zenquant 0.1.2

AQ-informed palette quantizer — perceptual masking meets color quantization
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283

extern crate alloc;
use alloc::vec;
use alloc::vec::Vec;

#[allow(unused_imports)]
use num_traits::Float;

use crate::oklab::srgb_to_oklab_l_fast;

/// Compute per-pixel AQ masking weights from an RGB image.
///
/// Returns weights in [0.1, 1.0] where:
/// - High weight (≈1.0) = smooth region → protect quality, allocate palette entries
/// - Low weight (≈0.1) = textured region → error is masked, allow more quantization
pub fn compute_masking_weights(pixels: &[rgb::RGB<u8>], width: usize, height: usize) -> Vec<f32> {
    let luminance = extract_luminance(pixels);
    let contrast = compute_local_contrast(&luminance, width, height);
    let block_w = width.div_ceil(4);
    let block_h = height.div_ceil(4);
    let block_masking = erode_to_blocks(&contrast, width, height, block_w, block_h);
    let per_pixel = upscale_bilinear(&block_masking, block_w, block_h, width, height);
    masking_to_weights(&per_pixel)
}

/// Compute masking weights from pre-computed OKLab values.
///
/// Extracts L from `labs[i].l` instead of converting sRGB per pixel.
pub fn compute_masking_weights_from_labs(
    labs: &[crate::oklab::OKLab],
    width: usize,
    height: usize,
) -> Vec<f32> {
    let luminance: Vec<f32> = labs.iter().map(|lab| lab.l).collect();
    let contrast = compute_local_contrast(&luminance, width, height);
    let block_w = width.div_ceil(4);
    let block_h = height.div_ceil(4);
    let block_masking = erode_to_blocks(&contrast, width, height, block_w, block_h);
    let per_pixel = upscale_bilinear(&block_masking, block_w, block_h, width, height);
    masking_to_weights(&per_pixel)
}

/// Same as above but for RGBA input.
pub fn compute_masking_weights_rgba(
    pixels: &[rgb::RGBA<u8>],
    width: usize,
    height: usize,
) -> Vec<f32> {
    let luminance = extract_luminance_rgba(pixels);
    let contrast = compute_local_contrast(&luminance, width, height);
    let block_w = width.div_ceil(4);
    let block_h = height.div_ceil(4);
    let block_masking = erode_to_blocks(&contrast, width, height, block_w, block_h);
    let per_pixel = upscale_bilinear(&block_masking, block_w, block_h, width, height);
    masking_to_weights(&per_pixel)
}

/// Extract OKLab L (lightness) channel from RGB pixels.
/// Uses fast_cbrt (~3x faster than f32::cbrt, ~22 bits precision).
fn extract_luminance(pixels: &[rgb::RGB<u8>]) -> Vec<f32> {
    pixels
        .iter()
        .map(|p| srgb_to_oklab_l_fast(p.r, p.g, p.b))
        .collect()
}

fn extract_luminance_rgba(pixels: &[rgb::RGBA<u8>]) -> Vec<f32> {
    pixels
        .iter()
        .map(|p| srgb_to_oklab_l_fast(p.r, p.g, p.b))
        .collect()
}

/// Compute local contrast: (L - avg_4_neighbors)², clamped to [0, 0.2].
fn compute_local_contrast(luminance: &[f32], width: usize, height: usize) -> Vec<f32> {
    let mut contrast = vec![0.0f32; luminance.len()];

    for y in 0..height {
        for x in 0..width {
            let idx = y * width + x;
            let center = luminance[idx];

            let mut sum = 0.0f32;
            let mut count = 0u32;

            if x > 0 {
                sum += luminance[idx - 1];
                count += 1;
            }
            if x + 1 < width {
                sum += luminance[idx + 1];
                count += 1;
            }
            if y > 0 {
                sum += luminance[idx - width];
                count += 1;
            }
            if y + 1 < height {
                sum += luminance[idx + width];
                count += 1;
            }

            let avg = if count > 0 {
                sum / count as f32
            } else {
                center
            };
            let diff = center - avg;
            contrast[idx] = (diff * diff).min(0.2);
        }
    }

    contrast
}

/// Min-biased erosion: for each block, take the weighted average of the 4 smallest contrast values.
/// Weights: [0.40, 0.25, 0.20, 0.15] — heavier min-bias for banding protection.
fn erode_to_blocks(
    contrast: &[f32],
    width: usize,
    height: usize,
    block_w: usize,
    block_h: usize,
) -> Vec<f32> {
    const WEIGHTS: [f32; 4] = [0.40, 0.25, 0.20, 0.15];

    let mut blocks = vec![0.0f32; block_w * block_h];

    for by in 0..block_h {
        for bx in 0..block_w {
            // Gather all contrast values in this 4x4 block (max 16 values)
            let mut values = [0.0f32; 16];
            let mut count = 0usize;
            let y_start = by * 4;
            let x_start = bx * 4;
            let y_end = (y_start + 4).min(height);
            let x_end = (x_start + 4).min(width);

            for y in y_start..y_end {
                for x in x_start..x_end {
                    values[count] = contrast[y * width + x];
                    count += 1;
                }
            }

            if count == 0 {
                continue;
            }

            // Sort ascending to find smallest values
            values[..count]
                .sort_unstable_by(|a, b| a.partial_cmp(b).unwrap_or(core::cmp::Ordering::Equal));

            // Weighted average of up to 4 smallest values
            let n = count.min(4);
            let mut weighted_sum = 0.0f32;
            let mut weight_sum = 0.0f32;
            for i in 0..n {
                weighted_sum += values[i] * WEIGHTS[i];
                weight_sum += WEIGHTS[i];
            }

            blocks[by * block_w + bx] = weighted_sum / weight_sum;
        }
    }

    blocks
}

/// Bilinear upscale from block grid to per-pixel resolution.
fn upscale_bilinear(
    blocks: &[f32],
    block_w: usize,
    block_h: usize,
    width: usize,
    height: usize,
) -> Vec<f32> {
    let mut output = vec![0.0f32; width * height];

    for y in 0..height {
        for x in 0..width {
            // Map pixel center to block grid coordinates
            // Block centers are at (bx * 4 + 2, by * 4 + 2)
            let bx_f = (x as f32 - 2.0) / 4.0;
            let by_f = (y as f32 - 2.0) / 4.0;

            let bx0 = (bx_f.floor() as isize).max(0) as usize;
            let by0 = (by_f.floor() as isize).max(0) as usize;
            let bx1 = (bx0 + 1).min(block_w - 1);
            let by1 = (by0 + 1).min(block_h - 1);

            let fx = (bx_f - bx0 as f32).clamp(0.0, 1.0);
            let fy = (by_f - by0 as f32).clamp(0.0, 1.0);

            let v00 = blocks[by0 * block_w + bx0];
            let v10 = blocks[by0 * block_w + bx1];
            let v01 = blocks[by1 * block_w + bx0];
            let v11 = blocks[by1 * block_w + bx1];

            let top = v00 * (1.0 - fx) + v10 * fx;
            let bot = v01 * (1.0 - fx) + v11 * fx;
            output[y * width + x] = top * (1.0 - fy) + bot * fy;
        }
    }

    output
}

/// Convert masking values to weights.
/// Low contrast (smooth) → high weight, high contrast (texture) → low weight.
/// K is tuned so typical images have mean weight ~0.5.
fn masking_to_weights(masking: &[f32]) -> Vec<f32> {
    const K: f32 = 8.0;

    masking
        .iter()
        .map(|&m| {
            let w = 0.1 + 0.9 / (1.0 + K * m.sqrt());
            w.clamp(0.1, 1.0)
        })
        .collect()
}

#[cfg(test)]
mod tests {
    use super::*;

    fn flat_image(r: u8, g: u8, b: u8, width: usize, height: usize) -> Vec<rgb::RGB<u8>> {
        vec![rgb::RGB { r, g, b }; width * height]
    }

    #[test]
    fn flat_image_high_weights() {
        let pixels = flat_image(128, 128, 128, 16, 16);
        let weights = compute_masking_weights(&pixels, 16, 16);
        assert_eq!(weights.len(), 256);
        // All pixels identical → zero contrast → weight should be ~1.0
        for &w in &weights {
            assert!(w > 0.95, "expected high weight for flat image, got {w}");
        }
    }

    #[test]
    fn checkerboard_low_weights() {
        let mut pixels = Vec::with_capacity(16 * 16);
        for y in 0..16 {
            for x in 0..16 {
                if (x + y) % 2 == 0 {
                    pixels.push(rgb::RGB { r: 0, g: 0, b: 0 });
                } else {
                    pixels.push(rgb::RGB {
                        r: 255,
                        g: 255,
                        b: 255,
                    });
                }
            }
        }
        let weights = compute_masking_weights(&pixels, 16, 16);
        assert_eq!(weights.len(), 256);
        let mean: f32 = weights.iter().sum::<f32>() / weights.len() as f32;
        // High contrast checkerboard → low mean weight
        assert!(
            mean < 0.5,
            "expected low mean weight for checkerboard, got {mean}"
        );
    }

    #[test]
    fn weights_in_valid_range() {
        let mut pixels = Vec::with_capacity(32 * 32);
        for i in 0..(32 * 32) {
            let v = (i % 256) as u8;
            pixels.push(rgb::RGB { r: v, g: v, b: v });
        }
        let weights = compute_masking_weights(&pixels, 32, 32);
        for &w in &weights {
            assert!(
                (0.1..=1.0).contains(&w),
                "weight {w} out of range [0.1, 1.0]"
            );
        }
    }
}