Gemini multi image capability seems to be unstable and fail very basic question

[ChenChengKuan]

Description of the bug:

Hello
I followed the sample code to analyze multiple images in the official doc

One weird thing I found is that the client cannot correctly identify the number of images uploaded so it only give the response on certain image

Followings are the images I test (I think any image should be able to reproduce):

fire
lighthouse
tree

img1 = Image.open("fire.jpg")
img2 = Image.open("lighthouse.jpg")
img3 = Image.open("tree.jpg")

MODEL_NAME = 'gemini-2.0-flash-001'
client = genai.Client(api_key="My api key")
response = client.models.generate_content(model=MODEL_NAME,
                                          contents=['Describe each image', img1, img2, img3])
print(response.text)

The output only contains the last image:
The image shows a single, large tree standing in a green field against a bright blue sky with scattered white clouds. The tree has a full, round crown of green leaves. The trunk of the tree is thick and gray. The field is a vibrant green, and the sky is a clear blue with wispy white clouds. The image is well-lit and has a clear, crisp focus.

When I further ask how many images do you see

response = client.models.generate_content(model=MODEL_NAME,
                                          contents=['How many images do you see?', img1, img2, img3])
print(response.text)
I see one image.

One workaround I found is to add system prompt to force it consider all images upload before answer, but this feel very unnatural

response = client.models.generate_content(model=MODEL_NAME,config=types.GenerateContentConfig(system_instruction="Consider all images uploaded by users before answering any question"),
                                          contents=['Describe each image', img1,img2,img3])

The output
Here are the descriptions of each image:

1. A bonfire that appears to be made out of branches is ablaze, with flames reaching high into the sky.
2. A lighthouse stands on a shore by the ocean, its light shining across the water under a purple sky with the setting sun.
3. A large green tree stands tall in a field of grass under a bright blue sky with fluffy white clouds.

When asking the number of the image uploaded:

response = client.models.generate_content(model=MODEL_NAME,config=types.GenerateContentConfig(system_instruction="Consider all images uploaded by users before answering any question"),
                                          contents=['How many images do you see?', img1,img2,img3])
print(response.text)
I see 3 images.

Actual vs expected behavior:

I think it should know how many images to analyze without using further prompting? It's feel awkward

Any other information you'd like to share?

What I have tried:

1. I found appending the correct number of image tag (<image>) at the end of question helps but not as stable as system prompt
2. Sometimes, when asked about how many images, it will add the cropped ones. It answer something like I saw one image and X cropped tiles or just give me the numbers of original + cropped (In this case, it's 3+15=18), just weird from user perspective, this is a bit hard to trigger but you can get one when trying multiple times

[Swini27]

Running the same command I am able to get response for any number of images.

img1 = Image.open("burj-khalifa1.jpg")
img2 = Image.open("burj-khalifa2.jpg")


MODEL_NAME = 'gemini-2.0-flash-001'
client = genai.Client(api_key= api)

response = client.models.generate_content(model=MODEL_NAME,
                                          contents=['Describe each image', img1, img2])
print(response.text)

Response

Here are the descriptions of the images:

**Image 1:**

*   **Subject:** An aerial view of the Dubai skyline at night.
*   **Features:** The Burj Khalifa is prominently featured, towering above the other skyscrapers. The city is brightly illuminated, with lights from buildings, streets, and circular road systems creating intricate patterns. The sky is a deep blue/purple at dusk.
*   **Overall Impression:** A dramatic and captivating view of a modern, vibrant city.

**Image 2:**

*   **Subject:** An aerial view of the Dubai skyline during sunset.
*   **Features:** The Burj Khalifa is visible in the haze, along with many other buildings. The sky is a bright orange, and most of the sky is tinted with the colors of the sunset.
*   **Overall Impression:** A calm and pretty aerial shot of Dubai.

Are you able to share some more information or defined steps to reproduce the issue.

[ChenChengKuan]

That's interesting, may I ask your genai python sdk version? (mine is 1.3.0). Btw way, I just found a interesting modification of my code and suddenly everything becomes normal.

Here is my modification (just put a list of image instead of feeding them like img1,img2,img3 as shown in official doc)

img1 = Image.open("fire.jpg")
img2 = Image.open("lighthouse.jpg")
img3 = Image.open("tree.jpg")

MODEL_NAME = 'gemini-2.0-flash-001'
client = genai.Client(api_key="My api key")
response = client.models.generate_content(model=MODEL_NAME,
                                          contents=['Describe each image', [img1, img2, img3]])
print(response.text)

Output

Here are the descriptions of each image:

**Image 1:**

This image shows a large bonfire against a dark night sky. The flames are bright orange and yellow, engulfing piles of wood and branches. The fire appears to be quite intense, with the flames reaching high into the frame.

**Image 2:**

This image shows a lighthouse at dusk/dawn. The sky is a beautiful blend of purple, pink, and orange hues, reflecting in the calm water. The lighthouse stands tall and is silhouetted against the colorful sky. The light from the top of the lighthouse shines brightly across the water. 

**Image 3:**

This image is a landscape shot featuring a single, large tree standing in a grassy field under a blue sky. The tree has a full, green canopy and is the focal point of the image. The sky has some fluffy white clouds, and the overall impression is one of tranquility and natural beauty.

[Swini27]

Currently i am on

Name: google-genai
Version: 1.5.0

The modification gets the work done, but it should work the other way too.

Edit- by mistake I wrote the version of some other venv , the code snippet in context was executed on V1.5.0 so, it should work for you if you upgrade the SDK

[ChenChengKuan]

Thanks for help @Swini27 It works well now!