Skip to main content

​
Short answer

Distorted prints with Nano Banana Pro are not a model failure — they come from two stacked causes: inherent randomness in single-shot generation plus overly assertive prompts. Use this order:
  1. First reproduce with the same prompt + reference images on imagen.apiyi.com to rule out client-side issues
  2. Rewrite the prompt: drop absolute phrases like “strictly lock” or “pixel-perfect restore”; use concrete color / position / preserve-item descriptions instead
  3. Upload clear, well-defined reference images using Base64 encoding (Nano Banana series does NOT support OpenAI-style URL uploads)
  4. Enable multi-roll retries: have your business code auto-retry the same prompt 1~3 times
  5. If still no luck, switch channel or model: banana pro defaults to AI Studio — switch to the Vertex group, or try the gpt-image-2 series (whose edit style stays closer to the source)

​
Why prints get distorted

AI image generation is single-shot atomic sampling — every call is an independent draw, with no “strict restore” toggle. The two most common culprits:
  • Overly assertive prompt phrasing: phrases like “strictly lock in place”, “pixel-perfect restore”, or “1:1 reproduce” can make the model interpret the request as “redraw it”, and the print gets redrawn along with everything else
  • Wrong reference-image upload format: Nano Banana series only supports Base64-encoded uploads. Pasting a URL directly into the OpenAI-compatible image_url field makes the model “not see” the reference, so the print is free to drift
Even with a perfect description, the model will still “wander” in roughly 5%~15% of single-shot samples. This is not a defect — it is inherent to generative models: the same prompt naturally yields different results across calls.

​
Step-by-step troubleshooting

1

Reproduce in the test tool to rule out client issues

Open imagen.apiyi.com/#generate and run the exact same prompt + reference images again:
  • Print is also distorted in the tool → most likely a prompt issue, go to Step 2
  • Print is preserved in the tool → debug your integration (image actually uploaded? parameters correct?), see Generated image differs too much from reference
2

Rewrite the prompt, drop absolute phrasing

Replace command-style language like “strictly restore” or “pixel-perfect” with concrete attribute descriptions (see “Prompt optimization examples” below).
3

Make sure reference images are uploaded correctly

  • Nano Banana series does NOT support OpenAI-format URL uploads — you must use Base64
  • Reference images should be clear and feature-rich: blurry or cluttered references force the model to “guess”, making prints more likely to drift
  • Single image ≤ 7MB (Gemini’s official limit) — lossless compression recommended before upload
  • Each prompt accepts up to 14 reference images; if you only need to preserve one print, upload just that close-up and raise its weight
4

Enable multi-roll retries (auto retry)

In your business code, implement 1~3 automatic retries for the same prompt. One failed attempt doesn’t mean the model is broken — more rolls significantly raise the hit rate.
5

Switch channel or model as fallback

  • Switch to the Vertex group (just pick the Vertex group when editing the token in the console — no code change) — Vertex’s moderation differs from AI Studio’s, and some outfit-swap requests that get falsely flagged on AI Studio pass cleanly on Vertex
  • Or try the gpt-image-2 series — its edit style stays closer to the source image, with less color drift and style shift, ideal when “stay close to the original” is the priority

​
Prompt optimization examples

Below is a bad-vs-good comparison table for the outfit swap + print preservation scenario: Improved full prompt example:
Edit this image. Complete two things: ① Change the two teapots inside the red box to matte pure black, preserving the original material texture and shape; ② Delete the red box outline itself. The color, position, dimension labels, and text of all other objects in the image must remain completely unchanged.
General principle: change one category per call. If your outfit swap needs “change color + change background + add text”, split it into multiple edits — each call’s success rate will be noticeably higher than one mega-instruction.

​
About overall red / warm color tint

If you notice the whole image skews red or warm after the swap, that’s a different problem — its usual cause is banana pro edit results in overall red tint. The causes and fixes are different from print distortion (switch to Vertex channel or fall back to gpt-image-2).

​
Contact support

If the steps above still don’t resolve it, contact support in the workspace or email [email protected] with:
  • Reproduction result of the same prompt in imagen.apiyi.com (screenshot or link)
  • The model name you’re using (e.g. gemini-3-pro-image-preview / Nano Banana 2)
  • API call timestamp + request ID (if available)