Transformers documentation

VideoLLaMA3

Transformers

You are viewing main version, which requires installation from source. If you'd like regular pip install, checkout the latest stable version (v5.14.0).

Join the Hugging Face community

and get access to the augmented documentation experience

Collaborate on models, datasets and Spaces

Faster examples with accelerated inference

Switch between documentation themes

to get started

This model was published in HF papers on 2025-01-22 and contributed to Hugging Face Transformers on 2025-10-13.

VideoLLaMA3

Overview

The VideoLLaMA3 model is a major update to VideoLLaMA2 from Alibaba DAMO Academy.

The abstract from the paper is as following:

In this paper, we propose VideoLLaMA 3, a more advanced multimodal foundation model for image and video understanding. The core design philosophy of VideoLLaMA3 is vision-centric. The meaning of “vision-centric” is two-fold: the vision-centric training paradigm and vision-centric framework design. The key insight of our vision-centric training paradigm is that high-quality image-text data is crucial for both image and video understanding. Instead of preparing massive video-text datasets, we focus on constructing large-scale, high-quality image-text datasets. VideoLLaMA3 has four training stages: 1) Vision Encoder Adaptation, which enables the vision encoder to accept images of variable resolutions as input; 2) Vision-Language Alignment, which jointly tunes the vision encoder, projector, and LLM with large-scale image-text data covering multiple types (including scene images, documents, and charts) as well as text-only data. 3) Multi-task Fine-tuning, which incorporates image-text SFT data for downstream tasks and video-text data to establish a foundation for video understanding. 4) Video-centric Fine-tuning, which further improves the model’s capability in video understanding. As for the framework design, to better capture fine-grained details in images, the pretrained vision encoder is adapted to encode images of varying sizes into vision tokens with corresponding numbers, rather than a fixed number of tokens. For video inputs, we reduce the number of vision tokens according to their similarity so that the representation of videos will be more precise and compact. Benefiting from vision-centric designs, VideoLLaMA3 achieves compelling performances in both image and video understanding benchmarks.

VideoLLaMA3 architecture. Taken from the technical report.

This model was contributed by lkhl.

Usage example

Single Media inference

The model can accept both images and videos as input. Here’s an example code for inference.

from transformers import AutoProcessor, VideoLlama3ForConditionalGeneration


# Load the model in half-precision on the available device(s)
model = VideoLlama3ForConditionalGeneration.from_pretrained("lkhl/VideoLLaMA3-2B-Image-HF", device_map="auto")
processor = AutoProcessor.from_pretrained("lkhl/VideoLLaMA3-2B-Image-HF")


conversation = [
    {
        "role":"user",
        "content":[
            {"type": "image", "image": "https://github.com/DAMO-NLP-SG/VideoLLaMA3/raw/refs/heads/main/assets/sora.png"},
            {"type": "text", "text": "Describe this image."}
        ]
    }
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device)

# Inference: Generation of the output
output_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids)]
output_text = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
print(output_text)



# Video
conversation = [
    {
        "role": "user",
        "content": [
            {"type": "video", "video": "https://github.com/DAMO-NLP-SG/VideoLLaMA3/raw/refs/heads/main/assets/cat_and_chicken.mp4"},
            {"type": "text", "text": "What happened in the video?"},
        ],
    }
]

inputs = processor.apply_chat_template(
    conversation,
    fps=1,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device)


# Inference: Generation of the output
output_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids)]
output_text = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
print(output_text)

Batch Mixed Media Inference

The model can batch inputs composed of mixed samples of various types such as images, videos, and text. Here is an example.

# Image
conversation1 = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "https://github.com/DAMO-NLP-SG/VideoLLaMA3/raw/refs/heads/main/assets/sora.png"},
            {"type": "text", "text": "Describe this image."}
        ]
    }
]

# Video
conversation2 = [
    {
        "role": "user",
        "content": [
            {"type": "video", "video": "https://github.com/DAMO-NLP-SG/VideoLLaMA3/raw/refs/heads/main/assets/cat_and_chicken.mp4"},
            {"type": "text", "text": "What happened in the video?"},
        ],
    }
]

# Text
conversation3 = [
    {
        "role": "user",
        "content": "What color is a banana?"
    }
]


conversations = [conversation1, conversation2, conversation3]
# Preparation for batch inference
inputs = processor.apply_chat_template(
    conversations,
    fps=1,
    add_generation_prompt=True,
    tokenize=True,
    padding=True,
    padding_side="left",
    return_dict=True,
    return_tensors="pt"
).to(model.device)


# Batch Inference
output_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids)]
output_text = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
print(output_text)

Flash-Attention 2 to speed up generation

First, make sure to install the latest version of Flash Attention 2:

pip install -U flash-attn --no-build-isolation

Also, you should have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of the flash attention repository. FlashAttention-2 can only be used when a model is loaded in torch.float16 or torch.bfloat16.

To load and run a model using Flash Attention-2, simply add attn_implementation="flash_attention_2" when loading the model as follows:

from transformers import VideoLlama3ForConditionalGeneration

model = VideoLlama3ForConditionalGeneration.from_pretrained(
    "lkhl/VideoLLaMA3-2B-Image-HF", , 
    attn_implementation="flash_attention_2",
 device_map="auto")

Transformers

VideoLLaMA3

Overview

Usage example

Single Media inference

Batch Mixed Media Inference

Flash-Attention 2 to speed up generation

VideoLlama3Config

class transformers.VideoLlama3Config

VideoLlama3VisionConfig

class transformers.VideoLlama3VisionConfig

VideoLlama3ImageProcessor

class transformers.VideoLlama3ImageProcessor

preprocess

VideoLlama3VideoProcessor

class transformers.VideoLlama3VideoProcessor

preprocess

VideoLlama3ImageProcessorPil

class transformers.VideoLlama3ImageProcessorPil

preprocess

VideoLlama3Processor

class transformers.VideoLlama3Processor

__call__

VideoLlama3Model

class transformers.VideoLlama3Model

forward

get_video_features

get_image_features

VideoLlama3VisionModel

class transformers.VideoLlama3VisionModel

forward

VideoLlama3ForConditionalGeneration

class transformers.VideoLlama3ForConditionalGeneration

forward

get_video_features

get_image_features

call