diff --git a/rtengine/color.cc b/rtengine/color.cc
index 3be412b6d..71fb097f6 100644
--- a/rtengine/color.cc
+++ b/rtengine/color.cc
@@ -1747,15 +1747,41 @@ void Color::Lab2XYZ(vfloat L, vfloat a, vfloat b, vfloat &x, vfloat &y, vfloat &
 }
 #endif // __SSE2__
 
+inline float Color::computeXYZ2Lab(float f)
+{
+    if (f < 0.f) {
+        return 327.68 * ((kappa * f / MAXVALF + 16.0) / 116.0);
+    } else if (f > 65535.f) {
+        return (327.68f * xcbrtf(f / MAXVALF));
+    } else {
+        return cachef[f];
+    }
+}
+
+
+inline float Color::computeXYZ2LabY(float f)
+{
+    if (f < 0.f) {
+        return 327.68 * (kappa * f / MAXVALF);
+    } else if (f > 65535.f) {
+        return 327.68f * (116.f * xcbrtf(f / MAXVALF) - 16.f);
+    } else {
+        return cachefy[f];
+    }
+}
+
+
 void Color::RGB2Lab(float *R, float *G, float *B, float *L, float *a, float *b, const float wp[3][3], int width)
 {
 
 #ifdef __SSE2__
+    vfloat minvalfv = F2V(0.f);
     vfloat maxvalfv = F2V(MAXVALF);
     vfloat c500v = F2V(500.f);
     vfloat c200v = F2V(200.f);
 #endif
     int i = 0;
+    
 #ifdef __SSE2__
     for(;i < width - 3; i+=4) {
         const vfloat rv = LVFU(R[i]);
@@ -1766,17 +1792,18 @@ void Color::RGB2Lab(float *R, float *G, float *B, float *L, float *a, float *b,
         const vfloat zv = F2V(wp[2][0]) * rv + F2V(wp[2][1]) * gv + F2V(wp[2][2]) * bv;
 
         vmask maxMask = vmaskf_gt(vmaxf(xv, vmaxf(yv, zv)), maxvalfv);
-        if (_mm_movemask_ps((vfloat)maxMask)) {
+        vmask minMask = vmaskf_lt(vminf(xv, vminf(yv, zv)), minvalfv);
+        if (_mm_movemask_ps((vfloat)maxMask) || _mm_movemask_ps((vfloat)minMask)) {
             // take slower code path for all 4 pixels if one of the values is > MAXVALF. Still faster than non SSE2 version
             for(int k = 0; k < 4; ++k) {
                 float x = xv[k];
                 float y = yv[k];
                 float z = zv[k];
-                float fx = (x <= 65535.f ? cachef[x] : (327.68f * xcbrtf(x / MAXVALF)));
-                float fy = (y <= 65535.f ? cachef[y] : (327.68f * xcbrtf(y / MAXVALF)));
-                float fz = (z <= 65535.f ? cachef[z] : (327.68f * xcbrtf(z / MAXVALF)));
+                float fx = computeXYZ2Lab(x);
+                float fy = computeXYZ2Lab(y);
+                float fz = computeXYZ2Lab(z);
 
-                L[i + k] = (y <= 65535.0f ? cachefy[y] : 327.68f * (116.f * xcbrtf(y / MAXVALF) - 16.f));
+                L[i + k] = computeXYZ2LabY(y);
                 a[i + k] = (500.f * (fx - fy) );
                 b[i + k] = (200.f * (fy - fz) );
             }
@@ -1800,11 +1827,11 @@ void Color::RGB2Lab(float *R, float *G, float *B, float *L, float *a, float *b,
         float z = wp[2][0] * rv + wp[2][1] * gv + wp[2][2] * bv;
         float fx, fy, fz;
 
-        fx = (x <= 65535.0f ? cachef[x] : (327.68f * xcbrtf(x / MAXVALF)));
-        fy = (y <= 65535.0f ? cachef[y] : (327.68f * xcbrtf(y / MAXVALF)));
-        fz = (z <= 65535.0f ? cachef[z] : (327.68f * xcbrtf(z / MAXVALF)));
+        fx = computeXYZ2Lab(x);
+        fy = computeXYZ2Lab(y);
+        fz = computeXYZ2Lab(z);
 
-        L[i] = (y <= 65535.0f ? cachefy[y] : 327.68f * (116.f * xcbrtf(y / MAXVALF) - 16.f));
+        L[i] = computeXYZ2LabY(y);
         a[i] = 500.0f * (fx - fy);
         b[i] = 200.0f * (fy - fz);
     }
@@ -1818,11 +1845,11 @@ void Color::XYZ2Lab(float X, float Y, float Z, float &L, float &a, float &b)
     float y = Y;
     float fx, fy, fz;
 
-    fx = (x <= 65535.0f ? cachef[x] : (327.68f * xcbrtf(x / MAXVALF)));
-    fy = (y <= 65535.0f ? cachef[y] : (327.68f * xcbrtf(y / MAXVALF)));
-    fz = (z <= 65535.0f ? cachef[z] : (327.68f * xcbrtf(z / MAXVALF)));
+    fx = computeXYZ2Lab(x);
+    fy = computeXYZ2Lab(y);
+    fz = computeXYZ2Lab(z);
 
-    L = (y <= 65535.0f ? cachefy[y] : 327.68f * (116.f * xcbrtf(y / MAXVALF) - 16.f));
+    L = computeXYZ2LabY(y);
     a = (500.0f * (fx - fy) );
     b = (200.0f * (fy - fz) );
 }
@@ -1854,11 +1881,11 @@ void Color::Yuv2Lab(float Yin, float u, float v, float &L, float &a, float &b, c
 
     gamutmap(X, Y, Z, wp);
 
-    float fx = (X <= 65535.0 ? cachef[X] : (327.68 * std::cbrt(X / MAXVALF)));
-    float fy = (Y <= 65535.0 ? cachef[Y] : (327.68 * std::cbrt(Y / MAXVALF)));
-    float fz = (Z <= 65535.0 ? cachef[Z] : (327.68 * std::cbrt(Z / MAXVALF)));
+    float fx = computeXYZ2Lab(X);
+    float fy = computeXYZ2Lab(Y);
+    float fz = computeXYZ2Lab(Z);
 
-    L = (Y <= 65535.0f ? cachefy[Y] : 327.68f * (116.f * xcbrtf(Y / MAXVALF) - 16.f));
+    L = computeXYZ2LabY(Y);
     a = (500.0 * (fx - fy) );
     b = (200.0 * (fy - fz) );
 }
diff --git a/rtengine/color.h b/rtengine/color.h
index 4b865f1d9..a0d8d93ee 100644
--- a/rtengine/color.h
+++ b/rtengine/color.h
@@ -97,6 +97,10 @@ private:
 #ifdef __SSE2__
     static vfloat hue2rgb(vfloat p, vfloat q, vfloat t);
 #endif
+
+    static float computeXYZ2Lab(float f);
+    static float computeXYZ2LabY(float f);
+    
 public:
 
     typedef enum Channel {
diff --git a/rtengine/curves.h b/rtengine/curves.h
index 01b17a02c..3762b3f1a 100644
--- a/rtengine/curves.h
+++ b/rtengine/curves.h
@@ -51,6 +51,9 @@ inline void setLutVal(const LUTf &lut, float &val)
 {
     if (!OOG(val)) {
         val = lut[std::max(val, 0.f)];
+    } else if (val < 0.f) {
+        float m = lut[0.f];
+        val += m;
     } else {
         float m = lut[MAXVALF];
         val += (m - MAXVALF);
@@ -61,8 +64,8 @@ inline void setLutVal(float &val, float lutval, float maxval)
 {
     if (!OOG(val)) {
         val = lutval;
-    } else {
-        val += (maxval - MAXVALF);
+    } else if (val > 0.f) {
+        val += maxval - MAXVALF;
     }
 }
 
diff --git a/rtengine/iplab2rgb.cc b/rtengine/iplab2rgb.cc
index e56e63dea..676213ef8 100644
--- a/rtengine/iplab2rgb.cc
+++ b/rtengine/iplab2rgb.cc
@@ -394,9 +394,9 @@ Imagefloat* ImProcFunctions::lab2rgbOut (LabImage* lab, int cx, int cy, int cw,
 
                 Color::xyz2srgb(x_, y_, z_, R, G, B);
 
-                setUnlessOOG(image->r(i - cy, j - cx), Color::gamma2curve[CLIP(R)]);
-                setUnlessOOG(image->g(i - cy, j - cx), Color::gamma2curve[CLIP(G)]);
-                setUnlessOOG(image->b(i - cy, j - cx), Color::gamma2curve[CLIP(B)]);
+                image->r(i - cy, j - cx) = Color::gamma2curve[CLIP(R)];
+                image->g(i - cy, j - cx) = Color::gamma2curve[CLIP(G)];
+                image->b(i - cy, j - cx) = Color::gamma2curve[CLIP(B)];
             }
         }
     }
diff --git a/rtengine/rt_math.h b/rtengine/rt_math.h
index 1918439ef..a8c835a98 100644
--- a/rtengine/rt_math.h
+++ b/rtengine/rt_math.h
@@ -141,7 +141,7 @@ constexpr std::uint8_t uint16ToUint8Rounded(std::uint16_t i)
 template <typename T>
 constexpr bool OOG(const T &val, const T &high=T(MAXVAL))
 {
-    return (val > high);
+    return (val < T(0)) || (val > high);
 }
 
 template <typename T>
diff --git a/rtgui/editorpanel.cc b/rtgui/editorpanel.cc
index 2f88e0c11..60dd25e7e 100644
--- a/rtgui/editorpanel.cc
+++ b/rtgui/editorpanel.cc
@@ -1979,10 +1979,6 @@ bool EditorPanel::saveImmediately (const Glib::ustring &filename, const SaveForm
     rtengine::procparams::ProcParams pparams;
     ipc->getParams (&pparams);
 
-    if (gimpPlugin) {
-        pparams.icm.gamma = "linear_g1.0";
-    }
-
     rtengine::ProcessingJob *job = rtengine::ProcessingJob::create (ipc->getInitialImage(), pparams);
 
     // save immediately