porn XXX Editorial Team. Recheck product controls and documentation before repeating the protocol.