We are experiencing a reproducible image-generation issue in which part of the main subject is duplicated without being requested in the prompt.
In some generated outputs, the model renders a second, unintended partial copy of the subject. In the attached example, the upper section of a shoe is shown twice: once in the correct main position and once again as an additional repeated partial rendering in the upper area of the image.
This duplicate content is not requested in the prompt and appears to be an unintended generation artifact rather than a creative variation.
We have observed this behavior not only through Replicate, but also through a Vertex AI API access path. This suggests that the issue may originate in the underlying model or image-generation pipeline rather than being specific to one provider integration.
The issue affects:
- Gemini 3 Pro Image
- Gemini 3.1 Flash Image
We started noticing this behavior after the transition from the previous Preview version to the current Image version.
Expected result
The model should generate a single coherent subject only, with no duplicated or repeated partial subject content unless explicitly requested in the prompt.
Actual result
A part of the subject is unintentionally repeated. In the attached example, the shoe upper is shown a second time in the image, even though the prompt does not ask for multiple shoes or duplicated product parts.
Conditions under which the issue occurs
- Access path / provider:
- Replicate
- Vertex AI API
- Affected model family:
- Gemini 3 Pro Image
- Gemini 3.1 Flash Image
- Input type: image-to-image generation
- Prompt type: single-product prompt
- Resolution: observed at 2K, and may also occur at higher resolutions
- Output format: occurs regardless of whether the output is set to PNG or JPG/JPEG
- No instruction is given to duplicate, repeat, mirror, or extend any part of the subject
Example prompt
create a new design of this shoe, plain white background
Reproduction code
import Replicate from "replicate";
const replicate = new Replicate();
const output = await replicate.run(
"google/nano-banana-pro",
{
input: {
allow_fallback_model: false,
aspect_ratio: "1:1",
image_input: [
"https://api.replicate.com/v1/files/MTFmYzdmYjEtMDI0NC00NmFmLWI4ZmEtYjA4M2JhMTQ3NDJl.jpg"
],
output_format: "png",
prompt: "create a new design of this shoe, plain white background",
resolution: "2K",
safety_filter_level: "block_only_high"
}
}
);
console.log(output);
Steps to reproduce
Use a single product reference image.
2. Run image-to-image generation through Replicate or Vertex AI API.
3. Use one of the affected image models.
4. Request a single redesigned product on a plain background.
5. Inspect the generated image.
6. In some cases, part of the subject appears a second time without instruction.
Additional observations
- The issue is reproducible across multiple access paths (Replicate and Vertex AI API).
- It affects Gemini 3 Pro Image and Gemini 3.1 Flash Image.
- The issue occurs even though only one single subject is requested.
- The model sometimes duplicates only a partial section of the object instead of generating a full second object.
- In the attached example, the upper part of the shoe is repeated.
- The issue occurs regardless of whether the output format is PNG or JPG/JPEG.
- We have observed this behavior at 2K resolution, so it is not limited to 4K-only rendering.
Regression information
We began noticing this issue after the transition from the previous Preview version to the current Image version.
Request
Please investigate whether this is caused by a subject duplication / partial replication bug in the image-generation or rendering pipeline.
Because the same behavior occurs through both Replicate and Vertex AI API, the root cause may be upstream in the underlying model or image-generation system.
