xinntao/Real-ESRGAN

Real-ESRGAN is a Python library for real-world image super-resolution that utilizes the SRVGGNetCompact architecture and supports both synthetic degradation and paired data training. It provides utilities for data preprocessing, model conversion to ONNX, and tiled inference for large images, along with GPU acceleration. The system also includes video super-resolution capabilities and deployment support for the Replicate platform.

This change introduces the Real-ESRGAN Python package, providing a new capability for real-world image super-resolution. It includes the \RealESRGANer\ helper class for inference, which supports tiled processing for large images, deep network interpolation (DNI), and GPU acceleration. The package registers new neural network architectures, specifically the \SRVGGNetCompact\ model and a \UNetDiscriminatorSN\, making them available for use via the basicsr registry. A training entry point is also provided to facilitate model training using the registered components.

realesrgan · high confidence

Added five new scripts to support data preprocessing and model deployment: extract\_subimages.py crops large images into overlapping sub-images for faster I/O; generate\_meta\_info.py and generate\_meta\_info\_pairdata.py create metadata text files for single and paired image datasets respectively, with an optional integrity check; generate\_multiscale\_DF2K.py creates multi-scale versions of ground-truth images using LANCZOS resampling; and pytorch2onnx.py converts trained PyTorch models to the ONNX format for broader compatibility.

scripts · high confidence

The repository is established as the Real-ESRGAN project (v0.3.0), replacing the previous generic template. This release introduces \inference\_realesrgan\_video.py\ for video super-resolution and \cog\_predict.py\ to enable deployment on the Replicate platform. The project license is updated from MIT to BSD 3-Clause, and the package name is set to \realesrgan\ for PyPI distribution.

(repo-wide) · high confidence

The data module now includes a new \RealESRGANPairedDataset\ class, enabling users to train the Real-ESRGAN and Real-ESRNet models using pre-existing low-quality and ground-truth image pairs. This complements the existing synthetic degradation pipeline by allowing direct loading of paired folders or LMDB/metadata-backed datasets, with support for standard augmentations like random cropping, flipping, and rotation.

realesrgan/data · high confidence

Added unit tests for the RealESRGAN and RealESRNet datasets (including LMDB and disk backends), model training loops (feed\_data, optimize\_parameters, validation), the UNetDiscriminatorSN architecture, and the RealESRGANer utility (including tile processing and various image type enhancements).

tests · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• First survey — no prior run to compare against. CAI 54.

• Code Health 94
• Architecture 94
• Maturity 40
• Readiness 52
• Security 70

• 138 commits — 127 feature/other, 11 fixes

• (root) — 114 commits
• (repo) — 7 commits
• .github/workflows — 2 commits
• docs/anime_comparisons.md — 2 commits
• options/train_realesrgan_x2plus.yml — 2 commits
• tests/data — 2 commits
• assets/teaser-text.png — 1 commit
• docs/FAQ.md — 1 commit
• docs/anime_video_model.md — 1 commit
• models/realesrgan_model.py — 1 commit
• realesrgan/data — 1 commit
• realesrgan/models — 1 commit
• scripts/extract_subimages.py — 1 commit
• scripts/generate_meta_info.py — 1 commit
• scripts/pytorch2onnx.py — 1 commit

• fix: Merge branch 'generate-meta-info-args-comma-fix' of https://github.com/rogachevai/Real-ESRGAN into rogachevai-generate-meta-info-args-comma-fix
• fix: Merge branch 'rogachevai-generate-meta-info-args-comma-fix'
• fix: add RealESRNet model, fix bug in exe file
• fix: comma fix
• fix: fix bug: extension
• fix: fix colab link error
• fix: fix colorspace bug & support multi-gpu and multi-processing (#312)
• fix: fix ffmpeg framerate bug
• fix: fix file_url bug
• fix: fix import bug in setup.py
• fix: update file organization, fix bug: scale argument in RRDB
• change: Add CODE_OF_CONDUCT.md
• change: Add Linux/MacOS executable files
• change: Add ReadMe for training (#3 )
• change: Add Replicate demo (#428)
• change: Add Training.md Simplified Chinese Version (#139)
• change: Add comparisons for the soon be released animevideo-v3 model (#301)
• change: Added Chinese README (#126)
• change: Added GPU selection feature to python inference (#321)
• change: Create LICENSE
• …and 118 more

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

xinntao/Real-ESRGAN was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.