Diagnose the failed region before spending another generation.

Mark the problem as anatomy, composition, identity, texture, or artifact. Extra fingers belong to anatomy. A hand cut by the frame belongs to composition. A changed face belongs to identity. Text or a watermark belongs to artifacts. Each class points toward a different correction.

Do not rewrite the full prompt after spotting one defect. Keep the parts that worked and change the control linked to the failed class. A smaller revision preserves evidence from the successful regions and makes the next output easier to compare.

Hands touching faces, crossed limbs, intertwined characters, reflective surfaces, and heavy occlusion require the model to resolve several boundaries. Simplify the action, widen the crop, or remove one prop. Restore complexity after the basic silhouette works.

For multiple subjects, describe each position and role with clear separation. A crowded prompt that alternates traits can mix identities and limbs. Start with one fictional adult subject when testing a new model or prompt structure.

A full-body request needs a canvas that can hold the figure. Portrait ratios provide vertical room, while a square crop can force limbs toward the edge. Name the shot and place the subject inside the frame with terms such as full figure, visible hands, or space above and below.

Composition instructions can conflict. A close-up and visible full body cannot occupy the same frame. Choose the information the viewer needs, then remove camera terms that ask for another crop.

A readable pose reference can guide skeleton and broad placement when text fails to describe the body arrangement. Choose a permitted source with clear limbs and limited occlusion. Keep appearance and scene direction in the prompt.

Pose guidance does not guarantee correct joints, hands, or contact. Inspect those regions in each output. If the model repeats the same failure, choose a simpler source or a model with stronger support for the control.

Add a short exclusion after you observe a repeat pattern, such as duplicate limbs, fused fingers, text, or watermark. Long copied lists can suppress desired details and hide the cause of a defect. Keep model-specific presets and remove terms without a named job.

Negative prompts influence learned patterns. They cannot solve a physically unclear action or contradictory frame. Pair them with a simpler positive prompt and structural guidance when the problem involves body arrangement.

Decide how many attempts the image deserves before changing the model, source, or composition. Repeating the same setup consumes tokens without testing a new idea. A stopping rule protects the budget and turns failure into a decision.

Save the closest output and the settings that produced it. You may return with a better reference or model. Do not publish a selected success as proof of general anatomy performance; keep the full sample when you make a product or comparison claim.

Reduce the image mentally to a dark silhouette. Check whether the head, torso, pelvis, limbs, hands, and major props separate clearly. If two limbs merge or an arm has no readable path from shoulder to hand, surface detail will not repair the structure. Regenerate with a clearer pose, less overlap, or a camera angle that exposes the joints.

Use a wider crop when the frame cuts through hands, feet, or active limbs. The model needs canvas space to resolve the complete form. A portrait crop can hide difficult anatomy intentionally, but it should not pretend to be a full-body result. Match the framing to the part of the body the project needs to show.

Hands touching faces, clothing, furniture, or another body create several overlapping constraints. Props add grip, scale, and perspective demands. If an interaction fails, remove the least important object, use one visible hand, or separate the subjects. Establish a stable pose before restoring secondary contact.

For two-character scenes, define each fictional adult separately and state the visible relationship in simple spatial terms such as standing side by side or facing each other. Avoid assigning several actions to both characters in one sentence. Test a neutral arrangement first, then increase motion while preserving the camera and character anchors.

Repeatedly regenerating a nearly correct image can change the face, palette, background, and composition while attempting to fix one hand. If the output is permitted and the defect is local, a rights-cleared editing workflow may be more controlled than a full rerun. Keep the original and record any manual or AI-assisted edits.

Do not hide serious structural failures with cropping when the intended use requires the complete pose. Review at full size and on the final display format. Text, fingers, reflections, jewelry, and background figures often reveal defects that a small preview conceals. Publish only after the actual export passes the same checks as the preview.

Tag failed outputs with a short cause such as hand count, limb merge, joint angle, foreshortening, crop, reflection, or multi-character overlap. Add the model, ratio, pose language, and reference state. After several sessions, the log shows which compositions repeatedly create trouble.

Turn recurring failures into production rules. You might reserve wider canvases for active full-body poses, avoid multiple hand props in the first pass, or use a pose reference for a specific camera angle. Review the rules when a model changes so old limitations do not become permanent superstition.