realesr-general-x4v3 Model Repository

This repository packages the official realesr-general-x4v3 checkpoints, shared provenance and quality fixtures, and reproducible exports for deployment runtimes. Format-specific artifacts, tools, documentation, and measurements are namespaced so additional backends can be added without changing the source model definition.

Model Definition

The published model is realesr-general-x4v3-dn0.5: native 4× super-resolution using deep network interpolation (DNI) 0.5 between two unmodified official Real-ESRGAN checkpoints:

realesr-general-x4v3.pth — standard restoration model.
realesr-general-wdn-x4v3.pth — weak-denoise model.

model-manifest.json records both checkpoint hashes, upstream URLs, the approved ONNX graph, scale, architecture, and golden-fixture hashes. Derived formats must preserve this model identity and pass their documented PyTorch-reference quality gates.

Available Formats

Format	Precision	Artifact	Validated target	Documentation	Measurements
TensorRT	FP32 with default TF32 tactics	`tensorrt_fp32/`	NVIDIA L4	Build and usage guide	Performance report

TensorRT plans are hardware and runtime specific. Consult the format guide before deploying a prebuilt artifact. Future formats should follow docs/adding-a-format.md.

Repository Layout

assets/                 shared sample image and numerical quality fixtures
docs/                   format, hardware, and contribution guides
reports/                measured results and machine-readable benchmark data
scripts/<format>/       conversion, validation, and benchmark entry points
<format_artifacts>/     deployable models, manifests, and checksums
model-manifest.json     canonical source provenance and approved graph metadata
*.pth                   original upstream PyTorch checkpoints
pyproject.toml          format-specific uv dependency groups
uv.lock                 reproducible Python dependency resolution

Large checkpoints, engines, ONNX graphs, fixtures, and generated images use Git LFS on the Hugging Face Hub.

Integrity and Quality Policy

Treat the root checkpoints and model-manifest.json as immutable, content-addressed source material.
Give each deployable format its own manifest, checksums, validation report, and reproducible conversion entry point.
Compare optimized output with the approved PyTorch FP32 reference before publishing performance results.
Record hardware, runtime versions, precision, latency methodology, and quality metrics alongside every benchmark.
Never replace an existing artifact silently; publish a new version and update the format table.

Getting Started

Install uv and Git LFS, clone the repository, and select a format from the table above. Each format guide lists its hardware requirements, dependency group, verification command, conversion procedure, and reproducible benchmark. Source checkpoint integrity can be checked without installing a GPU runtime:

uv run --locked --no-dev python -m scripts.verify_bundle

License and Upstream

Real-ESRGAN and the bundled checkpoints are distributed under the BSD 3-Clause license. See LICENSE and the upstream Real-ESRGAN repository.

Downloads last month: -; Downloads are not tracked for this model. How to track