Heron-NVILA-Lite-2B is a vision language model trained for Japanese, based on the NVILA-Lite architecture. This implementation is compatible with transformers.AutoModelForImageTextToText.

Model Overview

Developer: Turing Inc.
Vision Encoder: paligemma-siglip-so400m-patch14-448
Projector: mlp_downsample_2x2_fix
LLM: Qwen2.5-1.5B-Instruct
Supported Languages: Japanese, English

Setup

bash

# We have developed this implementation with `transformers==4.57.1`. Other versions of transformers may also work, but we have not tested them.
pip install "transformers[torch]==4.57.1" pillow

Usage

python

from transformers import AutoConfig, AutoModelForImageTextToText, AutoProcessor

MODEL_NAME = "turing-motors/Heron-NVILA-Lite-2B-hf"

# you can use config
config = AutoConfig.from_pretrained(MODEL_NAME, trust_remote_code=True)
model = AutoModelForImageTextToText.from_config(config, trust_remote_code=True)

# or directly from_pretrained
model = AutoModelForImageTextToText.from_pretrained(MODEL_NAME, trust_remote_code=True, device_map="auto")

# load processor
processor = AutoProcessor.from_pretrained(MODEL_NAME, trust_remote_code=True)

# show chat_template
print(processor.tokenizer.chat_template)

def generate_content(content: str, images: list | None = None, **kwargs) -> str:
    conversation = [{"role": "user", "content": content}]
    text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
    encoding = processor(text=text, images=images, return_tensors="pt").to(model.device)
    output = model.generate(**encoding, **kwargs)
    return processor.decode(output[0, len(encoding["input_ids"][0]):], skip_special_tokens=True)

# examples generate with raw text
response = generate_content("こんにちは")
print(response)
print("---" * 40)

# examples generate with text + image
from PIL import Image
import requests
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
response = generate_content("<image>\n画像を説明してください。", images=[image])
print(response)
print("---" * 40)

# examples generate using generation_config
from PIL import Image
import requests
from transformers import GenerationConfig
generation_config = {
    "max_new_tokens": 512,
    "temperature": 0.5,
    "do_sample": True,
}
generation_config = GenerationConfig(**generation_config)
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
response = generate_content(
    "<image>\n画像を説明してください。",
    images=[image],
    generation_config=generation_config
)
print(response)
print("---" * 40)

# examples generate with text + image + text + image + text
from PIL import Image
import requests
url_list = [
    "https://images.unsplash.com/photo-1694831404826-3400c48c188d?q=80&w=2070&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D",
    "https://images.unsplash.com/photo-1693240876439-473af88b4ed7?q=80&w=1974&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D"
]
images = [
   Image.open(requests.get(url, stream=True).raw).convert("RGB") for url in url_list
]
response = generate_content("".join([
    "<image>\n",
    "これは日本の画像です",
    "<image>\n",
    "これはオーストリアの画像です",
    "各画像の違いを説明して"]), images)
print(response)
print("---" * 40)

Training Summary

Stage	Training	Data Sources	Samples
Stage1	Projector	Japanese image text pairs, LLaVA-Pretrain	1.1M
Stage2	Projector, LLM	Filtered MOMIJI (CC-MAIN-2024-46, CC-MAIN-2024-51, CC-MAIN-2025-05)	13M
		Japanese image text pairs (subset), Japanese interleaved data (subset), mmc4-core (subset), coyo-700m (subset), wikipedia_ja, llava_pretrain_ja, stair_captions	20M
Stage3	Vision Encoder, Projector, LLM	llava-instruct-v1_5-en-subset-358k, llava-instruct-ja, japanese-photos-conv, ja-vg-vqa, synthdog-ja (subset), ai2d, synthdog-en, sherlock	1.1M

Evaluation

I used llm-jp-eval-mm for this evaluation. Scores for models other than Heron-NVILA-Lite and Sarashina2-Vision-14B were taken from llm-jp-eval-mm leaderboard as of March 2025 and the Asagi website. Heron-NVILA-Lite and Sarashina2-Vision-14B were evaluated using llm-as-a-judge with "gpt-4o-2024-05-13". Sarashina2-Vision-14B was evaluated on the official blog using "gpt-4o-2024-08-06"; please note that due to differing evaluation conditions, the results for Sarashina2-Vision-14B should be treated as reference only.

Model	LLM Size	Heron-Bench overall LLM (%)	JA-VLM-Bench-In-the-Wild LLM (/5.0)	JA-VG-VQA-500 LLM (/5.0)
Heron-NVILA-Lite-1B	0.5B	45.9	2.92	3.16
Heron-NVILA-Lite-2B	1.5B	52.8	3.52	3.50
Heron-NVILA-Lite-15B	14B	59.6	4.2	3.82
LLaVA-CALM2-SigLIP	7B	43.3	3.15	3.21
Llama-3-EvoVLM-JP-v2	8B	39.3	2.92	2.96
VILA-jp	13B	57.2	3.69	3.62
Asagi-14B	13B	55.8	3.44	3.84
Sarashina2-Vision-14B	13B	50.9	4.1	3.43
Qwen2-VL 7B Instruct	7B	55.5	3.61	3.6
GPT-4o	-	87.6	3.85	3.58

Risks and Limitations

This model is experimental and has not been thoroughly calibrated for ethical compliance or legal standards. Caution is advised for sensitive applications.

License

Model weights are licensed under Apache License 2.0.
Users must comply with OpenAI terms of use due to the inclusion of GPT-4-generated synthetic data.

Acknowledgements

This model is based on results obtained from a project, JPNP20017, subsidized by the New Energy and Industrial Technology Development Organization (NEDO).

I would like to acknowledge the use of the following open-source repositories: