Skip to content
LOPS Docs

OCR Operator

v0.2.1New

Overview

Section titled “Overview”

The OCR LOP extracts text from images using Optical Character Recognition. It sends image data to the vision_sidecar HTTP service for processing with either EasyOCR or PaddleOCR, keeping heavy ML inference out of TouchDesigner’s main process. Processing runs asynchronously so TouchDesigner stays responsive while OCR is in progress.

Requirements

Section titled “Requirements”
  • SideCar must be available with the vision_sidecar service. The operator communicates with it over HTTP and will attempt to start the service automatically if it is not already running.
  • The selected OCR library must be installed in the SideCar’s Python environment:
    • PaddleOCR: paddleocr, paddlepaddle-gpu
    • EasyOCR: easyocr
  • On first use, if the required library is missing, the operator will prompt you to install it (including PyTorch with CUDA support on Windows if needed).

Input/Output

Section titled “Input/Output”

Inputs

Section titled “Inputs”
  • Input 1 (TOP): The image to run OCR on. Can also be set via the Input TOP parameter on the OCR page.

Outputs

Section titled “Outputs”
  • Output 1 (DAT): Results table with columns for timestamp, text, u, v, width, height, and confidence. All coordinates are normalized (0-1).
  • Output 2 (TOP): The input image overlaid with bounding boxes around detected text regions. Line width and color are configurable under the Display / TOP Out2 section on the OCR page.

Usage

Section titled “Usage”
  1. Connect an image TOP to the OCR operator’s input (or set it via the Input TOP parameter)
  2. On the OCR page, select a Model Type — PaddleOCR or EasyOCR
  3. Adjust Min Confidence to filter out low-confidence detections
  4. Optionally lower Image Scale for faster processing on large images
  5. Pulse Process
  6. The Active indicator turns on while processing runs asynchronously
  7. Results appear in the output table and the Output Text parameter once complete. The second TOP output shows detected regions visually.

Enable Combine Text to join all detected text blocks into a single string in Output Text. When off, only the first detected block is shown.

Best Practices

Section titled “Best Practices”
  • Image Scale: For large images, lowering Image Scale reduces processing time without significantly impacting text detection quality. Start at 1.0 and reduce if processing is slow.
  • Model selection: PaddleOCR generally provides better accuracy for structured documents. EasyOCR supports more languages out of the box.
  • Confidence filtering: Raise Min Confidence to eliminate noisy detections in complex scenes. A value around 0.5 is a good starting point.

Troubleshooting

Section titled “Troubleshooting”
  • “No input TOP specified”: Make sure an image TOP is connected to the operator’s input or set via the Input TOP parameter.
  • “ML Server not available”: The vision_sidecar service could not be reached. The operator will queue the task and retry. Check that the SideCar is running and healthy.
  • Slow first run: The OCR model is loaded into memory on the first request. Subsequent requests are significantly faster.

Parameters

Section titled “Parameters”

OCR

Section titled “OCR”
Process (Process) op('ocr').par.Process Pulse
Default:
False
Input TOP (Inputtop) op('ocr').par.Inputtop TOP
Default:
"" (Empty String)
Status (Status) op('ocr').par.Status Str
Default:
"" (Empty String)
Active (Active) op('ocr').par.Active Toggle
Default:
False
Model Type (Modeltype) op('ocr').par.Modeltype Menu
Default:
paddle
Options:
paddle, easy
Min Confidence (Minconfidence) op('ocr').par.Minconfidence Float
Default:
0.0
Range:
0 to 1
Slider Range:
0 to 1
Image Scale (Imagescale) op('ocr').par.Imagescale Float
Default:
0.0
Range:
0.1 to 1
Slider Range:
0.1 to 1
Combine Text (Combinetext) op('ocr').par.Combinetext Toggle
Default:
False
Output Text (Outputtext) op('ocr').par.Outputtext Str
Default:
"" (Empty String)
Display / TOP Out2 Header
Line Width (Width) op('ocr').par.Width Float
Default:
0.0
Range:
0 to 1
Slider Range:
0 to 1
Color (Colorr) op('ocr').par.Colorr RGB
Default:
0.0
Range:
0 to 1
Slider Range:
0 to 1
Color (Colorg) op('ocr').par.Colorg RGB
Default:
0.0
Range:
0 to 1
Slider Range:
0 to 1
Color (Colorb) op('ocr').par.Colorb RGB
Default:
0.0
Range:
0 to 1
Slider Range:
0 to 1

Callbacks

Section titled “Callbacks”
Callbacks Header
Callback DAT (Callbackdat) op('ocr').par.Callbackdat DAT
Default:
ChatTD_callbacks
Edit Callbacks (Editcallbacksscript) op('ocr').par.Editcallbacksscript Pulse
Default:
False
Create Callbacks (Createpulse) op('ocr').par.Createpulse Pulse
Default:
False
onComplete (Oncomplete) op('ocr').par.Oncomplete Toggle
Default:
False
Textport Debug Callbacks (Debugcallbacks) op('ocr').par.Debugcallbacks Menu
Default:
Full Details
Options:
None, Errors Only, Basic Info, Full Details

Callbacks

Section titled “Callbacks”
Available Callbacks:
  • onComplete
Example Callback Structure: 
def onComplete(info):
# Called when OCR processing completes
# info contains:
#   text: list of detected text strings
#   results: list of dicts with text, confidence, u, v, width, height
#   processing_time: server processing time in seconds
combined = " ".join(info.get('text', []))
print(f"OCR found {len(info['results'])} text blocks: {combined[:100]}")
pass

Changelog

Section titled “Changelog”
v0.2.12026-03-26
  • Added OCR processing via vision_sidecar HTTP API - Supports Easy OCR and Paddle OCR models - Implemented results table and text output options
v0.2.02026-03-01
  • Refactor to call ml_server HTTP API directly instead of SideCar methods - Async image processing via TDAsyncIO - Base64 image encoding for HTTP transmission
  • Fix TD 32050+ freeze by using importlib.metadata for torch check - SideCar handles actual OCR inference
  • Initial commit
Section titled “Related Operators”