Benchmark Overview

This page summarizes representative benchmark results for the HitPaw Image Enhancement API. It is intended as a technical reference for developers who want to understand where different model families perform best.

For endpoint details and request examples, see the API Reference. For model behavior and visual examples, see Available Models.

What This Page Covers

The benchmark notes below focus on three common evaluation needs:

• Extreme upscaling on degraded inputs
• Recovery quality on real-world noisy images
• Structural stability on geometry-heavy scenes

These results are useful when choosing between generative restoration models and more conservative enhancement models.

Key Takeaways

• generative_4x is the strongest option in this set when the input is severely degraded and needs aggressive 4x reconstruction.
• face_v2_2x is the more texture-faithful portrait option when natural facial detail matters more than soft beautification.
• Different datasets stress different capabilities, so model selection should follow the actual failure mode of the source image rather than a single global score.

Benchmark Highlights

4x Super-Resolution on Challenging Inputs

The table below compares HitPaw generative_4x against Real-ESRGAN on three representative datasets.

Dataset Context	Baseline (Real-ESRGAN)	HitPaw (generative_4x)	Improvement
DIV2K (complex textures)	17.30 PSNR	21.70 PSNR	+4.40 dB (+25.4%)
RealSR (real-world noise)	22.49 PSNR	26.84 PSNR	+4.35 dB
Urban100 (geometric lines)	21.05 PSNR	22.99 PSNR	+1.94 dB

Interpretation:

• On heavily degraded 4x tasks, generative_4x shows the largest gain on texture-rich and real-world data.
• Urban100 still improves, but by a smaller margin, which is typical for geometry-heavy scenes where line stability matters as much as texture synthesis.

Portrait Fidelity

For portrait enhancement, face_v2_2x is the model in this benchmark set that best balances clarity with natural skin texture.

Model	Task	PSNR	SSIM	Practical Meaning
face_v2_2x	Portrait enhance	28.91	0.8148	Preserves facial texture better and avoids an overly smooth result

Interpretation:

• Use face_v2_* when realism, pores, wrinkles, eyelashes, and facial texture retention are important.
• Use face_* instead when the preferred output is softer and more beautified.

How To Read These Results

Benchmark numbers are directional, not absolute guarantees for every input.

• PSNR is helpful for measuring reconstruction accuracy against a reference image.
• SSIM is helpful for judging structural similarity and perceptual consistency.
• In production use, visual quality should still be checked alongside metrics, especially for generative restoration models.

In practice:

• Choose generative_* when the image is heavily compressed, very blurry, or missing detail.
• Choose high_fidelity_* when the source is already fairly good and you want conservative enhancement.
• Choose general_* for standard non-face enhancement.
• Choose face_v2_* or generative_portrait_* for human subjects, depending on how degraded the source is.

Dataset Notes

DIV2K

Best for evaluating texture reconstruction and detail recovery on high-quality reference imagery.

RealSR

Best for evaluating recovery on realistic degradation, including sensor noise and non-ideal capture conditions.

Urban100

Best for evaluating architectural detail, repeated structures, and line stability.

Visual Examples

DIV2K: Extreme Generative Recovery

Focus: reconstructing complex textures under 4x upscaling.

RealSR: Real-World Noise Processing

Focus: handling real-world noise and restoration artifacts.

Urban100: Structural Integrity

Focus: preserving straight lines and architectural structure.

Model Selection Guidance

Use this page together with the API model list when selecting a model:

• For severe low-resolution image recovery: generative_2x or generative_4x
• For realistic portrait recovery: face_v2_2x or face_v2_4x
• For highly degraded portraits: generative_portrait_1x, generative_portrait_2x, or generative_portrait_4x
• For standard non-face enhancement: general_2x or general_4x
• For conservative upscaling of already good images: high_fidelity_2x or high_fidelity_4x

Reference Links

• HitPaw MCP Server on GitHub

Ready to Start?

Get started by Purchasing an API Key Now to unlock full access to the HitPaw Enhancement API.