Table Transformer Structure Recognition
Open Weights • Released 2022-10-14 • Last Verified 2026-08-06
Table Transformer Structure Recognition is a high-performance multimodal vision-language model developed by Microsoft. Integrating advanced visual encoder networks with deep language models, Table Transformer Structure Recognition excels at visual document understanding (DocVQA), chart and diagram parsing, high-resolution optical character recognition (OCR), and spatial reasoning.
Plain English Summary (What is this model & who is it for?)
Think of Table Transformer Structure Recognition as an AI with eyes. You can upload photos, receipts, financial charts, or scanned documents, and ask it to read text, analyze visual contents, or answer questions about what it sees.
💡 Real-World Use Cases & Practical Examples
🚀 How to Run & Use This Model (Step-by-Step Guide)
Simple setup instructions for everyday users and developers.
Download a One-Click App (No Coding Required)
Download a free local AI launcher like LM Studio (lmstudio.ai) or Ollama (ollama.com) on your Mac, Windows, or Linux PC.
Load the Model
In LM Studio, search for "Table Transformer Structure Recognition". In Ollama, open your terminal and run "ollama run microsoft-table-transformer-structure-recognition".
Start Chatting or Generating
Type your text instructions or upload files into the app. The AI runs 100% privately on your hardware without internet requirement!
Developer API Integration
Developers can integrate Table Transformer Structure Recognition directly via Python (using Hugging Face transformers/diffusers) or connect via local OpenAI-compatible REST server (http://localhost:11434).
Benchmark Performance
Hardware Requirements for Local Running
Requires dedicated GPU with 16GB+ VRAM recommended for fast local inference.
Strengths
- •High-Resolution Document VQA & OCR
- •Multi-Chart & Diagram Structural Parsing
- •Spatial Object Detection & Annotation
- •Seamless Text-Image Input Fusion
Limitations & Weaknesses
- •Requires local GPU hardware for self-hosting
from transformers import AutoProcessor, AutoModelForVision2Seq
from PIL import Image
import torch
model_id = "microsoft/table-transformer-structure-recognition"
model = AutoModelForVision2Seq.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)
image = Image.open("sample.jpg")
inputs = processor(text="Analyze the contents of this image:", images=image, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=150)
print(processor.batch_decode(outputs, skip_special_tokens=True)[0])Pricing Overview
Prices subject to provider tiers and volume discounts. Check documentation for current token rates.
Model Tags
Did this model work for you?
Your feedback helps others find the right model.