Spaces:

lsmpp
/

PixText-Studio

Running on Zero

App Files Files Community

PixText-Studio / README.md

lsmpp

Detect and recognize Chinese glyphs individually from CRAFT region maps while preserving English word boxes

9d5e109 verified 1 day ago

preview code

Raw History Blame Contribute Delete

2.23 kB

---
title: PixText Studio
emoji: ✍️
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 6.30.0
python_version: 3.12
app_file: space.py
license: apache-2.0
models:
  - ATH-MaaS/PixText
---
# PixText Studio

Detect text, adjust its boxes, enter the text you want, and compare the repaired regions. CRAFT locates the text and PP-OCRv6 reads it. Chinese uses individual character boxes; English stays in word boxes.

Upload an image or choose the English and Chinese examples. Select only the regions you want to repair. You can move or resize any box, or draw a new one. MeanFlow defaults to four steps; Teacher defaults to 25. The result includes an image slider, matching region crops, a PNG download and editable region JSON.

## Run

Deploy this directory as a Gradio Space with ZeroGPU, or run locally on an Ampere-or-newer CUDA GPU:

```sh
python -m pip install -r requirements.txt
python space.py
```

Model packages download from [ATH-MaaS/PixText](https://huggingface.co/ATH-MaaS/PixText) when first used. The optional OCR files download on the first detection run and are checked against release SHA-256 hashes. For an existing installation, set `PIXTEXT_MODEL_ROOT` to the directory containing the `PixText-MeanFlow` and `PixText-Teacher` packages and `PIXTEXT_OCR_ROOT` to the OCR model directory. Both variables are optional.

Jobs are queued one at a time. Images are stored in isolated temporary sessions and removed after one hour of inactivity, when the next session is created. Nothing is uploaded to a dataset or sent to an external inference service.

The vendored inference code is from [PixText commit 7779419](https://github.com/yaoliliu/PixText/tree/7779419de634d9785ec23558942480db713f0f99). Model packages retain their manifests, font notices and SHA-256 checks. CRAFT weights come from [EasyOCR](https://github.com/JaidedAI/EasyOCR); recognition weights and configuration come from [PaddlePaddle](https://huggingface.co/PaddlePaddle/PP-OCRv6_medium_rec_onnx).

The Space uses Gradio Server to serve the custom editor and queue GPU requests. On ZeroGPU, PyTorch SDPA uses the original PixText token-routing mask without runtime compilation. A Dockerfile is also included for dedicated GPU hosting.

	---
	title: PixText Studio
	emoji: ✍️
	colorFrom: yellow
	colorTo: red
	sdk: gradio
	sdk_version: 6.30.0
	python_version: 3.12
	app_file: space.py
	license: apache-2.0
	models:
	- ATH-MaaS/PixText
	---

	# PixText Studio

	Detect text, adjust its boxes, enter the text you want, and compare the repaired regions. CRAFT locates the text and PP-OCRv6 reads it. Chinese uses individual character boxes; English stays in word boxes.

	Upload an image or choose the English and Chinese examples. Select only the regions you want to repair. You can move or resize any box, or draw a new one. MeanFlow defaults to four steps; Teacher defaults to 25. The result includes an image slider, matching region crops, a PNG download and editable region JSON.

	## Run

	Deploy this directory as a Gradio Space with ZeroGPU, or run locally on an Ampere-or-newer CUDA GPU:

	```sh
	python -m pip install -r requirements.txt
	python space.py
	```

	Model packages download from [ATH-MaaS/PixText](https://huggingface.co/ATH-MaaS/PixText) when first used. The optional OCR files download on the first detection run and are checked against release SHA-256 hashes. For an existing installation, set `PIXTEXT_MODEL_ROOT` to the directory containing the `PixText-MeanFlow` and `PixText-Teacher` packages and `PIXTEXT_OCR_ROOT` to the OCR model directory. Both variables are optional.

	Jobs are queued one at a time. Images are stored in isolated temporary sessions and removed after one hour of inactivity, when the next session is created. Nothing is uploaded to a dataset or sent to an external inference service.

	The vendored inference code is from [PixText commit 7779419](https://github.com/yaoliliu/PixText/tree/7779419de634d9785ec23558942480db713f0f99). Model packages retain their manifests, font notices and SHA-256 checks. CRAFT weights come from [EasyOCR](https://github.com/JaidedAI/EasyOCR); recognition weights and configuration come from [PaddlePaddle](https://huggingface.co/PaddlePaddle/PP-OCRv6_medium_rec_onnx).

	The Space uses Gradio Server to serve the custom editor and queue GPU requests. On ZeroGPU, PyTorch SDPA uses the original PixText token-routing mask without runtime compilation. A Dockerfile is also included for dedicated GPU hosting.