glsl-smaa

glsl-smaa

SMAA (Enhanced Subpixel Morphological Antialiasing) post-processing; WebGL (OpenGL ES 2.0) implementation with no fluff. All credit goes to the incredible work here (iryoku) and there (beakbeak).

Looking for WebGPU? See wgsl-smaa, which ships the same passes in WGSL.

Installation

npm install glsl-smaa

Usage

Three passes are required to apply the effect:

             |input|------------------·
                v                     |
      [ SMAA*EdgeDetection ]          |
                v                     |
            |edgesTex|                |
                v                     |
[ SMAABlendingWeightCalculation ]     |
                v                     |
            |blendTex|                |
                v                     |
  [ SMAANeighborhoodBlending ] <------·
                v
             |output|

You would typically set this up as part of a post-processing chain with framebuffers. The edges and weights targets must be cleared to [0, 0, 0, 0] every frame, and sampled with a linear filter.

See the example and its source for a setup with pex-context.

import {
  SMAATextures,
  PRESETS,
  SMAA_EDGES_VERT,
  SMAA_EDGES_FRAG,
} from "glsl-smaa";

const frag = /* glsl */ `#define SMAA_PRESET_HIGH
#define SMAA_EDGES_LUMA
${PRESETS}
${SMAA_EDGES_FRAG}`;

Each shader also ships as a plain file (eg. glsl-smaa/smaa-edges.frag) for bundlers that load shaders as raw text.

Lookup textures

SMAATextures.area and SMAATextures.search are the reference's precomputed textures as PNG data URIs. Upload them as is (no flipY), with linear filtering and no mipmaps.

Before 4.0.0, the search texture was stored flipped and had to be uploaded with flipY: true. It now matches the reference data: remove the flip.

Attributes

Attributes required for all vertex shaders:

Name	Type	Description
aPosition	vec2	Screen vertices (eg. a fullscreen triangle/quad)

Uniforms and defines

#define can be prepended to the shaders to override default settings. Typically:

const frag = /* glsl */ `#define SMAA_PRESET_HIGH
${PRESETS}
${SMAA_WEIGHTS_FRAG}
`;

smaa-edges.(vert|frag)

You can use one of the 3 following edge detection method:

• Depth: the fastest but it may miss some edges.

• Luma: more expensive than depth edge detection, but catches visible edges that depth edge detection can miss.

• Color: the most expensive one but catches chroma-only edges.

Defines:

# One of:
#define SMAA_EDGES_DEPTH
#define SMAA_EDGES_LUMA
#define SMAA_EDGES_COLOR

#define SMAA_THRESHOLD 0.1
#define SMAA_DEPTH_THRESHOLD (0.1 * SMAA_THRESHOLD)
#define SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR 2.0
#define SMAA_PREDICATION 0
#define SMAA_PREDICATION_THRESHOLD 0.01
#define SMAA_PREDICATION_SCALE 2.0
#define SMAA_PREDICATION_STRENGTH 0.4

# Optional:
#define SMAA_SRGB_INPUT

Name	Default [Range]	Description
SMAA_THRESHOLD	0.1 [0, 0.5]	Specifies the threshold or sensitivity to edges. Lowering this value you will be able to detect more edges at the expense of performance. 0.1 is a reasonable value, and allows to catch most visible edges. 0.05 is a rather overkill value, that allows to catch 'em all.
SMAA_DEPTH_THRESHOLD	0.1 * SMAA_THRESHOLD	Specific threshold for depth edge detection.
SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR	2.0	If there is an neighbor edge that has SMAA_LOCAL_CONTRAST_FACTOR times bigger contrast than current edge, current edge will be discarded. This allows to eliminate spurious crossing edges, and is based on the fact that, if there is too much contrast in a direction, that will hide perceptually contrast in the other neighbors.
SMAA_PREDICATION	0	Locally decreases the luma or color threshold if an edge is found in the depth buffer (so the global threshold can be higher).
SMAA_PREDICATION_THRESHOLD	0.01	Threshold used in the depth buffer. Depends on its content.
SMAA_PREDICATION_SCALE	2.0 [1, 5]	How much to scale the global threshold used for luma or color edge detection when using predication.
SMAA_PREDICATION_STRENGTH	0.4 [0, 1]	How much to locally decrease the threshold.
SMAA_SRGB_INPUT	undefined	Luma and color edge detection expect gamma-encoded input. Define it when sampling an sRGB texture, which decodes to linear, to re-encode the color before thresholding.

Uniforms:

Name	Type	Description
uTexelSize	vec2	1 / framebuffer size.
uColorTexture	sampler2D	The input color framebuffer. Luma and color edge detection only.
uDepthTexture	sampler2D	The input depth framebuffer. Depth edge detection and predication.

---

smaa-weights.(vert|frag)

Defines from presets.glsl:

# Use presets:
#define SMAA_PRESET_LOW
#define SMAA_PRESET_MEDIUM
#define SMAA_PRESET_HIGH
#define SMAA_PRESET_ULTRA

# or individually setting:
#define SMAA_THRESHOLD 0.1
#define SMAA_MAX_SEARCH_STEPS 16
#define SMAA_MAX_SEARCH_STEPS_DIAG 8
#define SMAA_CORNER_ROUNDING 25

# and optionally disable diagonal and corner detection:
#define SMAA_DISABLE_DIAG_DETECTION
#define SMAA_DISABLE_CORNER_DETECTION

Name	Default [Range]	Description
SMAA_THRESHOLD	0.1 [0, 0.5]	Should be the same as for the edge pass.
SMAA_MAX_SEARCH_STEPS	16 [0, 112]	Specifies the maximum steps performed in the horizontal/vertical pattern searches, at each side of the pixel. In number of pixels, it's actually the double. So the maximum line length perfectly handled by, for example 16, is 64 (by perfectly, we meant that longer lines won't look as good, but still antialiased.
SMAA_MAX_SEARCH_STEPS_DIAG	8 [0, 20]	Specifies the maximum steps performed in the diagonal pattern searches, at each side of the pixel. In this case we jump one pixel at time, instead of two.
SMAA_CORNER_ROUNDING	25 [0, 100]	Specifies how much sharp corners will be rounded.

Uniforms:

Name	Type	Description
uTexelSize	vec2	1 / framebuffer size.
uViewportSize	vec2	The framebuffer size.
uSubsampleIndices	vec4	[0, 0, 0, 0] for SMAA 1x. See Temporal supersampling.
uEdgesTexture	sampler2D	The framebuffer with edges.
uAreaTexture	sampler2D	SMAATextures.area.
uSearchTexture	sampler2D	SMAATextures.search.

---

smaa-blend.(vert|frag)

Defines:

#define SMAA_REPROJECTION 0
#define SMAA_DECODE_VELOCITY(sample) sample.rg

Name	Default	Description
SMAA_REPROJECTION	0	Packs the antialiased velocity into the output alpha, for the temporal resolve.
SMAA_DECODE_VELOCITY	sample.rg	Decodes the velocity texture. Velocity is in texture coordinates, from previous to current frame.

Uniforms:

Name	Type	Description
uTexelSize	vec2	1 / framebuffer size.
uColorTexture	sampler2D	The input color framebuffer.
uBlendTexture	sampler2D	The framebuffer with weights.
uVelocityTexture	sampler2D	The velocity framebuffer. SMAA_REPROJECTION only.

---

smaa-resolve.(vert|frag)

Optional temporal resolve pass, blending the current and previous frames.

Defines: SMAA_REPROJECTION, SMAA_DECODE_VELOCITY as above, and SMAA_REPROJECTION_WEIGHT_SCALE (30.0 [0, 80]): the higher, the sooner a moving pixel stops blending with its history.

Name	Type	Description
uCurrentColorTexture	sampler2D	Output of the current frame's blend pass.
uPreviousColorTexture	sampler2D	Output of the previous frame's blend pass.
uVelocityTexture	sampler2D	The velocity framebuffer. SMAA_REPROJECTION only.

---

Temporal supersampling (T2x)

1. Jitter the camera projection by a subpixel offset alternating every frame.
2. Run the three passes on each frame, setting uSubsampleIndices to the index matching the frame's jitter.
3. Resolve the current and previous blend outputs with smaa-resolve. With SMAA_REPROJECTION, both the blend and the resolve passes read a velocity texture, and the previous frame is reprojected.

Frame	Camera jitter (pixels, y up)	uSubsampleIndices
0	(0.25, 0.25)	[1, 1, 1, 0]
1	(-0.25, -0.25)	[2, 2, 2, 0]

WebGL's bottom-left texture origin runs the passes on a vertically mirrored image, so the jitter's y is the opposite of the reference table (see @SUBSAMPLE_INDICES in SMAA.hlsl).

License

Copyright (C) 2013 Jorge Jimenez (jorge@iryoku.com) Copyright (C) 2013 Jose I. Echevarria (joseignacioechevarria@gmail.com) Copyright (C) 2013 Belen Masia (bmasia@unizar.es) Copyright (C) 2013 Fernando Navarro (fernandn@microsoft.com) Copyright (C) 2013 Diego Gutierrez (diegog@unizar.es)

MIT. See license file.