My last couple of posts have gotten no support from Google Staff, so this is a Hail Mary.
This happens about 20% of the time. The reference images are completely ignored, and the prompt is largely ignored too.
Prompt:
## Illustration Instructions:
- Create a full-bleed printable page for a children's picture book.
- CRITICAL: The illustration must extend edge-to-edge with NO vignette effects
## Art Style
Warm, textured watercolour per the input images
## Character Reference Images
### INSTRUCTION:
- Maintain each character's appearance EXACTLY from the provided images.
### Input images
Mel: A 3 year old girl.
Dad: A 31 year old man.
Mum: A 31 year old woman.
## Scene Description
Interior of an airplane cabin. Mel sits in a window seat, pressing her face to the window with delight. Dad sits beside her smiling. Mum leans over from the aisle seat. Through the window, fluffy white clouds are visible. Soft cabin lighting. Close-up shot focusing on their row.
Reference images:
Using the Files API to add these files, setting display_name to the character names (Mel, Dad, Mum).
Generated 5 outputs from the exact same prompt and images. First one is just wildly inconsistent, the rest are fine. Is this a bug? Something with my prompting?
I use this: [Analyze image inputs strictly for the established Thick impasto brush strokes oil painting art style, color palette, and historical atmosphere. RESET SCENE GEOMETRY completely. Do not edit or add to the previous image’s composition. Create a completely fresh framing and camera perspective to depict the specific action described.] + your_prompt
I don’t describe what the images are, just what i want to see. Most of the time it accurately understands what to do. So your_promptshould be something like “Interior of an airplane cabin. a small girl sits in a window seat, pressing her face to the window with delight. Dad sits beside her smiling. Mum leans over from the aisle seat. Through the window, fluffy white clouds are visible. Soft cabin lighting. Close-up shot focusing on their row.“
I have tested your prompt with the reference images provided by you using Nano Banana Pro on my end. I have generated 5 output images. I am attaching the generated images for your reference.
Thanks @doneforaiur - your advice got me into something that worked. The model needed a bit of nudging upfront to just “think” through and establish the core rules before firing off into the prompt.
@Sonali_Kumari1 as I mentioned, the issue was occurring 20% of the time, so it was likely that after generating 5 images it wouldn’t appear.
@Sonali_Kumari1 after playing with this for weeks, I am now convinced that the model is just outright ignoring the input images 10-20% of the time. The issue above occurs over sufficient attempts, where reference images with the API (I’m using the files interface) are outright ignored.
Hi, I have the same issue. I’m using the AI Studio API (and I also tested Vertex AI, with the same result) in a production pipeline, and we are seeing the same behavior. Gemini 3 Pro Image is an excellent model for image generation, but something in Google’s pipeline seems to be unreliable, and images often do not arrive.
Because of this, we have had to add additional QA steps on our side. We run a check with Gemini Flash to confirm whether reference images were actually used. Otherwise, we would too often deliver incorrect outputs that do not use the intended assets.
I hope this gets fixed soon. It is very frustrating and it is costing us a lot of money. Have you found any workaround or fix? I would really appreciate hearing what worked for you.
It sounds like you’re now using a different prompt / process than one you noted at the top of the thread. If you’d like us advise more, could you please provide the updated prompt / example images?
My line of thinking is that, dependent somewhat on the temperature setting, any model will always deviate a little from one generation to the next. It could be an adjustment to the temperature, or just a more strict prompt is required.
That said, since generative models are stochastic, it’s impossible to remove chance / surprise entirely from the generation process.
My approach would usually be to create a prompt / settings which reduce the undesired variation to a minimum. I’d achieve this by creating a testing pipeline which can take a prompt / settings under test, & automatically perform x (e.g. 100) generations. I’d then iterate on the prompt / settings until I got as close as possible to the consistent output I wanted.
John,
This issue does not result in a subtle temperature variation. It’s a complete ignore of the reference images in about 10-20% of image generations through the API. I am experiencing the same issue and there seems to be no prompt change that can fix it. I have had to add a “regenerate” button to the resulted output for the user. When the same prompt with the same reference images is resent, the 2nd result is correct nearly 100% of the time
