diff --git a/libvo/gl_osd.c b/libvo/gl_osd.c
index 74053afd4a..5266dca4ac 100644
--- a/libvo/gl_osd.c
+++ b/libvo/gl_osd.c
@@ -32,7 +32,7 @@ struct osd_fmt_entry {
 // glBlendFunc() arguments
 static const int blend_factors[SUBBITMAP_COUNT][2] = {
     [SUBBITMAP_LIBASS] = {GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA},
-    [SUBBITMAP_RGBA] =   {GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA},
+    [SUBBITMAP_RGBA] =   {GL_ONE,       GL_ONE_MINUS_SRC_ALPHA},
 };
 
 static const struct osd_fmt_entry osd_to_gl3_formats[SUBBITMAP_COUNT] = {
diff --git a/libvo/vo_direct3d.c b/libvo/vo_direct3d.c
index 4f16c6a6b1..2b8f3429e4 100644
--- a/libvo/vo_direct3d.c
+++ b/libvo/vo_direct3d.c
@@ -2049,6 +2049,9 @@ static void draw_osd(d3d_priv *priv, struct sub_bitmaps *imgs)
 
         IDirect3DDevice9_SetTextureStageState(priv->d3d_device, 0,
                                             D3DTSS_ALPHAOP, D3DTOP_MODULATE);
+    } else {
+        IDirect3DDevice9_SetRenderState(priv->d3d_device, D3DRS_SRCBLEND,
+                                        D3DBLEND_ONE);
     }
 
     IDirect3DDevice9_SetFVF(priv->d3d_device, D3DFVF_EOSD_VERTEX);
@@ -2060,6 +2063,8 @@ static void draw_osd(d3d_priv *priv, struct sub_bitmaps *imgs)
                                           D3DTSS_COLORARG1, D3DTA_TEXTURE);
     IDirect3DDevice9_SetTextureStageState(priv->d3d_device, 0,
                                           D3DTSS_ALPHAOP, D3DTOP_SELECTARG1);
+    IDirect3DDevice9_SetRenderState(priv->d3d_device,
+                                    D3DRS_SRCBLEND, D3DBLEND_SRCALPHA);
 
     IDirect3DDevice9_SetTexture(priv->d3d_device, 0, NULL);
 
diff --git a/libvo/vo_vdpau.c b/libvo/vo_vdpau.c
index 34aa9824dc..94d37e3018 100644
--- a/libvo/vo_vdpau.c
+++ b/libvo/vo_vdpau.c
@@ -970,14 +970,21 @@ static void draw_eosd(struct vo *vo, int index)
         .blend_equation_alpha = VDP_OUTPUT_SURFACE_RENDER_BLEND_EQUATION_ADD,
     };
 
+    VdpOutputSurfaceRenderBlendState blend_state_premultiplied = blend_state;
+    blend_state_premultiplied.blend_factor_source_color =
+            VDP_OUTPUT_SURFACE_RENDER_BLEND_FACTOR_ONE;
+
     for (i = 0; i < sfc->render_count; i++) {
+        VdpOutputSurfaceRenderBlendState *blend = &blend_state;
+        if (sfc->format == VDP_RGBA_FORMAT_B8G8R8A8)
+            blend = &blend_state_premultiplied;
         vdp_st = vdp->
             output_surface_render_bitmap_surface(output_surface,
                                                  &sfc->targets[i].dest,
                                                  sfc->surface,
                                                  &sfc->targets[i].source,
                                                  &sfc->targets[i].color,
-                                                 &blend_state,
+                                                 blend,
                                                  VDP_OUTPUT_SURFACE_RENDER_ROTATE_0);
         CHECK_ST_WARNING("EOSD: Error when rendering");
     }
diff --git a/sub/img_convert.c b/sub/img_convert.c
index b8a769f052..888380cf70 100644
--- a/sub/img_convert.c
+++ b/sub/img_convert.c
@@ -166,6 +166,21 @@ bool osd_conv_ass_to_old_p(struct osd_conv_cache *c, struct sub_bitmaps *imgs)
     return true;
 }
 
+static void rgba_to_premultiplied_rgba(uint32_t *colors, size_t count)
+{
+    for (int n = 0; n < count; n++) {
+        uint32_t c = colors[n];
+        int b = c & 0xFF;
+        int g = (c >> 8) & 0xFF;
+        int r = (c >> 16) & 0xFF;
+        int a = (c >> 24) & 0xFF;
+        b = b * a / 255;
+        g = g * a / 255;
+        r = r * a / 255;
+        colors[n] = b | (g << 8) | (r << 16) | (a << 24);
+    }
+}
+
 bool osd_conv_idx_to_rgba(struct osd_conv_cache *c, struct sub_bitmaps *imgs)
 {
     struct sub_bitmaps src = *imgs;
@@ -179,18 +194,19 @@ bool osd_conv_idx_to_rgba(struct osd_conv_cache *c, struct sub_bitmaps *imgs)
     for (int n = 0; n < src.num_parts; n++) {
         struct sub_bitmap *d = &imgs->parts[n];
         struct sub_bitmap *s = &src.parts[n];
-        struct osd_bmp_indexed *sb = s->bitmap;
+        struct osd_bmp_indexed sb = *(struct osd_bmp_indexed *)s->bitmap;
+
+        rgba_to_premultiplied_rgba(sb.palette, 256);
 
         *d = *s;
         d->stride = s->w * 4;
         d->bitmap = talloc_size(c->parts, s->h * d->stride);
 
-        uint32_t *palette = sb->palette;
         uint32_t *outbmp = d->bitmap;
         for (int y = 0; y < s->h; y++) {
-            uint8_t *inbmp = sb->bitmap + y * s->stride;
+            uint8_t *inbmp = sb.bitmap + y * s->stride;
             for (int x = 0; x < s->w; x++)
-                *outbmp++ = palette[*inbmp++];
+                *outbmp++ = sb.palette[*inbmp++];
         }
     }
     return true;
diff --git a/sub/sub.h b/sub/sub.h
index 8649018029..649f196414 100644
--- a/sub/sub.h
+++ b/sub/sub.h
@@ -30,7 +30,7 @@ struct sub_render_params;
 enum sub_bitmap_format {
     SUBBITMAP_EMPTY = 0,// no bitmaps; always has num_parts==0
     SUBBITMAP_LIBASS,   // A8, with a per-surface blend color (libass.color)
-    SUBBITMAP_RGBA,     // B8G8R8A8 (MSB=A, LSB=B), can be scaled
+    SUBBITMAP_RGBA,     // B8G8R8A8 (MSB=A, LSB=B), scaled, premultiplied alpha
     SUBBITMAP_INDEXED,  // scaled, bitmap points to osd_bmp_indexed
     SUBBITMAP_OLD_PLANAR, // like previous, but bitmap points to old_osd_planar