diff --git a/include/fmt/format.h b/include/fmt/format.h
index f47f64278dea..8525b8237989 100644
--- a/include/fmt/format.h
+++ b/include/fmt/format.h
@@ -3674,21 +3674,39 @@ struct arg_join : detail::view {
       : begin(b), end(e), sep(s) {}
 };
 
+template <typename Char>
+using iterator_arg_mapper = detail::arg_mapper<FMT_BUFFER_CONTEXT(Char)>;
+
+template <typename It, typename Char>
+using iterator_format_type =
+    decltype(std::declval<iterator_arg_mapper<Char>>().map(
+        std::declval<typename std::iterator_traits<It>::value_type>()));
+
+template <typename It, typename Char>
+using iterator_formatter = conditional_t<
+    has_formatter<iterator_format_type<It, Char>,
+                  FMT_BUFFER_CONTEXT(Char)>::value,
+    formatter<iterator_format_type<It, Char>, Char>,
+    detail::fallback_formatter<iterator_format_type<It, Char>, Char>>;
+
 template <typename It, typename Sentinel, typename Char>
 struct formatter<arg_join<It, Sentinel, Char>, Char>
-    : formatter<typename std::iterator_traits<It>::value_type, Char> {
+    : iterator_formatter<It, Char> {
   template <typename FormatContext>
   auto format(const arg_join<It, Sentinel, Char>& value, FormatContext& ctx)
       -> decltype(ctx.out()) {
-    using base = formatter<typename std::iterator_traits<It>::value_type, Char>;
+    using base = iterator_formatter<It, Char>;
+
+    iterator_arg_mapper<Char> mapper;
+
     auto it = value.begin;
     auto out = ctx.out();
     if (it != value.end) {
-      out = base::format(*it++, ctx);
+      out = base::format(mapper.map(*it++), ctx);
       while (it != value.end) {
         out = detail::copy_str<Char>(value.sep.begin(), value.sep.end(), out);
         ctx.advance_to(out);
-        out = base::format(*it++, ctx);
+        out = base::format(mapper.map(*it++), ctx);
       }
     }
     return out;
diff --git a/test/format-test.cc b/test/format-test.cc
index 49886eaf3960..997945556e98 100644
--- a/test/format-test.cc
+++ b/test/format-test.cc
@@ -11,6 +11,7 @@
 #include <cfloat>
 #include <climits>
 #include <cmath>
+#include <cstddef>
 #include <cstring>
 #include <list>
 #include <memory>
@@ -1761,6 +1762,20 @@ TEST(FormatTest, JoinArg) {
 #endif
 }
 
+TEST(FormatTest, JoinByte) {
+  using fmt::join;
+#if __cplusplus >= 201703L
+  using byte = std::byte;
+#else
+  enum class byte : unsigned char {};
+#endif
+
+  std::vector<std::byte> v = {std::byte{0x1}, std::byte{0x2}};
+
+  EXPECT_EQ("(1, 2)", format("({})", join(v, ", ")));
+  EXPECT_EQ("(0x1, 0x2)", format("({0:#x})", join(v, ", ")));
+}
+
 template <typename T> std::string str(const T& value) {
   return fmt::format("{}", value);
 }
diff --git a/test/ostream-test.cc b/test/ostream-test.cc
index 7642c94b8438..bf6a228c87b2 100644
--- a/test/ostream-test.cc
+++ b/test/ostream-test.cc
@@ -188,6 +188,11 @@ TEST(OStreamTest, Join) {
   EXPECT_EQ("1, 2, 3", fmt::format("{}", fmt::join(v, v + 3, ", ")));
 }
 
+TEST(OStreamTest, JoinCustom) {
+  auto strs = std::vector<TestString>{TestString("foo"), TestString("bar")};
+  EXPECT_EQ("foo, bar", fmt::format("{}", fmt::join(strs, ", ")));
+}
+
 #if FMT_USE_CONSTEXPR
 TEST(OStreamTest, ConstexprString) {
   EXPECT_EQ("42", format(FMT_STRING("{}"), std::string("42")));