Implementation:Deepseek ai Janus Load Pil Images

Knowledge Sources	Janus
Domains	Computer_Vision, Preprocessing
Last Updated	2026-02-10 09:30 GMT

Overview

Concrete tool for extracting and loading PIL images from conversation message dictionaries provided by the Janus I/O utilities.

Description

The load_pil_images function iterates over a list of conversation message dicts, extracts any images field entries, and loads them as PIL Image objects. It supports both file paths and base64 data URIs. All images are converted to RGB format.

Usage

Import this function when you have a conversation dict with image references and need to convert them to PIL images for the VLChatProcessor.

Code Reference

Source Location

• Repository: Janus
• File: janus/utils/io.py
• Lines: L44-83

Signature

def load_pil_images(
    conversations: List[Dict[str, str]]
) -> List[PIL.Image.Image]:
    """
    Load PIL images from conversation messages.

    Supports file paths and base64 data URIs.

    Args:
        conversations: List of message dicts with optional "images" key
            containing file paths or "data:image" base64 strings.

    Returns:
        List of RGB PIL.Image.Image objects extracted from all messages.
    """

Import

from janus.utils.io import load_pil_images

I/O Contract

Inputs

Name	Type	Required	Description
conversations	List[Dict[str, str]]	Yes	Conversation message list; messages with "images" key will be processed

Outputs

Name	Type	Description
pil_images	List[PIL.Image.Image]	RGB-converted PIL images from all messages, in order

Usage Examples

Load From File Paths

from janus.utils.io import load_pil_images

conversation = [
    {
        "role": "User",
        "content": "<image_placeholder>\nDescribe this image.",
        "images": ["./examples/image.png"],
    },
    {"role": "Assistant", "content": ""},
]

pil_images = load_pil_images(conversation)
# Returns: [<PIL.Image.Image mode='RGB'>]

Load From Base64

import base64

with open("image.png", "rb") as f:
    b64 = base64.b64encode(f.read()).decode()

conversation = [
    {
        "role": "User",
        "content": "<image_placeholder>\nWhat is this?",
        "images": [f"data:image/png;base64,{b64}"],
    },
    {"role": "Assistant", "content": ""},
]

pil_images = load_pil_images(conversation)

Related Pages

Implements Principle

• Principle:Deepseek_ai_Janus_Image_Loading_and_Preprocessing