PaddleOCR-VL が公開されたので Gradio と Transformers を使ってOCRアプリを作る

実行画面

日本語もそこそこいけるようです。

Pythonスクリプト

from transformers import AutoProcessor, AutoModelForImageTextToText, TextIteratorStreamer
import gradio as gr
from threading import Thread

model = AutoModelForImageTextToText.from_pretrained(
    "PaddlePaddle/PaddleOCR-VL",
    dtype="bfloat16",
    device_map="cuda"
)

processor = AutoProcessor.from_pretrained("PaddlePaddle/PaddleOCR-VL")

def ocr_image(image):
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "url": image},
                {"type": "text", "text": "OCR:"},
            ]
        }
    ]

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

    streamer = TextIteratorStreamer(processor.tokenizer, skip_prompt=True, skip_special_tokens=False)

    generation_kwargs = dict(
        **inputs,
        max_new_tokens=100,
        streamer=streamer
    )

    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()

    partial_text = ""
    for new_text in streamer:
        partial_text += new_text
        clean_text = partial_text.replace("</s>", "").strip()
        yield clean_text

with gr.Blocks() as demo:
    gr.Interface(
        fn=ocr_image,
        inputs=gr.Image(type="filepath", label="画像をアップロード"),
        outputs=gr.Textbox(lines=10, max_lines=40, buttons=["copy"]),
        flagging_mode="never"
    )

demo.launch()

Python環境構築

pyproject.tomlを載せておきます。 （バージョンはあえて固定しています）

uvを使うとuv syncだけで環境構築できると思います。

[project]
name = "paddleocr"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = ">=3.13"
dependencies = [
    "accelerate==1.12.0",
    "einops==0.8.1",
    "gradio==6.1.0",
    "hf-xet==1.2.0",
    "torch==2.9.1+cu126",
    "transformers @ git+https://github.com/huggingface/transformers",
]

[[tool.uv.index]]
name = "torch-cuda"
url = "https://download.pytorch.org/whl/cu126"
explicit = true

[tool.uv.sources]
torch = [{ index = "torch-cuda" }]

補足

Flash Attention 2も使えました。

from threading import Thread

import gradio as gr
from transformers import (
    AutoModelForImageTextToText,
    AutoProcessor,
    TextIteratorStreamer,
)

model = AutoModelForImageTextToText.from_pretrained(
    "PaddlePaddle/PaddleOCR-VL",
    dtype="bfloat16",
    device_map="cuda",
    attn_implementation="flash_attention_2",
)

processor = AutoProcessor.from_pretrained("PaddlePaddle/PaddleOCR-VL")


def ocr_image(image):
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "url": image},
                {"type": "text", "text": "OCR:"},
            ],
        }
    ]

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

    streamer = TextIteratorStreamer(
        processor.tokenizer, skip_prompt=True, skip_special_tokens=False
    )

    generation_kwargs = dict(**inputs, max_new_tokens=1024, streamer=streamer)

    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()

    partial_text = ""
    for new_text in streamer:
        partial_text += new_text
        clean_text = partial_text.replace("</s>", "").strip()
        yield clean_text


with gr.Blocks() as demo:
    gr.Interface(
        fn=ocr_image,
        inputs=gr.Image(type="filepath", label="画像をアップロード"),
        outputs=gr.Textbox(lines=10, max_lines=40, buttons=["copy"]),
        flagging_mode="never",
    )

demo.launch()