How to use this page
The most frequently encountered issues, with solutions confirmed through actual testing. Scan the headings for your symptom, then apply the fix exactly as written — vague variations of these fixes are how the problems came back in our own testing.
Text / logo rendering
Text in generated images is distorted or garbled — AI models have real limits on precise text rendering. Simplification strategy: keep text to 8 characters or fewer, and add "clearly readable text, sharp letterforms" to your prompt. For brand-critical text, overlaying in Canva or Photoshop is more reliable.
English words are misspelled — you can emphasize spelling in the prompt: "spell exactly: P-R-O-D-U-C-T", or generate multiple times and pick the most accurate result. For important brand text, post-processing is the more stable long-term solution.
Character / subject consistency
The same character looks different across multiple images — option 1: keep the same reference image and the same core description across the whole batch instead of rephrasing between runs. Option 2: upload your first satisfactory result as the reference image for subsequent generations. Option 3: describe the character's appearance in detail every time (hair color, hairstyle, facial features, clothing). NB2 supports character consistency across up to 5 images.
Clothing or accessories change between batch generations — repeat the clothing description on every generation; you can't just say "same outfit as before." Models don't have true cross-generation memory. Write "wearing the same [color + style + specific details] as before."
Content safety restrictions
Real celebrity image generation is rejected — this is Google's content safety policy and cannot be bypassed. Use role descriptions instead of names: "a woman in her 30s, professional businesswoman, dark hair, confident expression."
NB2 rejects financial information modifications — since February 2026, NB2 has a dramatically higher rejection rate for requests involving invoices, ID documents, price tags, and similar financial content. Handle these through other means (e.g., edit directly in image editing software).
Video quality issues
Motion in video is unnatural or shaky — the most effective fix is simplifying the scene (simpler scenes produce smoother motion) and explicitly specifying camera movement. Vague descriptions like "show the product" produce unpredictable random motion. Use specific terms like "slow push-in" or "static camera."
Veo 3.1 audio doesn't match expectations — more specific audio descriptions help significantly: "soft piano with occasional string swell" beats "beautiful music," and "quiet café ambient with distant conversation" beats "café sounds." For content requiring precise audio, generate the video without audio and add it in post.
Seedance 2.0 generates videos with extra limbs or disappearing objects — a known issue in complex scenes (roughly 10% occurrence rate). Reduce the number of simultaneously moving objects, simplify subject actions, and keep prompts under 60 words. If it keeps happening, try Seedance 1.5 — slightly lower quality but more stable.
Other issues
Output style drifts across a batch — keep your core style description identical line to line, save your first successful prompt as a "seed" to build from, and don't rely on memory between runs to maintain style consistency.
For more prompt optimization techniques see the prompt-writing habits guide; for model-specific usage, see the individual model guides.