[Feature Request] Allow specifying custom default directory for downloaded models / document default model download path

[cqhung1412]

Summary

The existing issue #10234 asks “Where will the downloaded model be saved?” but ends without a clear resolution or documented behavior. The lack of documentation or config to override default download location is a pain point for users with limited home disk space or custom deployment scenarios.

I propose we add:

1. a documented default path (for all model types: detection, recognition, classification, layout, etc.)

2. a configuration option (argument, env var, or config file) to override that default, so users can redirect model downloads to a custom directory.

Background / Context

• In the current paddleocr wheel package, the parameters det_model_dir, rec_model_dir, cls_model_dir default to None, which triggers auto-download. As per PyPI metadata, when det_model_dir=None the builtin model is automatically downloaded to ~/.paddleocr/det (and similarly for ~/.paddleocr/rec and ~/.paddleocr/cls)
• But that behavior is not clearly documented in the user-facing docs, nor is there a way to change where those get saved (other than manually supplying those paths).
• There's a related issue: How to disable .paddleocr directory creation? (How to disable .paddleocr directory creation? #12123) which acknowledges that when you don't specify a model path, PaddleOCR will download the official model by default.

So configuring a stable, overrideable download path is a real need.

Proposal

In YAML config file or directly in pipeline init function, add a new attribute for general_model_dir (or another name)

from paddleocr import PPStructureV3

ocr_engine = PPStructureV3(general_model_dir="~/another/folder")

If this field is not None and other _dir fields are None, then download the models to this folder, with model dir creation of course.

Request

• Please reopen or link with Query: Default location of downloaded model files #10234, or forward this to the maintainers, so that this feature is tracked.
• It would be super helpful to have this documented and implemented as part of the core, so users don’t have to hack a workaround every time.

[liuhongen1234567]

Hello, you can upgrade PaddleOCR to version 3.0 or above. In PaddleOCR 3.0, you can specify the overall download directory for the model using the following command:

import os
os.environ["PADDLE_PDX_CACHE_HOME"] = r"XXX"

This command needs to be written before from paddleocr import PaddleOCR.