diff --git a/src/freedreno/ir3/ir3_nir.c b/src/freedreno/ir3/ir3_nir.c
index 7148ed0a221..200e53e6997 100644
--- a/src/freedreno/ir3/ir3_nir.c
+++ b/src/freedreno/ir3/ir3_nir.c
@@ -1537,6 +1537,7 @@ ir3_nir_lower_variant(struct ir3_shader_variant *so,
    progress |= OPT(s, nir_lower_mem_access_bit_sizes, &mem_bit_size_options);
    progress |= OPT(s, ir3_nir_lower_64b_global);
    progress |= OPT(s, ir3_nir_lower_64b_undef);
+   progress |= OPT(s, ir3_nir_lower_64b_image);
    progress |= OPT(s, nir_lower_int64);
    progress |= OPT(s, nir_lower_64bit_phis);
 
diff --git a/src/freedreno/ir3/ir3_nir.h b/src/freedreno/ir3/ir3_nir.h
index 46006e9848b..887170d68cf 100644
--- a/src/freedreno/ir3/ir3_nir.h
+++ b/src/freedreno/ir3/ir3_nir.h
@@ -54,6 +54,7 @@ uint8_t ir3_nir_vectorize_filter(const nir_instr *instr, const void *data);
 bool ir3_nir_lower_64b_undef(nir_shader *shader);
 bool ir3_nir_lower_64b_global(nir_shader *shader);
 bool ir3_nir_lower_64b_regs(nir_shader *shader);
+bool ir3_nir_lower_64b_image(nir_shader *shader);
 
 nir_mem_access_size_align ir3_mem_access_size_align(
    nir_intrinsic_op intrin, uint8_t bytes, uint8_t bit_size, uint32_t align,
diff --git a/src/freedreno/ir3/ir3_nir_lower_64b.c b/src/freedreno/ir3/ir3_nir_lower_64b.c
index 1688c7346a1..4cab8d04652 100644
--- a/src/freedreno/ir3/ir3_nir_lower_64b.c
+++ b/src/freedreno/ir3/ir3_nir_lower_64b.c
@@ -205,3 +205,71 @@ ir3_nir_lower_64b_regs(nir_shader *shader)
 
    return progress;
 }
+
+static void
+lower_64b_image_load(nir_builder *b, nir_intrinsic_instr *intr)
+{
+   nir_intrinsic_set_format(intr, PIPE_FORMAT_R32G32_UINT);
+   nir_intrinsic_set_dest_type(intr, nir_type_uint32);
+   unsigned ncomp = intr->num_components;
+   intr->num_components = 2;
+   intr->def.num_components = 2;
+   intr->def.bit_size = 32;
+
+   /* We only load the R component but the original load might need GBA as well,
+    * return the default (0, 0, 1) values.
+    */
+   b->cursor = nir_after_instr(&intr->instr);
+   nir_def *val_r64 = nir_pack_64_2x32(b, &intr->def);
+   nir_def *vec[4] = {val_r64, nir_imm_int64(b, 0), nir_imm_int64(b, 0),
+                      nir_imm_int64(b, 1)};
+   assert(ncomp <= ARRAY_SIZE(vec));
+   nir_def_rewrite_uses_after(&intr->def, nir_vec(b, vec, ncomp));
+}
+
+static void
+lower_64b_image_store(nir_builder *b, nir_intrinsic_instr *intr)
+{
+   nir_intrinsic_set_format(intr, PIPE_FORMAT_R32G32_UINT);
+   nir_intrinsic_set_src_type(intr, nir_type_uint32);
+   intr->num_components = 2;
+
+   b->cursor = nir_before_instr(&intr->instr);
+   nir_def *val_r64 = nir_channel(b, intr->src[3].ssa, 0);
+   nir_def *val_r32g32 = nir_unpack_64_2x32(b, val_r64);
+   nir_src_rewrite(&intr->src[3], val_r32g32);
+}
+
+static bool
+lower_64b_image(nir_builder *b, nir_intrinsic_instr *intr, void *data)
+{
+   switch (intr->intrinsic) {
+   case nir_intrinsic_bindless_image_load:
+   case nir_intrinsic_bindless_image_store: {
+      enum pipe_format format = nir_intrinsic_format(intr);
+      if (format != PIPE_FORMAT_R64_UINT && format != PIPE_FORMAT_R64_SINT) {
+         return false;
+      }
+      break;
+   }
+   default:
+      return false;
+   }
+
+   if (intr->intrinsic == nir_intrinsic_bindless_image_load) {
+      lower_64b_image_load(b, intr);
+   } else {
+      lower_64b_image_store(b, intr);
+   }
+
+   return true;
+}
+
+/* Lower r64u?i image loads/stores to r32g32u, which is our descriptor format.
+ */
+bool
+ir3_nir_lower_64b_image(nir_shader *shader)
+{
+   return nir_shader_intrinsics_pass(shader, lower_64b_image,
+                                     nir_metadata_control_flow, NULL);
+}