infer_unlimited_ocr

Unlimited-OCR by Baidu parses long document images into Markdown in a single OCR pass. It builds on the DeepSeek-OCR family and provides two single-image inference presets: Gundam for cropped high-throughput parsing and Base for standard full-image parsing.

The upstream Hugging Face inference code currently targets NVIDIA CUDA GPUs.

🚀 Use with Ikomia API

1. Install Ikomia API

We strongly recommend using a virtual environment. If you're not sure where to start, we offer a tutorial here.

pip install ikomia

2. Create your workflow

from ikomia.dataprocess.workflow import Workflow
from ikomia.utils.displayIO import display

# Init your workflow
wf = Workflow()

# Add algorithm
algo = wf.add_task(name="infer_unlimited_ocr", auto_connect=True)

# Run on your image
wf.run_on(url="https://raw.githubusercontent.com/NanoNets/Nanonets-OCR2/refs/heads/main/assets/bank_statement.jpg")

# Display input
display(algo.get_input(0).get_image())

# Display OCR text boxes
img_output = algo.get_output(0)
text_output = algo.get_output(1)
display(img_output.get_image_with_graphics(text_output), title="Unlimited OCR")

# Get cleaned OCR text
unlimited_output = algo.get_output(2)
print(unlimited_output.data["response"])

# Save output JSON
unlimited_output.save("unlimited_ocr_output.json")

☀️ Use with Ikomia Studio

Ikomia Studio offers a friendly UI with the same features as the API.

• If you haven't started using Ikomia Studio yet, download and install it from this page.
• For additional guidance on getting started with Ikomia Studio, check out this blog post.

📝 Set algorithm parameters

• model_name (string, default: baidu/Unlimited-OCR): Hugging Face model repo or local model path to load.
• cuda (bool, default: auto): Use CUDA. The upstream Hugging Face inference path requires an NVIDIA CUDA GPU.
• prompt (string, default: document parsing.): Text instruction appended after the <image> token.
• mode (enum, default: Gundam): Single-image preset.
  • Gundam: base_size=1024, image_size=640, crop_mode=True.
  • Base: base_size=1024, image_size=1024, crop_mode=False.
• max_length (int, default: 32768): Maximum generation length.
• no_repeat_ngram_size (int, default: 35): Repetition guard used by the model generation call. Set to 0 to disable.
• ngram_window (int, default: 128): Sliding window used with no_repeat_ngram_size. Set to 0 to use the preset default.

from ikomia.dataprocess.workflow import Workflow
from ikomia.utils.displayIO import display

# Init your workflow
wf = Workflow()

# Add algorithm
algo = wf.add_task(name="infer_unlimited_ocr", auto_connect=True)

algo.set_parameters({
    "prompt": "document parsing.",
    "mode": "Gundam",
    "max_length": "32768",
    "no_repeat_ngram_size": "35",
    "ngram_window": "128",
})

# Run on your image
wf.run_on(url="https://raw.githubusercontent.com/NanoNets/Nanonets-OCR2/refs/heads/main/assets/bank_statement.jpg")

# Show input
display(algo.get_input(0).get_image())

# Display OCR text boxes
img_output = algo.get_output(0)
text_output = algo.get_output(1)
display(img_output.get_image_with_graphics(text_output), title="Unlimited OCR")

# Save cleaned OCR text
unlimited_output = algo.get_output(2)
unlimited_output.save("unlimited_ocr_output.json")

🔍 Explore algorithm outputs

The plugin returns two outputs:

• algo.get_output(1): CTextIO fields with text and bounding boxes for display in Ikomia.
• algo.get_output(2): DataDictIO with cleaned OCR text.

response is cleaned for printing: model detection tags such as <|det|>text [x1, y1, x2, y2]<|/det|> are removed. Bounding boxes are available in the CTextIO output.

Do not print algo.get_output(1).data["response"]: output 1 is CTextIO and does not have a data attribute. Use algo.get_output(2).data["response"] for text.

from ikomia.dataprocess.workflow import Workflow
from ikomia.utils.displayIO import display

# Init your workflow
wf = Workflow()

# Add algorithm
algo = wf.add_task(name="infer_unlimited_ocr", auto_connect=True)

# Run on your image
wf.run_on(url="https://raw.githubusercontent.com/NanoNets/Nanonets-OCR2/refs/heads/main/assets/bank_statement.jpg")

# Display OCR text boxes
img_output = algo.get_output(0)
text_output = algo.get_output(1)
display(img_output.get_image_with_graphics(text_output), title="Unlimited OCR")

# Clean printable OCR text
data_output = algo.get_output(2)
print(data_output.data["response"])

⏩ Advanced usage

Tips for best results

• Use Gundam first for most document images.
• Try Base for pages where crop mode misses global layout relationships.
• Keep the default repetition settings for long documents; lower or disable them only if generation is unexpectedly truncated.
• Ensure images are sharp, upright, and high enough resolution for small text.

Prompt examples

• document parsing.
• Convert the document to markdown.
• Extract tables and text from this document.
• OCR this image.