Image Sharpen KJ(ImageSharpenKJ)

Description

GPU-accelerated image sharpening with multiple methods. Each method uses a different principle to enhance perceived sharpness or recover lost detail.

RCAS — AMD's Robust Contrast-Adaptive Sharpening (from FSR).
Single 5-tap cross filter that adapts to local contrast. Minimal artifacts, good for general use with little tuning.

Adaptive USM — Unsharp mask with local variance modulation.
Sharpens detail-rich areas more, flat/noisy areas less. More controllable than RCAS via radius and threshold parameters.

High-Pass — Extracts high-frequency detail and blends it back.
Gives a "clarity" enhancement feel. Uses radius to control the scale of detail.

Deconvolution — Richardson-Lucy iterative deconvolution.
Can recover actual lost detail from blur, not just enhance edges. Uses radius as the estimated blur kernel and iterations to control convergence.

Inputs

Common Inputs

• image – The input image (any image type). The output will match this type.
• method – Choose the sharpening algorithm from the dropdown: RCAS, Adaptive USM, High-Pass, or Deconvolution. The visible parameters change depending on this selection.

Method-Specific Parameters

RCAS

• strength (Float, 0–1, default 0.8) – 0 = no sharpening, 1 = full RCAS sharpening.

Adaptive USM

• strength (Float, 0–3, default 0.5) – Sharpening multiplier. Values above 1.0 give aggressive sharpening.
• radius (Float, 0.5–5, default 1.0, step 0.1) – Gaussian blur sigma for the unsharp mask. Larger values enhance coarser detail.
• threshold (Float, 0–1, default 0.05) – Noise gate. Higher values only sharpen areas with more texture/detail. 0 = sharpen everything.

High-Pass

• strength (Float, 0–3, default 0.5) – Blend factor for high-frequency detail. Values above 1.0 give a punchier effect.
• radius (Float, 0.5–5, default 1.0, step 0.1) – Gaussian blur sigma defining the frequency cutoff. Larger values enhance coarser detail.

Deconvolution

• strength (Float, 0–1, default 0.5) – Blend between original (0) and fully deconvolved (1).
• radius (Float, 0.5–5, default 1.0, step 0.1) – Sigma of the assumed Gaussian blur to reverse.
• iterations (Integer, 1–100, default 10) – Richardson-Lucy iterations. More iterations produce sharper results but are slower; diminishing returns are noticeable past ~20.

Outputs

• image – The sharpened image, of the same type as the input.

Usage Notes

• All methods run entirely on the GPU for fast processing.
• RCAS is the most "set and forget" option – try it first if you want a quick, artifact-free sharpen.
• Adaptive USM offers more control when you need to suppress noise: increase the threshold to protect flat areas and adjust the radius to match the detail scale you want to enhance.
• High-Pass works well as a clarity or texture enhancement. Use a small radius (0.5–1.5) for fine details and a larger radius for mid-range texture.
• Deconvolution is the most powerful but can produce ringing or halos if you over-iterate or guess the wrong blur kernel. Start with 10–20 iterations and a radius equal to the estimated blur of the original image.