AI image upscaling research is measured through a combination of image quality, model size, computational cost, and performance on controlled or real-world images. The statistics below show how those measures differ across benchmark settings, including 4K test images, camera-captured pairs, and standard super-resolution datasets.
Table of contents
- What efficient upscaling benchmarks measure
- Model size across super-resolution systems
- PSNR results on Set5 and Set14
- PSNR results on BSD100 and Urban100
- Why the image source matters
- Efficiency targets in the NTIRE 2024 challenge
What efficient upscaling benchmarks measure
The Efficient Perceptual Image Super Resolution: AIM 2025 Study and Benchmark placed explicit limits on the systems being compared. Model size was capped at 5 million parameters, while compute was capped at 2,000 GFLOPs. These constraints matter for people choosing an upscaler for ordinary hardware or a fast editing workflow: a model can be judged not only by its output, but also by the resources required to produce it.
The benchmark used inference inputs fixed at 960 × 540 pixels. Its test set contained 500 images at 4K resolution, and the benchmark applied multiple degradation types per image. That combination creates a demanding evaluation: the input dimensions are fixed for comparison, the reference images are high resolution, and each image can represent more than one kind of quality loss.
The benchmark’s top-performing approach outperformed Real-ESRGAN across all benchmark datasets, according to the same Efficient Perceptual Image Super Resolution: AIM 2025 Study and Benchmark. This is a benchmark result, not a universal ranking of every image or editing workflow. It does, however, show why efficiency-focused comparisons need to report both the operating limits and the datasets used.
Model size across super-resolution systems
The paper Transcending the Limit of Local Window listed parameter counts for ten systems in a cross-model super-resolution comparison. EDSR was listed at 43.0 million parameters, while RCAN had 15.6 million and SAN had 15.9 million. HAN was listed at 64.2 million parameters, and IPT was the largest of the listed systems at 116 million.
Several newer entries were substantially smaller than IPT. SwinIR was listed at 11.9 million parameters and EDT at 11.6 million. CAT-A and ART were each listed at 16.6 million parameters, while HAT was listed at 20.8 million. These figures describe the models in that comparison; they do not, by themselves, indicate which model will produce the most pleasing result on a particular photograph.
| Model | Parameters listed | Set5 PSNR at x3 |
|---|---|---|
| EDSR | 43.0M | 34.65 dB |
| RCAN | 15.6M | 34.74 dB |
| SwinIR | 11.9M | 34.97 dB |
| CAT-A | 16.6M | 35.06 dB |
| HAT | 20.8M | 35.07 dB |
The compact comparison illustrates why parameter count and image quality should be read together. In this table, the smallest listed model is not the lowest-scoring one on Set5 at x3, and the largest listed model is not automatically the only high-scoring option. The data supports comparison within the stated test, not a claim that parameter count alone determines quality.
PSNR results on Set5 and Set14
Peak signal-to-noise ratio, or PSNR, is reported in decibels. In the Transcending the Limit of Local Window comparison, all of the results in this section used an x3 enlargement setting. On Set5, EDSR scored 34.65 dB, RCAN 34.74 dB, SAN 34.75 dB, and HAN 34.75 dB. IPT scored 34.81 dB, while SwinIR and EDT each scored 34.97 dB. CAT-A scored 35.06 dB, and ART and HAT each scored 35.07 dB.
| Model | Set5 at x3 | Set14 at x3 |
|---|---|---|
| EDSR | 34.65 dB | 30.52 dB |
| RCAN | 34.74 dB | 30.65 dB |
| SAN | 34.75 dB | 30.59 dB |
| HAN | 34.75 dB | 30.67 dB |
| IPT | 34.81 dB | 30.85 dB |
| SwinIR | 34.97 dB | 30.93 dB |
| EDT | 34.97 dB | 30.89 dB |
| CAT-A | 35.06 dB | 31.04 dB |
| ART | 35.07 dB | 31.02 dB |
| HAT | 35.07 dB | 31.08 dB |
On Set14 at x3, EDSR scored 30.52 dB and RCAN 30.65 dB. SAN scored 30.59 dB, HAN 30.67 dB, and IPT 30.85 dB. SwinIR scored 30.93 dB, EDT 30.89 dB, CAT-A 31.04 dB, ART 31.02 dB, and HAT 31.08 dB. Within this reported comparison, HAT had the highest listed Set14 score, while HAT and ART shared the highest listed Set5 score.
These results are useful as controlled reference points for photo-editing readers. They do not mean that a 0.01 dB difference will be visible in every image. Texture, edges, noise, faces, and compression artifacts can affect how an enlarged image looks, while PSNR summarizes numerical similarity to a reference image.
PSNR results on BSD100 and Urban100
The same comparison reported results for BSD100. At x3, EDSR scored 29.25 dB, RCAN 29.32 dB, SAN 29.33 dB, and HAN 29.32 dB. IPT scored 29.38 dB, SwinIR 29.46 dB, and EDT 29.44 dB. CAT-A scored 29.52 dB, ART 29.51 dB, and HAT 29.54 dB.
Urban100 produced a wider separation among the listed scores. EDSR scored 28.80 dB, RCAN 29.09 dB, SAN 28.93 dB, and HAN 29.10 dB. IPT scored 29.49 dB. SwinIR scored 29.75 dB and EDT 29.72 dB. CAT-A scored 30.12 dB, ART 30.10 dB, and HAT 30.23 dB.
| Model | BSD100 at x3 | Urban100 at x3 |
|---|---|---|
| EDSR | 29.25 dB | 28.80 dB |
| RCAN | 29.32 dB | 29.09 dB |
| SAN | 29.33 dB | 28.93 dB |
| HAN | 29.32 dB | 29.10 dB |
| IPT | 29.38 dB | 29.49 dB |
| SwinIR | 29.46 dB | 29.75 dB |
| EDT | 29.44 dB | 29.72 dB |
| CAT-A | 29.52 dB | 30.12 dB |
| ART | 29.51 dB | 30.10 dB |
| HAT | 29.54 dB | 30.23 dB |
HAT had the highest listed score on both BSD100 and Urban100 in this table. Urban100 also shows why dataset identity matters: the scores are not interchangeable with Set5 or Set14 results. Each dataset represents a different test context, so a model’s position can be discussed only alongside the dataset name and enlargement setting.
Why the image source matters
Synthetic degradation benchmarks are not the only way to study upscaling. Toward Real-World Single Image Super-Resolution: A New Benchmark and a New Model described RealSR as using paired low-resolution and high-resolution images captured by adjusting a digital camera’s focal length. This setup links the two resolutions through camera capture rather than relying only on a generated degradation process.
The RealSR dataset was built with Canon 5D3 and Nikon D810 cameras. The model trained on RealSR generalized to Sony a7II devices and to mobile phones, according to the same source. Those device details are important for interpreting the result: a model trained with one set of cameras was evaluated for transfer to other capture hardware, including phones.
For a photo-editing workflow, this distinction changes what a benchmark can tell you. A controlled test with fixed input dimensions can reveal how systems compare under repeatable conditions. Camera-paired data can reveal whether a method remains useful when lens, sensor, processing, and capture conditions vary. Neither type of result should be treated as a guarantee for every image.
Efficiency targets in the NTIRE 2024 challenge
The Ninth NTIRE 2024 Efficient Super-Resolution Challenge Report described targets for the efficient super-resolution challenge. The target was about 26.90 dB PSNR on DIV2K_LSDIR_valid and about 26.99 dB PSNR on DIV2K_LSDIR_test. These are challenge targets associated with named validation and test sets, not general-purpose image-quality promises.
The report said that the main track optimized three metrics: runtime, FLOPs, and parameter count. That combination reflects a practical trade-off. Runtime describes how quickly an operation completes, FLOPs describe computational work, and parameter count describes model scale. A useful upscaler may need to balance all three instead of maximizing a single score.
Taken together, the benchmark statistics point to four questions for anyone comparing AI image upscaling tools: Which images were tested? What enlargement setting was used? What PSNR, parameter, or compute limits were reported? And was the evaluation based on synthetic degradation or camera-captured pairs? Keeping those details attached to each number makes the results more useful for real photo editing decisions.