Release Notes
Update 2025.07.25: V1.1.4 Released
Major Changes:
- Upgraded the Mathematical Formula Detection (MFD) and Mathematical Formula Recognition (MFR) models to version 1.5. All default configurations, documentation, and examples now use
mfd-1.5andmfr-1.5as the standard models.
主要变更:
- 数学公式检测(MFD)和数学公式识别(MFR)模型升级到 1.5 版本,所有默认配置、文档和示例均以
mfd-1.5和mfr-1.5为标准模型。
Update 2025.05.06: V1.1.3.2 Released
Major Changes:
- Fixed a potential error when processing transparent images, see #171 for details.
主要变更:
- 修复了处理透明图片时可能出现的错误,具体见 #171 。
Update 2025.04.27: V1.1.3.1 Released
Major Changes:
- Bugfix: Fixed the issue of model import related to VLM.
主要变更:
- 修复了 VLM 相关的模型导入问题。
Update 2025.04.15: V1.1.3 Released
Major Changes:
- Support for
VlmTableOCRandVlmTextFormulaOCRmodels based on the VLM interface (see LiteLLM documentation) allowing the use of closed-source VLM models. Installation command:pip install pix2text[vlm].- Usage examples can be found in tests/test_vlm.py and tests/test_pix2text.py.
主要变更:
- 支持基于 VLM 接口(具体参考 LiteLLM 文档)的
VlmTableOCR和VlmTextFormulaOCR模型,可使用闭源 VLM 模型。安装命令:pip install pix2text[vlm]。- 使用方式见 tests/test_vlm.py 和 tests/test_pix2text.py。
Update 2024.12.17: V1.1.2.3 Released
Major Changes:
- Bugfix: Fixed issues related to downloading models on Windows.
主要变更:
- 修复了在 Windows 环境下下载模型的问题。
Update 2024.12.11: V1.1.2.2 Released
Major Changes:
- Bugfix: Resolved issues related to serialization errors when handling ONNX Runtime session options by ensuring that non-serializable configurations are managed appropriately.
主要变更:
- 修复了与 ONNX Runtime session options 相关的序列化错误,通过确保不可序列化的配置信息在适当的管理下进行处理。
Update 2024.12.02: V1.1.2.1 Released
Major Changes:
- Fixed an error in
fetch_column_info()@DocYoloLayoutParser, thanks to Bin.
主要变更:
- 修复了 fetch_column_info()@DocYoloLayoutParser 中的错误,感谢网友 Bin 。
Update 2024.11.17: V1.1.2 Released
Major Changes:
- A new layout analysis model DocLayout-YOLO has been integrated, improving the accuracy of layout analysis.
- Bug fixes:
- When the text language is set to English only, a dedicated English OCR model is used to avoid including Chinese in the output.
- The processing logic for PNG images has been optimized, enhancing recognition performance.
主要变更:
- 版面分析模型加入 DocLayout-YOLO,提升版面分析的准确性。
- 修复 bugs:
- 在设置文本语言只有英语时,使用专门的英文 OCR 模型,避免输出中包含中文。
- 对 PNG 图片的处理逻辑进行了优化,提升了识别效果。
Update 2024.07.18: V1.1.1.2 Released
Major Changes:
- fix bugs:
主要变更:
- 修复 bugs:
Update 2024.06.24: V1.1.1.1 Released
Major Changes:
- Added a new parameter
static_resized_shapewhen initializingMathFormulaDetector, which is used to resize the input image to a fixed size. Some formats of models require fixed-size input images during inference, such asCoreML.
主要变更:
MathFormulaDetector初始化时加入了参数static_resized_shape, 用于把输入图片 resize 为固定大小。某些格式的模型在推理时需要固定大小的输入图片,如CoreML。
Update 2024.06.18: V1.1.1 Released
Major changes:
- Support the new mathematical formula detection models (MFD): breezedeus/pix2text-mfd (Mirror), which significantly improves the accuracy of formula detection.
See details: Pix2Text V1.1.1 Released, Bringing Better Mathematical Formula Detection Models | Breezedeus.com.
主要变更:
- 支持新的数学公式检测模型(MFD):breezedeus/pix2text-mfd (国内镜像),公式检测精度获得较大提升。
具体说明请见:Pix2Text V1.1.1 发布,带来更好的数学公式检测模型 | Breezedeus.com。
Update 2024.06.17:V1.1.0.7 Released
Major changes:
- adapted with cnstd>=1.2.4, thanks to @g1y5x3 .
主要变更:
- 适配 cnstd>=1.2.4 ,感谢 @g1y5x3 。
Update 2024.06.04:V1.1.0.6 Released
Major changes:
- Fix: The Text OCR incorrectly carried over the configuration from previous calls when it was called multiple times.
主要变更:
- 修复 bug:Text OCR 多次调用时错误沿用了之前的配置信息。
Update 2024.05.27:V1.1.0.5 Released
Major changes:
- Fixed bugs such as that in
._parse_remaining.
主要变更:
- 修复
._parse_remaining等 bug。
Update 2024.05.20:V1.1.0.4 Released
Major changes:
- set
table_as_imageasTrueifself.table_ocris not available. - fix typo: https://github.com/breezedeus/Pix2Text/pull/108 . Thanks to @billvsme.
主要变更:
- 如果
self.table_ocr不可用,将table_as_image设置为True。 - 修复拼写错误:https://github.com/breezedeus/Pix2Text/pull/108 。感谢 @billvsme。
Update 2024.05.19:V1.1.0.3 Released
Major changes:
- A new paid model,
mfr-plus, has been added, which offers better recognition for multi-line formulas. - When recognizing only English, CnOCR does not output Chinese.
- Bugs have been fixed.
主要变更:
- 加入新的付费模型:
mfr-plus,对多行公式的识别效果更好。 - 在只识别英文时,CnOCR 不输出中文。
- 修复 bugs。
Update 2024.05.10:V1.1.0.2 Released
Major changes:
- Fixed the error caused by empty lines in
merge_line_texts.
主要变更:
- 修复
merge_line_texts中空行导致的错误。
Update 2024.04.30:V1.1.0.1 Released
Major changes:
- Fix the exception occurring when saving files on Windows.
主要变更:
- 修复 Windows 下存储文件时出现的异常。
Update 2024.04.28:V1.1 Released
Major changes:
- Added layout analysis and table recognition models, supporting the conversion of images with complex layouts into Markdown format. See examples: Pix2Text Online Documentation / Examples.
- Added support for converting entire PDF files to Markdown format. See examples: Pix2Text Online Documentation / Examples.
- Enhanced the interface with more features, including adjustments to existing interface parameters.
- Launched the Pix2Text Online Documentation.
主要变更:
- 加入了版面分析和表格识别模型,支持把复杂排版的图片转换为 Markdown 格式,示例见:Pix2Text 在线文档/Examples。
- 支持把整个 PDF 文件转换为 Markdown 格式,示例见:Pix2Text 在线文档/Examples。
- 加入了更丰富的接口,已有接口的参数也有所调整。
- 上线了 Pix2Text 在线文档。
Update 2024.03.30:V1.0.2.3 Released
Major changes:
- Fixed the issue caused by
merge_line_texts, see details at: https://github.com/breezedeus/Pix2Text/issues/84. - Optimized the post-processing logic to handle some abnormal sequences.
主要变更:
- 修复
merge_line_texts带来的错误,具体见:https://github.com/breezedeus/Pix2Text/issues/84 。 - 优化了后处理逻辑,处理部分不正常的序列。
Update 2024.03.18:V1.0.2.2 Released
Major changes:
- The previously used
output_logitsargument is incompatible with transformers < 4.38.0, replaced by theoutput_scoresargument. https://github.com/breezedeus/Pix2Text/issues/81 - Fixed a bug in
serve.pythat was not compatible with the new pix2text version.
主要变更:
- 之前使用的
output_logits参数不兼容 transformers < 4.38.0,换为output_scores参数。 https://github.com/breezedeus/Pix2Text/issues/81 - 修复
serve.py中未兼容新版接口的 bug。
Update 2024.03.15:V1.0.2.1 Released
Major Changes:
- Fixed mishandling of LaTeX expressions during post-processing, such as replacing
\rightarrowwitharrow. - Added
rec_configparameter to.recognize_text()and.recognize_formula()methods for passing additional parameters for recognition.
主要变更:
- 修复对 LaTeX 表达式进行后处理时引入的误操作,如
\rightarrow被替换为arrow。 - 对
.recognize_text()和.recognize_formula()加入了rec_config参数,以便传入用于识别的额外参数。
Update 2024.03.14:V1.0.2 Released
Major Changes:
- Optimized the recognition process, improving the recognition of boundary punctuation that may have been missed before.
- Enhanced the LaTeX recognition results by restoring the formula tags to the formulas.
- Adjusted the output format of the recognition results, adding the
return_textparameter to control whether to return only text or more detailed information. When returning more detailed information, confidence scorescoreand position informationpositionwill also be provided. Thanks to @hiroi-sora for the suggestion: https://github.com/breezedeus/Pix2Text/issues/67.
主要变更:
- 优化了识别的逻辑,以前可能漏识的边界标点现在可以比较好的识别。
- 对 Latex 识别结果进行了优化,把公式的 tag 还原到公式中。
- 调整了识别结果的输出格式,增加了参数
return_text来控制结果是只返回文本还是更丰富的信息。当返回更丰富信息时,会返回置信度score以及位置信息position。感谢 @hiroi-sora 的建议:https://github.com/breezedeus/Pix2Text/issues/67 。
Update 2024.03.03:发布 V1.0.1
主要变更:
- 修复在 CUDA 环境下使用
LatexOCR时出现的错误,具体见:https://github.com/breezedeus/Pix2Text/issues/65#issuecomment-1973037910 ,感谢 @MSZ-006NOC。
Update 2024.02.26:发布 V1.0
主要变更:
- 数学公式识别(MFR)模型使用新架构,在新的数据集上训练,获得了 SOTA 的精度。具体说明请见:Pix2Text V1.0 新版发布:最好的开源公式识别模型 | Breezedeus.com。
Update 2024.01.10:发布 V0.3
主要变更:
支持识别
80+种语言,详细语言列表见 支持的语言列表;模型自动下载增加国内站点;
优化对检测 boxes 的合并逻辑。
Update 2023.12.21:发布 V0.2.3.3
主要变更:
- fix: bugfixed from @hiroi-sora , thanks much.
Update 2023.09.10:发布 V0.2.3.2
主要变更:
- fix: 去掉
consts.py无用的CATEGORY_MAPPINGS。
Update 2023.07.14:发布 V0.2.3.1
主要变更:
- 修复了
self.recognize_by_clf返回结果中不包含line_number字段导致merge_line_texts报错的bug。
Update 2023.07.03:发布 V0.2.3
主要变更:
- 优化了对检测出的boxes的排序逻辑,以及对混合图片的处理逻辑,使得最终识别效果更符合直觉。具体参考:Pix2Text 新版公式识别模型 | Breezedeus.com 。
- 修复了模型文件自动下载的功能。HuggingFace似乎对下载文件的逻辑做了调整,导致之前版本的自动下载失败,当前版本已修复。但由于HuggingFace国内被墙,国内下载仍需 梯子(VPN)。
- 更新了各个依赖包的版本号。
Update 2023.06.20:发布新版 MFD 模型
主要变更:
- 基于新标注的数据,重新训练了 MFD YoloV7 模型,目前新模型已部署到 P2T网页版 。具体说明见:Pix2Text (P2T) 新版公式检测模型 | Breezedeus.com 。
- 之前的 MFD YoloV7 模型已开放给星球会员下载,具体说明见:P2T YoloV7 数学公式检测模型开放给星球会员下载 | Breezedeus.com 。
Update 2023.02.19:发布 V0.2.2.1
主要变更:
- 修复bug。
Update 2023.02.19:发布 V0.2.2
主要变更:
- 修复旋转框导致的识别结果错误;
- 去掉代码中不小心包含的
breakpoint()。
[Yanked] Update 2023.02.19:发布 V0.2.1
主要变更:
Update 2023.02.03:发布 V0.2
主要变更:
- 利用 CnSTD 新版的数学公式检测(Mathematical Formula Detection,简称 MFD)能力,P2T V0.2 支持识别既包含文字又包含公式的混合图片。
Update 2022.10.21:发布 V0.1.1
主要变更:
- Fix: remove the character which causes error on Windows
Update 2022.09.11:发布 V0.1
- 初版发布