diff --git a/README.md b/README.md
index 315d006..d0a9651 100644
--- a/README.md
+++ b/README.md
@@ -16,6 +16,16 @@ Fast encoder/decoder for [QOI image format](https://qoiformat.org/), implemented
 - `no_std` support.
 - Roundtrip-tested vs the reference C implementation; fuzz-tested.
 
+### Note about this fork
+
+This fork implements a slight improvement to the original specs, which leaves unused the `QOI_OP_RGBA` chunk flag with RGB.
+
+Here, we use this flag for the new `QOI_OP_RUN2` chunk. It's like the `QOI_OP_RUN` chunk, but followed by two bytes representing `run` (BE). (only for RGB, as the flag is already used for RGBA)
+
+The decoder remains fully compatible with the original one (except when using `QOI_OP_RGBA` in a RGB image). The encoder is fully compatible for RGBA, not for RGB (except using the `reference` feature).
+
+Why this? Because it enables significant improvements for compressing images with large uniform areas (such as screen captures), or for encoding a diff-filtered video stream where successive frames often have identical regions. (see [syeve](https://framagit.org/ZettaScript/syeve) for the video encoding)
+
 ### Examples
 
 ```rust
diff --git a/src/consts.rs b/src/consts.rs
index 2281c24..0e39c4b 100644
--- a/src/consts.rs
+++ b/src/consts.rs
@@ -4,6 +4,7 @@ pub const QOI_OP_LUMA: u8 = 0x80; // 10xxxxxx
 pub const QOI_OP_RUN: u8 = 0xc0; // 11xxxxxx
 pub const QOI_OP_RGB: u8 = 0xfe; // 11111110
 pub const QOI_OP_RGBA: u8 = 0xff; // 11111111
+pub const QOI_OP_RUN2: u8 = 0xff; // 11111111
 
 pub const QOI_MASK_2: u8 = 0xc0; // (11)000000
 
diff --git a/src/decode.rs b/src/decode.rs
index 019dac2..a4343d0 100644
--- a/src/decode.rs
+++ b/src/decode.rs
@@ -49,9 +49,21 @@ where
                 px.update_rgb(*r, *g, *b);
                 data = dtail;
             }
-            [QOI_OP_RGBA, r, g, b, a, dtail @ ..] if RGBA => {
-                px.update_rgba(*r, *g, *b, *a);
-                data = dtail;
+            [QOI_OP_RGBA, dtail @ ..] => {
+                if RGBA {
+                    if let [r, g, b, a, dtail @ ..] = dtail {
+                        px.update_rgba(*r, *g, *b, *a);
+                        data = dtail;
+                    }
+                } else if let [b1, b2, dtail @ ..] = dtail {
+                    *px_out = px.into();
+                    let run = (u16::from_be_bytes([*b1, *b2]) as usize).min(pixels.len());
+                    let (phead, ptail) = pixels.split_at_mut(run); // can't panic
+                    phead.fill(px.into());
+                    pixels = ptail;
+                    data = dtail;
+                    continue;
+                }
             }
             [b1 @ QOI_OP_RUN..=QOI_OP_RUN_END, dtail @ ..] => {
                 *px_out = px.into();
diff --git a/src/encode.rs b/src/encode.rs
index 0ed8476..ce2b36f 100644
--- a/src/encode.rs
+++ b/src/encode.rs
@@ -6,7 +6,7 @@ use std::io::Write;
 
 use bytemuck::Pod;
 
-use crate::consts::{QOI_HEADER_SIZE, QOI_OP_INDEX, QOI_OP_RUN, QOI_PADDING, QOI_PADDING_SIZE};
+use crate::consts::{QOI_HEADER_SIZE, QOI_OP_INDEX, QOI_OP_RUN, QOI_PADDING, QOI_PADDING_SIZE, QOI_OP_RUN2};
 use crate::error::{Error, Result};
 use crate::header::Header;
 use crate::pixel::{Pixel, SupportedChannels};
@@ -26,7 +26,7 @@ where
     let mut index = [Pixel::new(); 256];
     let mut px_prev = Pixel::new().with_a(0xff);
     let mut hash_prev = px_prev.hash_index();
-    let mut run = 0_u8;
+    let mut run = 0_u16;
     let mut px = Pixel::<N>::new().with_a(0xff);
     let mut index_allowed = false;
 
@@ -36,8 +36,20 @@ where
         px.read(chunk);
         if px == px_prev {
             run += 1;
+            #[cfg(not(feature = "reference"))]
+            if N == 3 {
+                if run == 65535 || unlikely(i == n_pixels - 1) {
+                    let [b1, b2] = (run-1).to_be_bytes();
+                    buf = buf.write_many(&[QOI_OP_RUN2, b1, b2])?;
+                    run = 0;
+                }
+            } else if run == 62 || unlikely(i == n_pixels - 1) {
+                buf = buf.write_one(QOI_OP_RUN | (run as u8 - 1))?;
+                run = 0;
+            }
+            #[cfg(feature = "reference")]
             if run == 62 || unlikely(i == n_pixels - 1) {
-                buf = buf.write_one(QOI_OP_RUN | (run - 1))?;
+                buf = buf.write_one(QOI_OP_RUN | (run as u8 - 1))?;
                 run = 0;
             }
         } else {
@@ -45,15 +57,18 @@ where
                 #[cfg(not(feature = "reference"))]
                 {
                     // credits for the original idea: @zakarumych (had to be fixed though)
-                    buf = buf.write_one(if run == 1 && index_allowed {
-                        QOI_OP_INDEX | hash_prev
+                    buf = if run == 1 && index_allowed {
+                        buf.write_one(QOI_OP_INDEX | hash_prev)?
+                    } else if N == 4 || run < 63 {
+                        buf.write_one(QOI_OP_RUN | (run as u8 - 1))?
                     } else {
-                        QOI_OP_RUN | (run - 1)
-                    })?;
+                        let [b1, b2] = (run-1).to_be_bytes();
+                        buf.write_many(&[QOI_OP_RUN2, b1, b2])?
+                    };
                 }
                 #[cfg(feature = "reference")]
                 {
-                    buf = buf.write_one(QOI_OP_RUN | (run - 1))?;
+                    buf = buf.write_one(QOI_OP_RUN | (run as u8 - 1))?;
                 }
                 run = 0;
             }