New Node: Sharpen node

[Zadfar]

This PR adds a new sharpen node to the Raster: Filter category, allowing users to sharpen images using unsharp masking.

Changes made

• Added new sharpen node with three inputs (amount, radius and threshold)
• Implemented sharpen_algorithm() which uses the unsharp mask technique to sharpen images
• Made a new fn for PremultipliedGammaPixel called to_unpremultiplied_channels(), as this was required by both the sharpening algorithm and gaussian blur algorithm

Notes on implementation

• This implementation borrows heavily from how photoshop did unsharp mask (inputs and ranges), i thought this might be more intuitive for users
• I opted to use the gaussian blur in gamma space as this provided the output most similar to unsharp mask tools from other software.
• There is a hard limit on the threshold variable (255), this comes from the maximum constrast difference in 8bit images
• The soft limits of 100%, 50px and 30 for amount, radius and threshold are arbitrary and can be changed if necessary

Part of #912

[cubic-dev-ai]

1 issue found across 1 file

Confidence score: 4/5

• In node-graph/nodes/raster/src/filter.rs, the threshold fade begins applying sharpening below the configured threshold, so the default threshold of 30 does not behave as an exact cutoff; verify or adjust the fade mask boundaries.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="node-graph/nodes/raster/src/filter.rs">

<violation number="1" location="node-graph/nodes/raster/src/filter.rs:397">
P2: The threshold fade (`threshold * 0.75`) begins applying sharpening below the configured threshold, because the mask only reaches 0 at `diff.abs() <= threshold - fade_width`. At the default threshold of 30 this sharpens pixels that differ by ~8/255, and at high threshold values the deviation grows — contradicting the parameter's documented 'before sharpening is applied' behavior. Consider gating the mask at 0 for `diff.abs() <= threshold` while keeping fade only above the threshold (e.g. ramp from `threshold` to `threshold + fade_width`) if the intent is an unsharp-mask-style hard gate.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: The threshold fade (threshold * 0.75) begins applying sharpening below the configured threshold, because the mask only reaches 0 at diff.abs() <= threshold - fade_width. At the default threshold of 30 this sharpens pixels that differ by ~8/255, and at high threshold values the deviation grows — contradicting the parameter's documented 'before sharpening is applied' behavior. Consider gating the mask at 0 for diff.abs() <= threshold while keeping fade only above the threshold (e.g. ramp from threshold to threshold + fade_width) if the intent is an unsharp-mask-style hard gate.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At node-graph/nodes/raster/src/filter.rs, line 397:

<comment>The threshold fade (`threshold * 0.75`) begins applying sharpening below the configured threshold, because the mask only reaches 0 at `diff.abs() <= threshold - fade_width`. At the default threshold of 30 this sharpens pixels that differ by ~8/255, and at high threshold values the deviation grows — contradicting the parameter's documented 'before sharpening is applied' behavior. Consider gating the mask at 0 for `diff.abs() <= threshold` while keeping fade only above the threshold (e.g. ramp from `threshold` to `threshold + fade_width`) if the intent is an unsharp-mask-style hard gate.</comment>

<file context>
@@ -341,3 +384,40 @@ fn median_quickselect(values: &mut [f32]) -> f32 {
+	let amount = amount / 100.;
+	let threshold = threshold / 255.;
+	// Width of the linear transition around the threshold
+	let threshold_fade_width = threshold * 0.75;
+
+	let sharpen_channel = |orig: f32, blur: f32| -> f32 {
</file context>

[Zadfar]

The unsharp mask style hard gate mentioned here produces output that differs from the output of other software like photoshop and gimp. The threshold_fade_width is a hacky way to get output that looks similar.

[cubic-dev-ai]

All reported issues were addressed across 1 file (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

1 issue found across 1 file

Confidence score: 3/5

• In node-graph/nodes/raster/src/filter.rs, sharpening compares gamma-encoded premultiplied original RGB against un-premultiplied blurred values, which can produce incorrect sharpening results; make the premultiplication and color-space representation consistent before comparison.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="node-graph/nodes/raster/src/filter.rs">

<violation number="1" location="node-graph/nodes/raster/src/filter.rs:417">
P2: The sharpening compares gamma values of the premultiplied original against un-premultiplied blur values. original.to_gamma_srgb_channels() gamma-encodes the premultiplied linear RGB stored in Color, while blurred.to_unpremultiplied_channels() is un-premultiplied. For semi-transparent pixels these representations diverge (the original is scaled by alpha, the blur is not), so the diff and threshold mask are computed on mismatched data, which can produce fringing/halos at soft alpha edges. Un-premultiply the original channels as well before computing the diff, so both sides are unassociated.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: The sharpening compares gamma values of the premultiplied original against un-premultiplied blur values. original.to_gamma_srgb_channels() gamma-encodes the premultiplied linear RGB stored in Color, while blurred.to_unpremultiplied_channels() is un-premultiplied. For semi-transparent pixels these representations diverge (the original is scaled by alpha, the blur is not), so the diff and threshold mask are computed on mismatched data, which can produce fringing/halos at soft alpha edges. Un-premultiply the original channels as well before computing the diff, so both sides are unassociated.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At node-graph/nodes/raster/src/filter.rs, line 417:

<comment>The sharpening compares gamma values of the premultiplied original against un-premultiplied blur values. original.to_gamma_srgb_channels() gamma-encodes the premultiplied linear RGB stored in Color, while blurred.to_unpremultiplied_channels() is un-premultiplied. For semi-transparent pixels these representations diverge (the original is scaled by alpha, the blur is not), so the diff and threshold mask are computed on mismatched data, which can produce fringing/halos at soft alpha edges. Un-premultiply the original channels as well before computing the diff, so both sides are unassociated.</comment>

<file context>
@@ -341,3 +384,47 @@ fn median_quickselect(values: &mut [f32]) -> f32 {
+	};
+
+	for (original, blurred) in buffer.data.iter_mut().zip(&blurred_image.data) {
+		let [original_r, original_g, original_b, original_a] = original.to_gamma_srgb_channels();
+		let [blurred_r, blurred_g, blurred_b, _] = blurred.to_unpremultiplied_channels();
+
</file context>

[Zadfar]

This is how it is handled in gaussian blur. so fixing this here would require a fix there as well. I dont know enough about this particular issue and the chances it will occur while using normally.