修复 text layout

Branch logic

• If the user has not provided image 使用 bad text, ask 用于 it before generating the final clean image.
• If a reference image is provided, preserve the subject identity, proportions, 和 important details before changing style 或 scene.
• If the first result fails 视觉 inspection, fix the smallest broken area first instead of regenerating the whole image.
• If the first tool path fails, switch to the next listed option: GPT Image 2, Recraft.
• If the task depends on masks, text, 或 object edges, inspect those details at full size before accepting the 输出.
• If the image will be 已发布 commercially 或 publicly, run the risk guardrails before export.

When not to use

• Do not use this as proof that the generated clean image is factually real 或 legally approved.
• Do not use it when the user needs professional legal, medical, safety, 或 regulated-compliance judgment.
• Do not publish the result until 权利, consent, authenticity, 和 platform constraints have been checked.

• The final asset clearly matches the requested clean image.
• The main subject still reflects the original image 使用 bad text.
• Text, hands, logos, edges, 和 产品细节 are not visibly broken.
• The result is usable in the target channel 且不 another full regeneration.
• Proofread all 可见 text, labels, logos, 和 small typography before using the final image.