# Image understanding Source: --- Gemini models are built to be multimodal from the ground up, unlocking a wide range of image processing and computer vision tasks including but not limited to image captioning, classification, and visual question answering without having to train specialized ML models. **Tip:** In addition to their general multimodal capabilities, Gemini models (2.0 and newer) offer **improved accuracy** for specific use cases like object detection and segmentation, through additional training. See the Capabilities section for more details. ## Passing images to Gemini You can provide images as input to Gemini using two methods: * Passing inline image data: Ideal for smaller files (total request size less than 20MB, including prompts). * Uploading images using the File API: Recommended for larger files or for reusing images across multiple requests. ### Passing inline image data You can pass inline image data in the request to `generateContent`. You can provide image data as Base64 encoded strings or by reading local files directly (depending on the language). The following example shows how to read an image from a local file and pass it to `generateContent` API for processing. from google.genai import types with open('path/to/small-sample.jpg', 'rb') as f: image_bytes = f.read() response = client.models.generate_content( model='gemini-2.5-flash', contents=[ types.Part.from_bytes( data=image_bytes, mime_type='image/jpeg', ), 'Caption this image.' ] ) print(response.text) You can also fetch an image from a URL, convert it to bytes, and pass it to `generateContent` as shown in the following examples. from google import genai from google.genai import types import requests image_path = "https://goo.gle/instrument-img" image_bytes = requests.get(image_path).content image = types.Part.from_bytes( data=image_bytes, mime_type="image/jpeg" ) client = genai.Client() response = client.models.generate_content( model="gemini-2.5-flash", contents=["What is this image?", image], ) print(response.text) **Note:** Inline image data limits your total request size (text prompts, system instructions, and inline bytes) to 20MB. For larger requests, upload image files using the File API. Files API is also more efficient for scenarios that use the same image repeatedly. ### Uploading images using the File API For large files or to be able to use the same image file repeatedly, use the Files API. The following code uploads an image file and then uses the file in a call to `generateContent`. See the [Files API guide](/gemini-api/docs/files) for more information and examples. from google import genai client = genai.Client() my_file = client.files.upload(file="path/to/sample.jpg") response = client.models.generate_content( model="gemini-2.5-flash", contents=[my_file, "Caption this image."], ) print(response.text) ## Prompting with multiple images You can provide multiple images in a single prompt by including multiple image `Part` objects in the `contents` array. These can be a mix of inline data (local files or URLs) and File API references. from google import genai from google.genai import types client = genai.Client() # Upload the first image image1_path = "path/to/image1.jpg" uploaded_file = client.files.upload(file=image1_path) # Prepare the second image as inline data image2_path = "path/to/image2.png" with open(image2_path, 'rb') as f: img2_bytes = f.read() # Create the prompt with text and multiple images response = client.models.generate_content( model="gemini-2.5-flash", contents=[ "What is different between these two images?", uploaded_file, # Use the uploaded file reference types.Part.from_bytes( data=img2_bytes, mime_type='image/png' ) ] ) print(response.text) ## Object detection From Gemini 2.0 onwards, models are further trained to detect objects in an image and get their bounding box coordinates. The coordinates, relative to image dimensions, scale to [0, 1000]. You need to descale these coordinates based on your original image size. from google import genai from google.genai import types from PIL import Image import json client = genai.Client() prompt = "Detect the all of the prominent items in the image. The box_2d should be [ymin, xmin, ymax, xmax] normalized to 0-1000." image = Image.open("/path/to/image.png") config = types.GenerateContentConfig( response_mime_type="application/json" ) response = client.models.generate_content(model="gemini-2.5-flash", contents=[image, prompt], config=config ) width, height = image.size bounding_boxes = json.loads(response.text) converted_bounding_boxes = [] for bounding_box in bounding_boxes: abs_y1 = int(bounding_box["box_2d"][0]/1000 * height) abs_x1 = int(bounding_box["box_2d"][1]/1000 * width) abs_y2 = int(bounding_box["box_2d"][2]/1000 * height) abs_x2 = int(bounding_box["box_2d"][3]/1000 * width) converted_bounding_boxes.append([abs_x1, abs_y1, abs_x2, abs_y2]) print("Image size: ", width, height) print("Bounding boxes:", converted_bounding_boxes) **Note:** The model also supports generating bounding boxes based on custom instructions, such as: "Show bounding boxes of all green objects in this image". It also support custom labels like "label the items with the allergens they can contain". For more examples, check following notebooks in the [Gemini Cookbook](https://github.com/google-gemini/cookbook): * [2D spatial understanding notebook](https://colab.research.google.com/github/google-gemini/cookbook/blob/main/quickstarts/Spatial_understanding.ipynb) * [Experimental 3D pointing notebook](https://colab.research.google.com/github/google-gemini/cookbook/blob/main/examples/Spatial_understanding_3d.ipynb) ## Segmentation Starting with Gemini 2.5, models not only detect items but also segment them and provide their contour masks. The model predicts a JSON list, where each item represents a segmentation mask. Each item has a bounding box ("`box_2d`") in the format `[y0, x0, y1, x1]` with normalized coordinates between 0 and 1000, a label ("`label`") that identifies the object, and finally the segmentation mask inside the bounding box, as base64 encoded png that is a probability map with values between 0 and 255. The mask needs to be resized to match the bounding box dimensions, then binarized at your confidence threshold (127 for the midpoint). **Note:** For better results, disable [thinking](/gemini-api/docs/thinking) by setting the thinking budget to 0. See code sample below for an example. from google import genai from google.genai import types from PIL import Image, ImageDraw import io import base64 import json import numpy as np import os client = genai.Client() def parse_json(json_output: str): # Parsing out the markdown fencing lines = json_output.splitlines() for i, line in enumerate(lines): if line == "```json": json_output = "\n".join(lines[i+1:]) # Remove everything before "```json" output = json_output.split("```")[0] # Remove everything after the closing "```" break # Exit the loop once "```json" is found return json_output def extract_segmentation_masks(image_path: str, output_dir: str = "segmentation_outputs"): # Load and resize image im = Image.open(image_path) im.thumbnail([1024, 1024], Image.Resampling.LANCZOS) prompt = """ Give the segmentation masks for the wooden and glass items. Output a JSON list of segmentation masks where each entry contains the 2D bounding box in the key "box_2d", the segmentation mask in key "mask", and the text label in the key "label". Use descriptive labels. """ config = types.GenerateContentConfig( thinking_config=types.ThinkingConfig(thinking_budget=0) # set thinking_budget to 0 for better results in object detection ) response = client.models.generate_content( model="gemini-2.5-flash", contents=[prompt, im], # Pillow images can be directly passed as inputs (which will be converted by the SDK) config=config ) # Parse JSON response items = json.loads(parse_json(response.text)) # Create output directory os.makedirs(output_dir, exist_ok=True) # Process each mask for i, item in enumerate(items): # Get bounding box coordinates box = item["box_2d"] y0 = int(box[0] / 1000 * im.size[1]) x0 = int(box[1] / 1000 * im.size[0]) y1 = int(box[2] / 1000 * im.size[1]) x1 = int(box[3] / 1000 * im.size[0]) # Skip invalid boxes if y0 >= y1 or x0 >= x1: continue # Process mask png_str = item["mask"] if not png_str.startswith("data:image/png;base64,"): continue # Remove prefix png_str = png_str.removeprefix("data:image/png;base64,") mask_data = base64.b64decode(png_str) mask = Image.open(io.BytesIO(mask_data)) # Resize mask to match bounding box mask = mask.resize((x1 - x0, y1 - y0), Image.Resampling.BILINEAR) # Convert mask to numpy array for processing mask_array = np.array(mask) # Create overlay for this mask overlay = Image.new('RGBA', im.size, (0, 0, 0, 0)) overlay_draw = ImageDraw.Draw(overlay) # Create overlay for the mask color = (255, 255, 255, 200) for y in range(y0, y1): for x in range(x0, x1): if mask_array[y - y0, x - x0] > 128: # Threshold for mask overlay_draw.point((x, y), fill=color) # Save individual mask and its overlay mask_filename = f"{item['label']}_{i}_mask.png" overlay_filename = f"{item['label']}_{i}_overlay.png" mask.save(os.path.join(output_dir, mask_filename)) # Create and save overlay composite = Image.alpha_composite(im.convert('RGBA'), overlay) composite.save(os.path.join(output_dir, overlay_filename)) print(f"Saved mask and overlay for {item['label']} to {output_dir}") # Example usage if __name__ == "__main__": extract_segmentation_masks("path/to/image.png") Check the [segmentation example](https://colab.research.google.com/github/google-gemini/cookbook/blob/main/quickstarts/Spatial_understanding.ipynb#scrollTo=WQJTJ8wdGOKx) in the cookbook guide for a more detailed example. ![A table with cupcakes, with the wooden and glass objects highlighted](/static/gemini-api/docs/images/segmentation.jpg) An example segmentation output with objects and segmentation masks ## Supported image formats Gemini supports the following image format MIME types: * PNG - `image/png` * JPEG - `image/jpeg` * WEBP - `image/webp` * HEIC - `image/heic` * HEIF - `image/heif` ## Capabilities All Gemini model versions are multimodal and can be utilized in a wide range of image processing and computer vision tasks including but not limited to image captioning, visual question and answering, image classification, object detection and segmentation. Gemini can reduce the need to use specialized ML models depending on your quality and performance requirements. Some later model versions are specifically trained improve accuracy of specialized tasks in addition to generic capabilities: * **Gemini 2.0 models** are further trained to support enhanced object detection. * **Gemini 2.5 models** are further trained to support enhanced segmentation in addition to object detection. ## Limitations and key technical information ### File limit Gemini 2.5 Pro/Flash, 2.0 Flash, 1.5 Pro, and 1.5 Flash support a maximum of 3,600 image files per request. ### Token calculation * **Gemini 1.5 Flash and Gemini 1.5 Pro** : 258 tokens if both dimensions <= 384 pixels. Larger images are tiled (min tile 256px, max 768px, resized to 768x768), with each tile costing 258 tokens. * **Gemini 2.0 Flash and Gemini 2.5 Flash/Pro** : 258 tokens if both dimensions <= 384 pixels. Larger images are tiled into 768x768 pixel tiles, each costing 258 tokens. ## Tips and best practices * Verify that images are correctly rotated. * Use clear, non-blurry images. * When using a single image with text, place the text prompt _after_ the image part in the `contents` array. ## What's next This guide shows you how to upload image files and generate text outputs from image inputs. To learn more, see the following resources: * [Files API](/gemini-api/docs/files): Learn more about uploading and managing files for use with Gemini. * [System instructions](/gemini-api/docs/text-generation#system-instructions): System instructions let you steer the behavior of the model based on your specific needs and use cases. * [File prompting strategies](/gemini-api/docs/files#prompt-guide): The Gemini API supports prompting with text, image, audio, and video data, also known as multimodal prompting. * [Safety guidance](/gemini-api/docs/safety-guidance): Sometimes generative AI models produce unexpected outputs, such as outputs that are inaccurate, biased, or offensive. Post-processing and human evaluation are essential to limit the risk of harm from such outputs.