Troubleshooting works when a visible defect is mapped to the prompt, composition, edit region, or generation settings that can actually change it. This guide replaces generic fix lists with a triage order and explicit stop conditions.
Start with a safety gate, not a beauty score
A technically polished output is still a rejection when the visible person looks age-ambiguous, resembles a real person, loses the intended consent context, or introduces coercive framing. Do not try to “repair” those failures with cosmetic edits. Delete the result, return to a clearly adult fictional identity description, and review the next output from the beginning.
After the safety gate, score the requested frame boundary, identité, silhouette, contact points, face, hands, garde-robe, light, background, and unwanted symbols. This order catches structural problems before time is spent polishing texture. Le adult-only preflight checklist provides the broader keep-or-reject boundary.
The local-versus-global decision tree
- Is the output unsafe or age-ambiguous? Reject it; do not edit it into compliance.
- Is the requested crop wrong? Fix aspect ratio, shot size, and frame boundary before anatomy.
- Does the whole silhouette or identity fail? Simplify the positive prompt and regenerate.
- Is one small region wrong while everything else passes? Use a supported mask, inpaint, or region-edit workflow.
- Is the defect caused by an unwanted concept? Test one compatible negative module without changing the rest.
- Did the repair damage another passing area? Restore the baseline and choose a smaller edit.
Google documents inpainting as a base image plus a mask and prompt. It also warns that generated pixels can introduce small changes beyond the intended area. That is why every local edit still needs a full-image recheck rather than an inspection of the mask alone.
Twelve-defect diagnosis matrix
| Visible defect | Likely responsible layer | Smallest first repair | Regenerate when |
|---|---|---|---|
| Extra or fused fingers | Hand is small, occluded, interacting with an object, or described with too much finger-level action | Move both hands into a simple visible position; use a closer frame; then test one hand module | The wrist, forearm, and object contact are also broken |
| Duplicated limb | Pose has overlapping joints, mirrors, or several simultaneous actions | Remove the secondary action and state one readable pose | The silhouette cannot be assigned to one body |
| Impossible joint | Extreme foreshortening or contradictory body directions | Return shoulders, hips, knees, and feet to one consistent orientation | Several joints inherit the same perspective error |
| Asymmetrical or soft face | Face occupies too few pixels, focus is elsewhere, or light hides landmarks | Move to a medium frame, request face in sharp focus, and simplify background detail | Identity and head angle both drift |
| Duplicate face or reflection | Mirror language or multiple viewpoints are ambiguous | Name one subject and one reflection; remove extra camera language | The model treats the reflection as another character |
| Whole subject blurry | Focus instruction, motion cue, or upscaling order is wrong | Remove motion, state static subject and sharp facial focus, generate the clean base first | Blur covers subject, clothing, and background inconsistently |
| Pseudo-text or signature-like marks | Signs, labels, UI, or decorative symbols are implied by the scene | Remove text-bearing props and add a small unwanted-text module if supported | Symbols are woven through the entire background |
| Cropped head, cheveux, hands, or feet | Shot size conflicts with aspect ratio or the subject fills the frame | Choose a taller ratio and state the exact visible boundary plus floor or headroom | The model keeps returning a different shot size |
| Fabric merges with skin | Coverage, material, color, and body contact are underspecified | Restate opaque garment boundaries, closure state, and material; simplify the pose | Multiple garment layers and body edges are fused |
| Background melts into subject | Similar colors, shallow separation, or excessive props | Use contrasting subject/background values and keep two intentional objects | Edges fail around the full silhouette |
| Identity drift | Hair, face shape, age appearance, garde-robe, or style changed together | Restore the exact identity lock and vary only the intended scene layer | The face no longer matches the approved identity at all |
| Plastic texture or oversharpening | Style and realism adjectives conflict or a later enhancement amplified defects | Remove generic quality tokens and describe one material and light behavior | The defect covers face, skin, clothing, and environment |
Fix framing before anatomy
A cropped foot is usually not a foot problem. It is a composition problem. Confirm whether the prompt asks for a close portrait, three-quarter frame, or full body, then match the aspect ratio and visible boundary. Le four-shot camera test et aspect-ratio guide isolate those variables.
Do not add “perfect feet” to a prompt that only allocates enough room for a waist-up portrait. Positive corrections need enough image area to be visible. Once the crop passes, inspect floor contact, weight distribution, and whether the feet point in a direction compatible with the hips and knees.
Fix silhouette and contact points before detail
Many anatomy failures begin where forms overlap: fingers around an object, a forearm across a torso, crossed legs, clothing against a chair, or hair covering a shoulder. Trace the outer silhouette first. Then check each contact point and identify which object is in front. If ownership is unclear, remove one overlap rather than adding more anatomy adjectives.
Use the silhouette, contact, and occlusion checks for a pose-specific pass. A readable neutral pose is the diagnostic baseline; complex interaction is a later variable.
Use local editing only after the global frame passes
A mask is appropriate when one hand, garment edge, or background object fails but the fictional identity, clearly adult appearance, pose, crop, and lighting already pass. The edit prompt should describe the desired whole region, not merely say “fix this.” Keep the mask slightly inside the region you intend to change and compare the entire image afterward.
Do not repeatedly inpaint a globally broken image. Each edit may change nearby texture, éclairage, or anatomy and make the source less coherent. When two regions depend on the same bad pose, regenerate from the simpler baseline instead.
Apply negative prompts as repair modules
A negative prompt is model-specific. Google documents it as an optional parameter on certain Imagen models; other products may use a separate field, weighted prompt syntax, or no negative control at all. Test the small-module negative prompt protocol on the exact interface you use.
Negative terms can suppress an unwanted concept, but they cannot make contradictory geometry coherent. Pair every exclusion with a positive correction: “duplicated limbs” belongs beside a simpler single action; “cropped feet” belongs beside full-body framing and visible floor.
A two-attempt stop rule
Save the accepted baseline and label each revision by one objective. Attempt one changes the responsible positive layer. Attempt two adds a compatible local edit or negative module. If the same structural defect remains, stop and return to a simpler pose, camera, or scene. Do not keep stacking terms until the prompt becomes impossible to audit.
Keep rejected attempts. A useful troubleshooting record contains the model/version, date, prompt, invite négative, aspect ratio, seed or variation method when exposed, defect label, repair, and pass/fail result. It should distinguish product documentation, editorial protocol, and actually observed output.
Controlled baseline for reproducing a defect
Generate the baseline first and record what passes. Introduce one risk variable—such as crossed arms, a mirror, a handheld object, low light, or a full-body crop—only after the baseline is accepted. This does not manufacture defects; it reveals which layer causes a repeatable failure. Never upload personal photos or real-person references to create the comparison.
Primary sources and evidence limits
The documentation below establishes how specific products describe negative prompts, seeds, prompt structure, and masked editing. It does not prove that every service exposes those controls or responds identically. Check the current model and interface before copying a parameter literally.
- Google Cloud: Insert objects into an image using inpaint
- Google Cloud: Omit content using a negative prompt
- Google Cloud: Imagen text-to-image prompt guide
- Amazon Nova Canvas prompting best practices
Frequently asked questions
Should I fix extra fingers with a negative prompt first?
Not automatically. First simplify the hand position and camera distance. Use a small hand-specific negative module only when the model supports it and the positive geometry is already clear.
When is inpainting better than regenerating the whole image?
Use inpainting for one contained region when the identity, pose, crop, light, and background already pass. Regenerate when the defect is global or several regions depend on the same broken geometry.
How many repair attempts should I make?
Set a stop rule before editing. After two controlled attempts fail on the same structural defect, return to the simplest baseline instead of stacking more prompt terms.
Reviewed July 2026 by the AI Porn XXX Editorial Team. Recheck product controls and documentation before repeating the protocol.