InternVL Family: Closing the Gap to Commercial Multimodal Models with Open-Source Suites —— A Pioneering Open-Source Alternative to GPT-5

• 2024/10/21: We release the Mini-InternVL series. These models achieve impressive performance with minimal size: the 4B model achieves 90% of the performance with just 5% of the model size. For more details, please check our project page and document.
• 2024/08/01: The Chartmimic team evaluated the InternVL2 series models on their benchmark. The InternVL2-26B and 76B models achieved the top two performances among open-source models, with the InternVL2 76B model surpassing GeminiProVision and exhibiting comparable results to Claude-3-opus.
• 2024/08/01: InternVL2-Pro achieved the SOTA performance among open-source models on the CharXiv dataset, surpassing many closed-source models such as GPT-4V, Gemini 1.5 Flash, and Claude 3 Sonnet.
• 2024/07/24: The MLVU team evaluated InternVL-1.5 on their benchmark. The average performance on the multiple-choice task was 50.4%, while the performance on the generative tasks was 4.02. The performance on the multiple-choice task ranked #1 among all open-source MLLMs.
• 2024/07/04: We release the InternVL2 series. InternVL2-Pro achieved a 62.0% accuracy on the MMMU benchmark, matching the performance of leading closed-source commercial models like GPT-4o.
• 2024/07/18: InternVL2-40B achieved SOTA performance among open-source models on the Video-MME dataset, scoring 61.2 when inputting 16 frames and 64.4 when inputting 32 frames. It significantly outperforms other open-source models and is the closest open-source model to GPT-4o mini.
• 2024/07/18: InternVL2-Pro achieved the SOTA performance on the DocVQA and InfoVQA benchmarks.
• 2024/06/19: We propose Needle In A Multimodal Haystack (MM-NIAH), the first benchmark designed to systematically evaluate the capability of existing MLLMs to comprehend long multimodal documents.
• 2024/05/30: We release ShareGPT-4o, a large-scale dataset that we plan to open-source with 200K images, 10K videos, and 10K audios with detailed descriptions.
• 2024/05/28: Thanks to the lmdeploy team for providing AWQ quantization support. The 4-bit model is available at OpenGVLab/InternVL-Chat-V1-5-AWQ.
• 2024/05/13: InternVL 1.0 can now be used as the text encoder for diffusion models to support multilingual generation natively in over 110 languages worldwide. See MuLan for more details.
• 2024/04/18: InternVL-Chat-V1-5 has been released at HF link, approaching the performance of GPT-4V and Gemini Pro on various benchmarks like MMMU, DocVQA, ChartQA, MathVista, etc.
• 2024/02/27: InternVL is accepted by CVPR 2024 (Oral)! 🎉
• 2024/02/21: InternVL-Chat-V1-2-Plus achieved SOTA performance on MathVista (59.9), MMBench (83.8), and MMVP (58.7). See our blog for more details.
• 2024/02/12: InternVL-Chat-V1-2 has been released. It achieves 51.6 on MMMU val and 82.3 on MMBench test. For more details, please refer to our blog and SFT data. The model is now available on HuggingFace, and both training / evaluation data and scripts are open-sourced.
• 2024/01/24: InternVL-Chat-V1-1 is released, it supports Chinese and has stronger OCR capability, see here.
• 2024/01/16: We release our customized mmcv/mmsegmentation/mmdetection code, integrated with DeepSpeed, which can be used for training large-scale detection and segmentation models.

• Linear-Probe Image Classification [see details]

  ViT-22B uses the private JFT-3B dataset.

  method	#param	IN-1K	IN-ReaL	IN-V2	IN-A	IN-R	IN-Sketch
  OpenCLIP-G	1.8B	86.2	89.4	77.2	63.8	87.8	66.4
  DINOv2-g	1.1B	86.5	89.6	78.4	75.9	78.8	62.5
  EVA-01-CLIP-g	1.1B	86.5	89.3	77.4	70.5	87.7	63.1
  MAWS-ViT-6.5B	6.5B	87.8	-	-	-	-	-
  ViT-22B*	21.7B	89.5	90.9	83.2	83.8	87.4	-
  InternViT-6B (ours)	5.9B	88.2	90.4	79.9	77.5	89.8	69.1

• Semantic Segmentation [see details]

  method	decoder	#param (train/total)	crop size	mIoU
  OpenCLIP-G (frozen)	Linear	0.3M / 1.8B	512	39.3
  ViT-22B (frozen)	Linear	0.9M / 21.7B	504	34.6
  InternViT-6B (frozen)	Linear	0.5M / 5.9B	504	47.2 (+12.6)
  ViT-22B (frozen)	UperNet	0.8B / 22.5B	504	52.7
  InternViT-6B (frozen)	UperNet	0.4B / 6.3B	504	54.9 (+2.2)
  ViT-22B	UperNet	22.5B / 22.5B	504	55.3
  InternViT-6B	UperNet	6.3B / 6.3B	504	58.9 (+3.6)

• Zero-Shot Image Classification [see details]

  method	IN-1K	IN-A	IN-R	IN-V2	IN-Sketch	ObjectNet
  OpenCLIP-G	80.1	69.3	92.1	73.6	68.9	73.0
  EVA-02-CLIP-E+	82.0	82.1	94.5	75.7	71.6	79.6
  ViT-22B*	85.9	90.1	96.0	80.9	-	87.6
  InternVL-C (ours)	83.2	83.8	95.5	77.3	73.9	80.6

• Multilingual Zero-Shot Image Classification [see details]

  EN: English, ZH: Chinese, JP: Japanese, Ar: Arabic, IT: Italian

  method	IN-1K (EN)	IN-1K (ZH)	IN-1K (JP)	IN-1K (AR)	IN-1K (IT)
  Taiyi-CLIP-ViT-H	-	54.4	-	-	-
  WuKong-ViT-L-G	-	57.5	-	-	-
  CN-CLIP-ViT-H	-	59.6	-	-	-
  AltCLIP-ViT-L	74.5	59.6	-	-	-
  EVA-02-CLIP-E+	82.0	-	-	-	41.2
  OpenCLIP-XLM-R-H	77.0	55.7	53.1	37.0	56.8
  InternVL-C (ours)	83.2	64.5	61.5	44.9	65.7

• Zero-Shot Video Classification

  method	#frame	K400	K600	K700
  OpenCLIP-G	1	65.9	66.1	59.2
  EVA-02-CLIP-E+	1	69.8	69.3	63.4
  InternVL-C (ours)	1	71.0	71.3	65.7
  ViCLIP	8	75.7	73.5	66.4
  InternVL-C (ours)	8	79.4	78.8	71.5

• English Zero-Shot Image-Text Retrieval [see details]

  model	Flickr30K						COCO						avg
  	image-to-text			text-to-image			image-to-text			text-to-image
  	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10
  OpenCLIP-G	92.9	99.3	99.8	79.5	95.0	97.1	67.3	86.9	92.6	51.4	74.9	83.0	85.0
  EVA-02-CLIP-E+	93.9	99.4	99.8	78.8	94.2	96.8	68.8	87.8	92.8	51.1	75.0	82.7	85.1
  EVA-CLIP-8B	95.6	99.6	99.9	80.8	95.5	97.6	70.3	89.3	93.9	53.0	76.0	83.4	86.2
  InternVL-C (ours)	94.7	99.6	99.9	81.7	96.0	98.2	70.6	89.0	93.5	54.1	77.3	84.6	86.6
  InternVL-G (ours)	95.7	99.7	99.9	85.0	97.0	98.6	74.9	91.3	95.2	58.6	81.3	88.0	88.8

• Chinese Zero-Shot Image-Text Retrieval [see details]

  model	Flickr30K-CN						COCO-CN						avg
  	image-to-text			text-to-image			image-to-text			text-to-image
  	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10
  CN-CLIP-ViT-H	81.6	97.5	98.8	71.2	91.4	95.5	63.0	86.6	92.9	69.2	89.9	96.1	86.1
  OpenCLIP-XLM-R-H	86.1	97.5	99.2	71.0	90.5	94.9	70.0	91.5	97.0	66.1	90.8	96.0	87.6
  InternVL-C (ours)	90.3	98.8	99.7	75.1	92.9	96.4	68.8	92.0	96.7	68.9	91.9	96.5	89.0
  InternVL-G (ours)	92.9	99.4	99.8	77.7	94.8	97.3	71.4	93.9	97.7	73.8	94.4	98.1	90.9

• Multilingual Zero-Shot Image-Text Retrieval on XTD [see details]

  method	EN	ES	FR	ZH	IT	KO	RU	JP	average
  AltCLIP	95.4	94.1	92.9	95.1	94.2	94.4	91.8	91.7	93.7
  OpenCLIP-XLM-R-H	97.3	96.1	94.5	94.7	96.0	90.2	93.9	94.0	94.6
  InternVL-C (ours)	97.3	95.7	95.1	95.6	96.0	92.2	93.3	95.5	95.1
  InternVL-G (ours)	98.6	97.7	96.5	96.7	96.9	95.1	94.8	96.1	96.6

import torch
from PIL import Image
from transformers import AutoModel, CLIPImageProcessor

model = AutoModel.from_pretrained(
    'OpenGVLab/InternViT-6B-448px-V2_5',
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True).cuda().eval()

image = Image.open('./examples/image1.jpg').convert('RGB')

image_processor = CLIPImageProcessor.from_pretrained('OpenGVLab/InternViT-6B-448px-V1-5')

pixel_values = image_processor(images=image, return_tensors='pt').pixel_values
pixel_values = pixel_values.to(torch.bfloat16).cuda()

outputs = model(pixel_values)

import torch
from PIL import Image
from transformers import AutoModel, CLIPImageProcessor
from transformers import AutoTokenizer


model = AutoModel.from_pretrained(
    'OpenGVLab/InternVL-14B-224px',
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True).cuda().eval()

image_processor = CLIPImageProcessor.from_pretrained('OpenGVLab/InternVL-14B-224px')

tokenizer = AutoTokenizer.from_pretrained(
    'OpenGVLab/InternVL-14B-224px', use_fast=False, add_eos_token=True)
tokenizer.pad_token_id = 0  # set pad_token_id to 0

images = [
    Image.open('./examples/image1.jpg').convert('RGB'),
    Image.open('./examples/image2.jpg').convert('RGB'),
    Image.open('./examples/image3.jpg').convert('RGB')
]
prefix = 'summarize:'
texts = [
    prefix + 'a photo of a red panda',  # English
    prefix + '一张熊猫的照片',  # Chinese
    prefix + '二匹の猫の写真'  # Japanese
]

pixel_values = image_processor(images=images, return_tensors='pt').pixel_values
pixel_values = pixel_values.to(torch.bfloat16).cuda()
input_ids = tokenizer(texts, return_tensors='pt', max_length=80,
                      truncation=True, padding='max_length').input_ids.cuda()

# InternVL-C
logits_per_image, logits_per_text = model(
    image=pixel_values, text=input_ids, mode='InternVL-C')
probs = logits_per_image.softmax(dim=-1)
# tensor([[9.9609e-01, 5.2185e-03, 6.0070e-08],
#         [2.2949e-02, 9.7656e-01, 5.9903e-06],
#         [3.2932e-06, 7.4863e-05, 1.0000e+00]], device='cuda:0',
#        dtype=torch.bfloat16, grad_fn=<SoftmaxBackward0>)

# InternVL-G
logits_per_image, logits_per_text = model(
    image=pixel_values, text=input_ids, mode='InternVL-G')
probs = logits_per_image.softmax(dim=-1)
# tensor([[9.9609e-01, 3.1738e-03, 3.6322e-08],
#         [8.6060e-03, 9.9219e-01, 2.8759e-06],
#         [1.7583e-06, 3.1233e-05, 1.0000e+00]], device='cuda:0',
#        dtype=torch.bfloat16, grad_fn=<SoftmaxBackward0>)

# please set add_eos_token to False for generation
tokenizer.add_eos_token = False
image = Image.open('./examples/image1.jpg').convert('RGB')
pixel_values = image_processor(images=image, return_tensors='pt').pixel_values
pixel_values = pixel_values.to(torch.bfloat16).cuda()

tokenized = tokenizer("English caption:", return_tensors='pt')
pred = model.generate(
    pixel_values=pixel_values,
    input_ids=tokenized.input_ids.cuda(),
    attention_mask=tokenized.attention_mask.cuda(),
    num_beams=5,
    min_new_tokens=8,
)
caption = tokenizer.decode(pred[0].cpu(), skip_special_tokens=True).strip()
# English caption: a red panda sitting on top of a wooden platform

Here, we take the smaller OpenGVLab/InternVL2_5-8B as an example:

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

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

def build_transform(input_size):
    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):
    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):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    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 the target width and height
    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 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 the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).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(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

# If you have an 80G A100 GPU, you can put the entire model on a single GPU.
# Otherwise, you need to load a model using multiple GPUs, please refer to the `Multiple GPUs` section.
path = 'OpenGVLab/InternVL2_5-8B'
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)

# set the max number of tiles in `max_num`
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=False)

# pure-text conversation (纯文本对话)
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'Can you tell me a story?'
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# single-image single-round conversation (单图单轮对话)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')

# single-image multi-round conversation (单图多轮对话)
question = '<image>\nPlease describe the image in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'Please write a poem according to the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# multi-image multi-round conversation, combined images (多图多轮对话，拼接图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)

question = '<image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# multi-image multi-round conversation, separate images (多图多轮对话，独立图像)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]

question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               num_patches_list=num_patches_list,
                               history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'What are the similarities and differences between these two images.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               num_patches_list=num_patches_list,
                               history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# batch inference, single image per sample (单图批处理)
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)

questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
responses = model.batch_chat(tokenizer, pixel_values,
                             num_patches_list=num_patches_list,
                             questions=questions,
                             generation_config=generation_config)
for question, response in zip(questions, responses):
    print(f'User: {question}\nAssistant: {response}')

# video multi-round conversation (视频多轮对话)
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
    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):
    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_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
    for frame_index in frame_indices:
        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

video_path = './examples/red-panda.mp4'
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
pixel_values = pixel_values.to(torch.bfloat16).cuda()
video_prefix = ''.join([f'Frame-{i+1}: <image>\n' for i in range(len(num_patches_list))])
question = video_prefix + 'What is the red panda doing?'
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               num_patches_list=num_patches_list, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

question = 'Describe this video in detail.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
                               num_patches_list=num_patches_list, history=history, return_history=True)
print(f'User: {question}\nAssistant: {response}')