Fleming-VL-38B

Fleming-VL-38B

<p align="center" style="margin: 0;"> <a href="https://github.com/UbiquantAI/Fleming-VL" aria-label="GitHub Repository" style="text-decoration:none;"> <span style="display:inline-flex;align-items:center;gap:.35em;"> <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" width="16" height="16" aria-hidden="true" style="vertical-align:text-bottom;fill:currentColor;"> <path d="M8 0C3.58 0 0 3.58 0 8c0 3.54 2.29 6.53 5.47 7.59.4.07.55-.17.55-.38 0-.19-.01-.82-.01-1.49-2.01.37-2.53-.49-2.69-.94-.09-.23-.48-.94-.82-1.13-.28-.15-.68-.52-.01-.53.63-.01 1.08.58 1.23.82.72 1.21 1.87.87 2.33.66.07-.52.28-.87.51-1.07-1.78-.2-3.64-.89-3.64-3.95 0-.87.31-1.59.82-2.15-.08-.2-.36-1.02.08-2.12 0 0 .67-.21 2.2.82.64-.18 1.32-.27 2-.27.68 0 1.36.09 2 .27 1.53-1.04 2.2-.82 2.2-.82.44 1.1.16 1.92.08 2.12.51.56.82 1.27.82 2.15 0 3.07-1.87 3.75-3.65 3.95.29.25.54.73.54 1.48 0 1.07-.01 1.93-.01 2.2 0 .21.15.46.55.38A8.013 8.013 0 0016 8c0-4.42-3.58-8-8-8Z"/> </svg> <span>GitHub</span> </span> </a> <span style="margin:0 .75em;opacity:.6;">•</span> <a href="https://arxiv.org/abs/2509.15279" aria-label="Paper">📑&nbsp;Paper</a> </p>

Highlights

📖 Model Overview

Fleming-VL is a multimodal reasoning model for medical scenarios that can process and analyze various types of medical data including 2D images, 3D volumetric data, and video sequences. The model performs step-by-step analysis of complex multimodal medical problems and produces reliable answers. Building upon the GRPO reasoning paradigm, Fleming-VL extends the capabilities to handle diverse medical imaging modalities while maintaining strong reasoning performance.

Model Features:

• Multimodal Processing Supports various medical data types including 2D images (X-rays, pathology slides), 3D volumes (CT/MRI scans), and videos (ultrasound, endoscopy, surgical recordings);
• Medical Reasoning Performs step-by-step chain-of-thought reasoning to analyze complex medical problems, combining visual information with medical knowledge to provide reliable diagnostic insights.

📦 Releases

• Fleming-VL-8B —— Trained on InternVL3-8B
  🤗 UbiquantAI/Fleming-VL-8B
• Fleming-VL-38B —— Trained on InternVL3-38B
  🤗 UbiquantAI/Fleming-VL-38B

📊 Performance

<div align="center"> <figure> <img src="images/main_benchmark.png" alt="Main Benchmark Results" width="60%"> <figcaption><b>Figure 1.</b> Main Benchmark Results.</figcaption> </figure> </div> <div align="center"> <figure> <img src="images/vqa.png" alt="General Medical Vqa" width="60%"> <figcaption><b>Figure 2.</b> General Medical VQA.</figcaption> </figure> </div> <div align="center"> <figure> <img src="images/report.png" alt="Medical Report Generation" width="60%"> <figcaption><b>Figure 3.</b> Medical Report Generation.</figcaption> </figure> </div> <div align="center"> <figure> <img src="images/video_3d.png" alt="Video and 3D understanding" width="60%"> <figcaption><b>Figure 4.</b> Video and 3D Understanding.</figcaption> </figure> </div>

🔧 Quick Start

# Fleming-VL-8B Multi-Modal Inference Script

# This script demonstrates three inference modes:
# 1. Single image inference
# 2. Video inference (frame-by-frame)
# 3. 3D medical image (CT/MRI) inference from .npy files

# Model: UbiquantAI/Fleming-VL-8B
# Based on: InternVL_chat-1.2 template


from transformers import AutoTokenizer, AutoModel
from torchvision.transforms.functional import InterpolationMode
from decord import VideoReader, cpu
from PIL import Image
import torchvision.transforms as T
import numpy as np
import torch
import os


# ============================================================================
# Configuration
# ============================================================================

MODEL_PATH = "UbiquantAI/Fleming-VL-38B"

# Prompt template for reasoning-based responses
REASONING_PROMPT = (
    "A conversation between User and Assistant. The user asks a question, "
    "and the Assistant solves it. The assistant first thinks about the "
    "reasoning process in the mind and then provides the user a concise "
    "final answer in a short word or phrase. The reasoning process and "
    "answer are enclosed within <think> </think> and <answer> </answer> "
    "tags, respectively, i.e., <think> reasoning process here </think>"
    "<answer> answer here </answer>"
)

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)


# ============================================================================
# Image Preprocessing Functions
# ============================================================================

def build_transform(input_size):
    """Build image transformation pipeline."""
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform


def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    """Find the closest aspect ratio from target ratios."""
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    """
    Dynamically preprocess image by splitting into tiles based on aspect ratio.
    
    Args:
        image: PIL Image
        min_num: Minimum number of tiles
        max_num: Maximum number of tiles
        image_size: Size of each tile
        use_thumbnail: Whether to add a thumbnail image
    
    Returns:
        List of preprocessed PIL Images
    """
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # Calculate possible tile configurations
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) 
        for i in range(1, n + 1) 
        for j in range(1, n + 1) 
        if i * j <= max_num and i * j >= min_num
    )
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # Find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size
    )

    # Calculate target dimensions
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # Resize and split the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    
    assert len(processed_images) == blocks
    
    # Add thumbnail if requested
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    
    return processed_images


# ============================================================================
# Utility Functions
# ============================================================================

def load_model(model_path, use_flash_attn=True):
    """
    Load the vision-language model and tokenizer.
    
    Args:
        model_path: Path to the pretrained model
        use_flash_attn: Whether to use flash attention (default: True)
    
    Returns:
        tuple: (model, tokenizer)
    """
    model = AutoModel.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        low_cpu_mem_usage=True,
        use_flash_attn=use_flash_attn,
        trust_remote_code=True
    ).eval().cuda()
    
    tokenizer = AutoTokenizer.from_pretrained(
        model_path,
        trust_remote_code=True,
        use_fast=False
    )
    
    return model, tokenizer


# ============================================================================
# Image Inference
# ============================================================================

def inference_single_image(model, tokenizer, image_path, question, 
                          prompt=REASONING_PROMPT, input_size=448, max_num=12):
    """
    Perform inference on a single image.
    
    Args:
        model: Loaded vision-language model
        tokenizer: Loaded tokenizer
        image_path: Path to the input image
        question: Question to ask about the image
        prompt: System prompt template
        input_size: Input image size (default: 448)
        max_num: Maximum number of tiles (default: 12)
    
    Returns:
        str: Model response
    """
    # Load and preprocess image using InternVL's dynamic preprocessing
    image = Image.open(image_path).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(img) for img in images]
    pixel_values = torch.stack(pixel_values).to(torch.bfloat16).cuda()
    
    # Prepare question with prompt and image token
    full_question = f"{prompt}\n<image>\n{question}"
    # print("###",full_question)
    
    # Generate response
    generation_config = dict(max_new_tokens=2048, do_sample=False)
    response = model.chat(tokenizer, pixel_values, full_question, generation_config)
    
    return response


# ============================================================================
# Video Inference
# ============================================================================

def get_frame_indices(bound, fps, max_frame, first_idx=0, num_segments=32):
    """
    Calculate evenly distributed frame indices for video sampling.
    
    Args:
        bound: Tuple of (start_time, end_time) in seconds, or None for full video
        fps: Frames per second of the video
        max_frame: Maximum frame index
        first_idx: First frame index to consider
        num_segments: Number of frames to sample
    
    Returns:
        np.array: Array of frame indices
    """
    if bound:
        start, end = bound[0], bound[1]
    else:
        start, end = -100000, 100000
    
    start_idx = max(first_idx, round(start * fps))
    end_idx = min(round(end * fps), max_frame)
    seg_size = float(end_idx - start_idx) / num_segments
    
    frame_indices = np.array([
        int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
        for idx in range(num_segments)
    ])
    
    return frame_indices


def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
    """
    Load and preprocess video frames.
    
    Args:
        video_path: Path to the video file
        bound: Time boundary tuple (start, end) in seconds
        input_size: Input image size (default: 448)
        max_num: Maximum number of tiles per frame (default: 1)
        num_segments: Number of frames to extract
    
    Returns:
        tuple: (pixel_values tensor, list of num_patches per frame)
    """
    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
    max_frame = len(vr) - 1
    fps = float(vr.get_avg_fps())
    
    pixel_values_list = []
    num_patches_list = []
    transform = build_transform(input_size=input_size)
    
    frame_indices = get_frame_indices(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
    
    for frame_index in frame_indices:
        # Extract and preprocess frame
        img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
        img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
        pixel_values = [transform(tile) for tile in img]
        pixel_values = torch.stack(pixel_values)
        num_patches_list.append(pixel_values.shape[0])
        pixel_values_list.append(pixel_values)
    
    pixel_values = torch.cat(pixel_values_list)
    return pixel_values, num_patches_list


def inference_video(model, tokenizer, video_path, video_duration, question, 
                   prompt=REASONING_PROMPT, input_size=448, max_num=1):
    """
    Perform inference on a video by sampling frames.
    
    Args:
        model: Loaded vision-language model
        tokenizer: Loaded tokenizer
        video_path: Path to the video file
        video_duration: Duration of video in seconds
        question: Question to ask about the video
        prompt: System prompt template
        input_size: Input image size (default: 448)
        max_num: Maximum number of tiles per frame (default: 1)
    
    Returns:
        str: Model response
    """
    # Sample frames from video (1 frame per second)
    num_segments = int(video_duration)
    pixel_values, num_patches_list = load_video(
        video_path, bound=None, input_size=input_size, 
        max_num=max_num, num_segments=num_segments
    )
    pixel_values = pixel_values.to(torch.bfloat16).cuda()
    
    # Create image token prefix for all frames
    video_prefix = ''.join([f'<image>\n' for _ in range(len(num_patches_list))])
    
    # Prepare question with prompt and image tokens
    full_question = f"{prompt}\n{video_prefix}{question}"
    
    # Generate response
    generation_config = dict(max_new_tokens=1024, do_sample=False)
    response, history = model.chat(
        tokenizer,
        pixel_values,
        full_question,
        generation_config,
        num_patches_list=num_patches_list,
        history=None,
        return_history=True
    )
    
    return response


# ============================================================================
# 3D Medical Image (NPY) Inference
# ============================================================================

def normalize_image(image):
    """
    Normalize image array to 0-255 range.
    
    Args:
        image: NumPy array of image data
    
    Returns:
        np.array: Normalized image as uint8
    """
    img_min = np.min(image)
    img_max = np.max(image)
    
    if img_max - img_min == 0:
        return np.zeros_like(image, dtype=np.uint8)
    
    return ((image - img_min) / (img_max - img_min) * 255).astype(np.uint8)


def convert_npy_to_images(npy_path, input_size=448, max_num=1, num_slices=11):
    """
    Convert 3D medical image (.npy) to multiple 2D RGB images.
    
    Expected input shape: (32, 256, 256) or (1, 32, 256, 256)
    Extracts evenly distributed slices and converts to RGB format.
    
    Args:
        npy_path: Path to the .npy file
        input_size: Input image size (default: 448)
        max_num: Maximum number of tiles per slice (default: 1)
        num_slices: Number of slices to extract (default: 11)
    
    Returns:
        tuple: (pixel_values tensor, list of num_patches per slice) or False if error
    """
    try:
        # Load .npy file
        data = np.load(npy_path)
        
        # Handle shape (1, 32, 256, 256) -> (32, 256, 256)
        if data.ndim == 4 and data.shape[0] == 1:
            data = data[0]
        
        # Validate shape
        if data.shape != (32, 256, 256):
            print(f"Warning: {npy_path} has shape {data.shape}, expected (32, 256, 256), skipping")
            return False
        
        # Select evenly distributed slices from 32 slices
        indices = np.linspace(0, 31, num_slices, dtype=int)
        
        transform = build_transform(input_size=input_size)
        pixel_values_list = []
        num_patches_list = []
        
        # Process each selected slice
        for idx in indices:
            # Get slice
            slice_img = data[idx]
            
            # Normalize to 0-255
            normalized = normalize_image(slice_img)
            
            # Convert grayscale to RGB by stacking
            rgb_img = np.stack([normalized, normalized, normalized], axis=-1)
            
            # Convert to PIL Image
            img = Image.fromarray(rgb_img)
            
            # Preprocess with InternVL's dynamic preprocessing
            img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
            pixel_values = [transform(tile) for tile in img]
            pixel_values = torch.stack(pixel_values)
            num_patches_list.append(pixel_values.shape[0])
            pixel_values_list.append(pixel_values)
        
        pixel_values = torch.cat(pixel_values_list)
        return pixel_values, num_patches_list
    
    except Exception as e:
        print(f"Error processing {npy_path}: {str(e)}")
        return False


def inference_3d_medical_image(model, tokenizer, npy_path, question, 
                              prompt=REASONING_PROMPT, input_size=448, max_num=1):
    """
    Perform inference on 3D medical images stored as .npy files.
    
    Args:
        model: Loaded vision-language model
        tokenizer: Loaded tokenizer
        npy_path: Path to the .npy file (shape: 32x256x256)
        question: Question to ask about the image
        prompt: System prompt template
        input_size: Input image size (default: 448)
        max_num: Maximum number of tiles per slice (default: 1)
    
    Returns:
        str: Model response or None if error
    """
    # Convert 3D volume to multiple 2D slices
    result = convert_npy_to_images(npy_path, input_size=input_size, max_num=max_num)
    
    if result is False:
        return None
    
    pixel_values, num_patches_list = result
    pixel_values = pixel_values.to(torch.bfloat16).cuda()
    
    # Create image token prefix for all slices
    image_prefix = ''.join([f'<image>\n' for _ in range(len(num_patches_list))])
    
    # Prepare question with prompt and image tokens
    full_question = f"{prompt}\n{image_prefix}{question}"
    
    # Generate response
    generation_config = dict(max_new_tokens=1024, do_sample=False)
    response, history = model.chat(
        tokenizer,
        pixel_values,
        full_question,
        generation_config,
        num_patches_list=num_patches_list,
        history=None,
        return_history=True
    )
    
    return response


# ============================================================================
# Main Execution Examples
# ============================================================================

def main():
    """
    Main function demonstrating all three inference modes.
    """
    
    # ========================================================================
    # Example 1: Single Image Inference
    # ========================================================================
    print("\n" + "="*80)
    print("EXAMPLE 1: Single Image Inference")
    print("="*80)
    
    image_path = "./resource/1.jpg"
    question = ' What type of abnormality is present in this image?'
    
    model, tokenizer = load_model(MODEL_PATH, use_flash_attn=True)
    response = inference_single_image(model, tokenizer, image_path, question)
    
    print(f"\nUser: {question}")
    print(f"Assistant: {response}")
    
    # Clean up GPU memory
    del model, tokenizer
    torch.cuda.empty_cache()
    
    # ========================================================================
    # Example 2: Video Inference
    # ========================================================================
    print("\n" + "="*80)
    print("EXAMPLE 2: Video Inference")
    print("="*80)
    
    video_path = "./resource/video.mp4"
    video_duration = 6  # seconds
    question = "Please describe the video."
    
    model, tokenizer = load_model(MODEL_PATH, use_flash_attn=False)
    response = inference_video(model, tokenizer, video_path, video_duration, question)
    
    print(f"\nUser: {question}")
    print(f"Assistant: {response}")
    
    # Clean up GPU memory
    del model, tokenizer
    torch.cuda.empty_cache()
    
    # ========================================================================
    # Example 3: 3D Medical Image Inference
    # ========================================================================
    print("\n" + "="*80)
    print("EXAMPLE 3: 3D Medical Image Inference")
    print("="*80)
    
    npy_path = "./resource/test.npy"
    question = "What device is observed on the chest wall?"
    
    # Example cases:
    # Case 1: /path/to/test_1016_d_2.npy
    #   Question: "Where is the largest lymph node observed?"
    #   Answer: "Right hilar region."
    #
    # Case 2: /path/to/test_1031_a_2.npy
    #   Question: "What device is observed on the chest wall?"
    #   Answer: "Pacemaker."
    
    model, tokenizer = load_model(MODEL_PATH, use_flash_attn=False)
    response = inference_3d_medical_image(model, tokenizer, npy_path, question)
    
    if response:
        print(f"\nUser: {question}")
        print(f"Assistant: {response}")
    else:
        print("\nError: Failed to process 3D medical image")
    
    # Clean up GPU memory
    del model, tokenizer
    torch.cuda.empty_cache()


if __name__ == "__main__":
    main()

⚠️ Safety Statement

This project is for research and non-clinical reference only; it must not be used for actual diagnosis or treatment decisions.
The generated reasoning traces are an auditable intermediate process and do not constitute medical advice.
In medical scenarios, results must be reviewed and approved by qualified professionals, and all applicable laws, regulations, and privacy compliance requirements in your region must be followed.

📚 Citation

@misc{flemingvl,
      title={Fleming-VL: Towards Universal Medical Visual Reasoning with Multimodal LLMs}, 
      author={Yan Shu and Chi Liu and Robin Chen and Derek Li and Bryan Dai},
      year={2025},
      eprint={2511.00916},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.00916}, 
}