diff --git a/out/dice.qoi b/out/dice.qoi
index 9f6317c..ce84029 100644
Binary files a/out/dice.qoi and b/out/dice.qoi differ
diff --git a/out/kodim23.qoi b/out/kodim23.qoi
new file mode 100644
index 0000000..078918d
Binary files /dev/null and b/out/kodim23.qoi differ
diff --git a/src/common.rs b/src/common.rs
index edc0863..a287501 100644
--- a/src/common.rs
+++ b/src/common.rs
@@ -74,25 +74,44 @@ pub fn hash ([r, g, b, a]: [u8; 4]) -> usize {
      (r as usize * 3 + g as usize * 5 + b as usize * 7 + a as usize * 11) % 64
 }
 
-pub fn calc_diff ([r0, g0, b0]: [u8; 3], [r1, g1, b1]: [u8; 3]) -> [i16; 3] {
-    [idiff!(r0, r1), idiff!(g0, g1), idiff!(b0, b1)]
+pub fn calc_diff ([r0, g0, b0, _]: [i16; 4], [r1, g1, b1, _]: [i16; 4]) -> [i16; 3] {
+    [diff(r0, r1), diff(g0, g1), diff(b0, b1)]
 }
 
-pub fn diff_ops (left: [u8; 3], right: [u8; 3]) -> Option<Ops> {
+pub fn diff_ops (left: [u8; 4], right: [u8; 4]) -> Option<Ops> {
+    if left[3] != right[3] { return None; }
+    let left: [i16; 4] = left.map(i16::from);
+    let right: [i16; 4] = right.map(i16::from);
+
     let diff = calc_diff(left, right);
+
     if diff.iter().all(|d| (-2..2).contains(d)) {
         Some(Ops::Diff(diff.map(|e| e as i8)))
-    } else if (-32..32).contains(&diff[1]) && (-8..8).contains(&(diff[0] - diff[1])) && (-8..8).contains(&(diff[2] - diff[1])) {
-        Some(Ops::Luma([(diff[0] - diff[1]) as i8, diff[1] as i8, (diff[2] - diff[1]) as i8]))
+    } else if let Some(luma) = new_luma(diff) {
+        Some(luma)
     } else { None }
 }
 
+pub fn new_luma ([r, g, b]: [i16; 3]) -> Option<Ops> {
+    let drdg = r - g;
+    let dbdg = b - g;
+    if (-32..32).contains(&g) && (-8..8).contains(&drdg) && (-8..8).contains(&dbdg) {
+        return Some(Ops::Luma([drdg as i8, g as i8, dbdg as i8]));
+    }
+    return None;
+}
+
+fn diff (lhs: i16, rhs: i16) -> i16 {
+    let mut diff = lhs - rhs;
+    if diff.abs() > 128 {
+        diff = 256 - rhs + lhs + 1 * match diff.is_negative() {
+            true => 1,
+            false => -1
+        };
+    }
+    diff
+}
+
 pub fn cast_with_bias (num: i8, bias: i8) -> u8 {
     (num + bias) as u8
 }
-
-pub macro idiff {
-    ($t1: expr, $t2: expr) => {
-        $t1 as i16 - $t2 as i16
-    }
-}
diff --git a/src/encoder.rs b/src/encoder.rs
index 7330596..e929e6b 100644
--- a/src/encoder.rs
+++ b/src/encoder.rs
@@ -1,7 +1,5 @@
 use std::mem::discriminant;
-
 use itertools::Itertools;
-
 use crate::common::*;
 
 
@@ -9,11 +7,11 @@ impl Ops {
     // @TODO don't alloc here, get a mut reference and extend here
     fn format (self) -> Vec<u8> {
         // cast_diff
-        let cd = |diff: i8| cast_with_bias(diff, -2);
+        let cd = |diff: i8| cast_with_bias(diff, 2);
         // cast_luma_green
-        let clg = |diff: i8| cast_with_bias(diff, -32);
+        let clg = |diff: i8| cast_with_bias(diff, 32);
         // cast_luma_red_blue
-        let clrb = |diff: i8| cast_with_bias(diff, -8);
+        let clrb = |diff: i8| cast_with_bias(diff, 8);
 
         match self {
             Ops::Nop => vec![],
@@ -45,6 +43,7 @@ fn encode_body (header: Header, data: &[u8]) -> Option<Vec<u8>> {
     let mut last = [0u8, 0, 0, 255];
     let mut out = vec![];
     let mut cur = Ops::Nop;
+    let mut counter = 0;
 
     for mut chunk in &data.iter().chunks(header.channels.num()) {
         let (&r, &g, &b) = chunk.next_tuple()?;
@@ -75,7 +74,7 @@ fn encode_body (header: Header, data: &[u8]) -> Option<Vec<u8>> {
             // INDEX -> DIFF -> LUMA -> RGB(A)
             if known[hash] == rgba {
                 cur = Ops::Index(hash);
-            } else if let Some(ops) = diff_ops([last[0], last[1], last[2]], [r, g, b]) {
+            } else if let Some(ops) = diff_ops([r, g, b, a], [last[0], last[1], last[2], last[3]]) {
                 cur = ops;
             } else {
                 cur = header.channels.ops_color(&rgba[..header.channels.num()])?;
@@ -86,6 +85,11 @@ fn encode_body (header: Header, data: &[u8]) -> Option<Vec<u8>> {
         }
     }
 
+    // if an image buffer ends with a run, flush it as well
+    if discriminant(&cur) == discriminant(&Ops::Run(0)) {
+        out.extend_from_slice(&cur.format());
+    }
+
     Some(out)
 }
 
@@ -129,4 +133,17 @@ mod test {
 
         let _write = std::fs::write("out/1kb.qoi", &encoded);
     }
+
+    #[test]
+    fn kodim23 () {
+        let file = imageproc::image::io::Reader::open("in/kodim23.png").unwrap().decode().unwrap();
+        let file = file.into_rgb8().into_raw();
+        let header = Header { width: 768, height: 512, channels: Channels::RGB, colorspace: Colorspace::Srgb };
+
+
+        let inst = Instant::now();
+        let encoded = encode_full(header, &file).unwrap();
+        println!("kodim23 encoding done in {} ms", inst.elapsed().as_millis());
+        let _write = std::fs::write("out/kodim23.qoi", &encoded);
+    }
 }