Arm/esrgan-int8-xnnpack-executorch-graviton-g4

RRDB_PSNR_x4 optimized for Arm-based Cloud CPU

RRDB_PSNR_x4, a 4x image super-resolution model, optimized for Arm-based Cloud CPU systems and delivered as an ExecuTorch .pte artifact.

Summary

This repository contains an Arm-optimized version of kadirnar/RRDB_PSNR_x4 for 4x image super-resolution. The model is provided in the ExecuTorch .pte format running on the ExecuTorch runtime, targeting Cloud CPU systems.

This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on Urban100 and measured performance on a representative evaluation target.

The optimization is a mixed-precision INT8 recipe: the bulk of the RRDB backbone is quantized to INT8, while the upsampling tail and the first three RRDB body blocks are held in FP32, and post-conversion graph surgery removes quantize/dequantize pairs that the XNNPACK quantizer inserts around concatenation nodes inside those FP32 blocks.

Key results

Area	Result
Model format	ExecuTorch .pte
Target device class	Cloud CPU
Reference device	AWS Graviton G4 (Neoverse-V2, Ubuntu 24.04.4 LTS)
Primary performance result	704.42 ms p50 latency per 128x128 tile, 1.42 frames/sec
Accuracy result	26.81 dB PSNR on Urban100
Size / memory result	23.176 MB, 2.76x smaller than the FP32 baseline

Original model

Field	Value
Original model	kadirnar/RRDB_PSNR_x4
Original source	Hugging Face
Original developer	ESRGAN (Xintao Wang et al.)
Original model card	kadirnar/RRDB_PSNR_x4
Original license	Apache-2.0

Model files

File	Description
esrgan_x4_graviton_executorch_optimized.pte	Arm-optimized model for deployment
example.py	Minimal inference example
pyproject.toml	Pinned runtime dependencies for example.py, resolved with uv
uv.lock	Locked dependency resolution for pyproject.toml
config.yaml	Model I/O contract used by the example
benchmarks/	FP32 baseline and Arm-optimized benchmark records

Performance

Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.

Reference configuration

Field	Value
Device / platform	AWS Graviton G4
CPU / accelerator	Neoverse-V2, 16 cores, 32 GB system memory, CPU execution
OS	Linux Ubuntu 24.04.4 LTS
Runtime	ExecuTorch 1.1.0
Backend / delegate	XNNPACK, KleidiAI
Batch size	1
Precision	INT8 static PTQ — per-channel symmetric weights, per-tensor affine activations (mixed precision: upsampling tail and first three RRDB body blocks held in FP32)
Runs	10 warmup + 100 measured

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
p50 latency	1762.73 ms	704.42 ms	2.50x faster
p90 latency	1764.72 ms	779.70 ms	2.26x faster
p99 latency	1771.11 ms	819.72 ms	2.16x faster
Throughput	0.57 frames/sec	1.42 frames/sec	2.49x
Model load time	66.29 ms	36.67 ms	1.81x faster
Time to first inference	1968.65 ms	906.88 ms	2.17x faster
Model size	64.073 MB	23.176 MB	2.76x smaller
Peak memory	799.74 MB	795.07 MB	1.01x less

Latency is measured per 128x128 input tile at a 4x scale factor.

Accuracy

Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.

Evaluation setup

Field	Value
Dataset	Urban100
Split	N/A
Number of samples	100
Metric(s)	PSNR (dB), SSIM
Evaluation runtime	ExecuTorch

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
PSNR (dB)	27.03	26.81	-0.22
SSIM	0.8176	0.8109	-0.0067

Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.

For this release, Arm used:

Optimization area	Applied?	Notes
Model conversion	Yes	Converted to the ExecuTorch .pte format for the ExecuTorch runtime
Quantization	Yes	Mixed-precision INT8 static PTQ: per-channel symmetric weights and per-tensor affine activations for the quantized portion of the RRDB backbone, calibrated on 100 randomly selected Urban100 samples; the upsampling tail and the first three RRDB body blocks are held in FP32 to keep accuracy within acceptable ranges
Runtime/backend selection	Yes	XNNPACK delegate with KleidiAI kernels, CPU execution backend
Graph/runtime compatibility updates	Yes	Performed as part of the ExecuTorch export pipeline: post-conversion graph surgery removes the quantize/dequantize pairs the XNNPACK quantizer inserts around concatenation nodes inside the blocks that are held in FP32
Accuracy validation	Yes	Compared against the original model or published baseline
Performance validation	Yes	Measured on the reference Arm platform

The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.

Using this model

Install dependencies

Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:

uv python install
uv sync --frozen

Run the example

uv run example.py

The ordinary command runs inference without writing files. To save the generated image and summary, provide explicit output paths:

uv run example.py --output output.png --summary-output summary.json

Expected input

Property	Value
Input shape	[1, 3, 128, 128]
Input type	float32
Input range	[0.0, 1.0]
Preprocessing	Convert an RGB image to a [0, 1] float32 tensor; images larger than one tile are split into 128x128 tiles automatically

Expected output

Property	Value
Output shape	[1, 3, 512, 512]
Output type	float32
Postprocessing	Clamp to [0.0, 1.0] at a 4x spatial scale factor; adjacent tiles overlap by 8 pixels and are blended by averaging

Intended use

This model is intended for developers evaluating 4x image super-resolution workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

• Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
• With the pinned runtime, TorchAO reports that its optional C++ extensions are incompatible with the installed Torch version, ExecuTorch reports that InternalConsistency verification is unavailable, and CPUinfo may fall back from /sys/devices/soc0/image_version to MIDR. These warnings did not affect XNNPACK registration, model loading, or inference on the reference platform.
• Accuracy was evaluated on Urban100 and may not generalize to all domains.
• This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
• This repository is not a replacement for the original model documentation.

Additional notes

The INT8 recipe is mixed precision, and both parts of it are required together:

• Layers held in FP32: the upsampling tail (upconv1, upconv2, HRconv, conv_last), which directly synthesizes output pixel values, and the first three RRDB body blocks (body.0, body.1, body.2), whose early feature representations span a wide dynamic range. The exclusion filter matches exact path components, so body.1 does not also select body.10 through body.19.
• Graph surgery: the XNNPACK quantizer annotates every concatenation node unconditionally, which bypasses the per-layer exclusion filter. Without removing those quantize/dequantize pairs after conversion, convolutions downstream of the concatenation nodes inside body.0 through body.2 still receive INT8 activations despite the blocks being nominally FP32, and PSNR is measurably worse.

Calibration uses 100 randomly sampled Urban100 images as 128x128 center crops, matching the model's export tile size.

• Sample input: sample_input.jpg is derived from Living room (Unsplash) by Jarosław Ceborski, via Wikimedia Commons (CC0 1.0).

About this version

Original Model: kadirnar/RRDB_PSNR_x4 by ESRGAN (Xintao Wang et al.) - Repository

Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.

Converted/optimized by: Arm

License: The Original Model and the Optimized Model are subject to Apache-2.0.

This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.

No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.

Original Model and Documentation

For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.

Licenses and Third-Party Terms

Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

Purpose of this Release

The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.

Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.

To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.

You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.

Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.