diff --git a/source/blender/compositor/realtime_compositor/COM_context.hh b/source/blender/compositor/realtime_compositor/COM_context.hh
index 486d01f6580..945aa17f6d8 100644
--- a/source/blender/compositor/realtime_compositor/COM_context.hh
+++ b/source/blender/compositor/realtime_compositor/COM_context.hh
@@ -45,9 +45,14 @@ class Context {
   /* Get the node tree used for compositing. */
   virtual const bNodeTree &get_node_tree() const = 0;
 
-  /* True if compositor should do write file outputs, false if only running for viewing. */
+  /* True if the compositor should write file outputs, false otherwise. */
   virtual bool use_file_output() const = 0;
 
+  /* True if the compositor should write the composite output, otherwise, the compositor is assumed
+   * to not support the composite output and just displays its viewer output. In that case, the
+   * composite output will be used as a fallback viewer if no other viewer exists */
+  virtual bool use_composite_output() const = 0;
+
   /* True if color management should be used for texture evaluation. */
   virtual bool use_texture_color_management() const = 0;
 
@@ -66,10 +71,14 @@ class Context {
    * region. */
   virtual rcti get_compositing_region() const = 0;
 
-  /* Get the texture representing the output where the result of the compositor should be
-   * written. This should be called by output nodes to get their target texture. */
+  /* Get the texture where the result of the compositor should be written. This should be called by
+   * the composite output node to get its target texture. */
   virtual GPUTexture *get_output_texture() = 0;
 
+  /* Get the texture where the result of the compositor viewer should be written. This should be
+   * called by viewer output nodes to get their target texture. */
+  virtual GPUTexture *get_viewer_output_texture() = 0;
+
   /* Get the texture where the given render pass is stored. This should be called by the Render
    * Layer node to populate its outputs. */
   virtual GPUTexture *get_input_texture(int view_layer, const char *pass_name) = 0;
diff --git a/source/blender/compositor/realtime_compositor/COM_scheduler.hh b/source/blender/compositor/realtime_compositor/COM_scheduler.hh
index 7255007dc06..a0dd8ddb1ff 100644
--- a/source/blender/compositor/realtime_compositor/COM_scheduler.hh
+++ b/source/blender/compositor/realtime_compositor/COM_scheduler.hh
@@ -8,6 +8,8 @@
 
 #include "NOD_derived_node_tree.hh"
 
+#include "COM_context.hh"
+
 namespace blender::realtime_compositor {
 
 using namespace nodes::derived_node_tree_types;
@@ -18,6 +20,6 @@ using Schedule = VectorSet<DNode>;
 /* Computes the execution schedule of the node tree. This is essentially a post-order depth first
  * traversal of the node tree from the output node to the leaf input nodes, with informed order of
  * traversal of dependencies based on a heuristic estimation of the number of needed buffers. */
-Schedule compute_schedule(const DerivedNodeTree &tree);
+Schedule compute_schedule(const Context &context, const DerivedNodeTree &tree);
 
 }  // namespace blender::realtime_compositor
diff --git a/source/blender/compositor/realtime_compositor/intern/evaluator.cc b/source/blender/compositor/realtime_compositor/intern/evaluator.cc
index 47b873d207f..7ee9618dfdd 100644
--- a/source/blender/compositor/realtime_compositor/intern/evaluator.cc
+++ b/source/blender/compositor/realtime_compositor/intern/evaluator.cc
@@ -72,7 +72,7 @@ void Evaluator::compile_and_evaluate()
     return;
   }
 
-  const Schedule schedule = compute_schedule(*derived_node_tree_);
+  const Schedule schedule = compute_schedule(context_, *derived_node_tree_);
 
   CompileState compile_state(schedule);
 
diff --git a/source/blender/compositor/realtime_compositor/intern/scheduler.cc b/source/blender/compositor/realtime_compositor/intern/scheduler.cc
index 25ce9cf2cd2..e0e43d2a765 100644
--- a/source/blender/compositor/realtime_compositor/intern/scheduler.cc
+++ b/source/blender/compositor/realtime_compositor/intern/scheduler.cc
@@ -13,6 +13,7 @@
 #include "BKE_node.hh"
 #include "BKE_node_runtime.hh"
 
+#include "COM_context.hh"
 #include "COM_scheduler.hh"
 #include "COM_utilities.hh"
 
@@ -72,55 +73,88 @@ static const DTreeContext *find_active_context(const DerivedNodeTree &tree)
   return find_active_context_recursive(&tree.root_context(), NODE_INSTANCE_KEY_BASE);
 }
 
-/* Return the output node which is marked as NODE_DO_OUTPUT. If multiple types of output nodes are
- * marked, then the preference will be CMP_NODE_VIEWER > CMP_NODE_SPLITVIEWER > CMP_NODE_COMPOSITE.
- * If no output node exists, a null node will be returned. */
-static DNode find_output_in_context(const DTreeContext *context)
+/* Add the viewer node which is marked as NODE_DO_OUTPUT in the given context to the given stack.
+ * If multiple types of viewer nodes are marked, then the preference will be CMP_NODE_VIEWER >
+ * CMP_NODE_SPLITVIEWER. If no viewer nodes were found, composite nodes can be added as a fallback
+ * viewer node. */
+static bool add_viewer_nodes_in_context(const DTreeContext *context, Stack<DNode> &node_stack)
 {
-  const bNodeTree &tree = context->btree();
-
-  for (const bNode *node : tree.nodes_by_type("CompositorNodeViewer")) {
+  for (const bNode *node : context->btree().nodes_by_type("CompositorNodeViewer")) {
     if (node->flag & NODE_DO_OUTPUT) {
-      return DNode(context, node);
+      node_stack.push(DNode(context, node));
+      return true;
     }
   }
 
-  for (const bNode *node : tree.nodes_by_type("CompositorNodeSplitViewer")) {
+  for (const bNode *node : context->btree().nodes_by_type("CompositorNodeSplitViewer")) {
     if (node->flag & NODE_DO_OUTPUT) {
-      return DNode(context, node);
+      node_stack.push(DNode(context, node));
+      return true;
     }
   }
 
-  for (const bNode *node : tree.nodes_by_type("CompositorNodeComposite")) {
+  /* The active Composite node was already added, no need to add it again, see the next block. */
+  if (!node_stack.is_empty() && node_stack.peek()->type == CMP_NODE_COMPOSITE) {
+    return false;
+  }
+
+  /* No active viewers exist in this context, try to add the Composite node as a fallback viewer if
+   * it was not already added. */
+  for (const bNode *node : context->btree().nodes_by_type("CompositorNodeComposite")) {
     if (node->flag & NODE_DO_OUTPUT) {
-      return DNode(context, node);
+      node_stack.push(DNode(context, node));
+      return true;
     }
   }
 
-  return DNode();
+  return false;
 }
 
-/* Compute the output node whose result should be computed. This node is the output node that
- * satisfies the requirements in the find_output_in_context function. First, the active context is
- * searched for an output node, if non was found, the root context is search. For more information
- * on what contexts mean here, see the find_active_context function. */
-static DNode compute_output_node(const DerivedNodeTree &tree)
+/* Add the output nodes whose result should be computed to the given stack. This includes File
+ * Output, Composite, and Viewer nodes. Viewer nodes are a special case, as only the nodes that
+ * satisfies the requirements in the add_viewer_nodes_in_context function are added. First, the
+ * active context is searched for viewer nodes, if non were found, the root context is searched.
+ * For more information on what contexts mean here, see the find_active_context function. */
+static void add_output_nodes(const Context &context,
+                             const DerivedNodeTree &tree,
+                             Stack<DNode> &node_stack)
 {
+  const DTreeContext &root_context = tree.root_context();
+
+  /* Only add File Output nodes if the context supports them. */
+  if (context.use_file_output()) {
+    for (const bNode *node : root_context.btree().nodes_by_type("CompositorNodeOutputFile")) {
+      node_stack.push(DNode(&root_context, node));
+    }
+  }
+
+  /* Only add the Composite output node if the context supports composite outputs. The active
+   * Composite node may still be added as a fallback viewer output below. */
+  if (context.use_composite_output()) {
+    for (const bNode *node : root_context.btree().nodes_by_type("CompositorNodeComposite")) {
+      if (node->flag & NODE_DO_OUTPUT) {
+        node_stack.push(DNode(&root_context, node));
+        break;
+      }
+    }
+  }
+
   const DTreeContext *active_context = find_active_context(tree);
+  const bool viewer_was_added = add_viewer_nodes_in_context(active_context, node_stack);
 
-  const DNode node = find_output_in_context(active_context);
-  if (node) {
-    return node;
+  /* An active viewer was added, no need to search further. */
+  if (viewer_was_added) {
+    return;
   }
 
-  /* If the active context is the root one and no output node was found, we consider this node tree
-   * to have no output node, even if one of the non-active descendants have an output node. */
+  /* If the active context is the root one and no viewer nodes were found, we consider this node
+   * tree to have no viewer nodes, even if one of the non-active descendants have viewer nodes. */
   if (active_context->is_root()) {
-    return DNode();
+    return;
   }
 
-  /* The active context doesn't have an output node, search in the root context as a fallback. */
-  return find_output_in_context(&tree.root_context());
+  /* The active context doesn't have a viewer node, search in the root context as a fallback. */
+  add_viewer_nodes_in_context(&tree.root_context(), node_stack);
 }
 
 /* A type representing a mapping that associates each node with a heuristic estimation of the
@@ -177,12 +211,12 @@ using NeededBuffers = Map<DNode, int>;
  *   implementation because it rarely affects the output and is done by very few nodes.
  * - The compiler may decide to compiler the schedule differently depending on runtime information
  *   which we can merely speculate at scheduling-time as described above. */
-static NeededBuffers compute_number_of_needed_buffers(DNode output_node)
+static NeededBuffers compute_number_of_needed_buffers(Stack<DNode> &output_nodes)
 {
   NeededBuffers needed_buffers;
 
-  /* A stack of nodes used to traverse the node tree starting from the output node. */
-  Stack<DNode> node_stack = {output_node};
+  /* A stack of nodes used to traverse the node tree starting from the output nodes. */
+  Stack<DNode> node_stack = output_nodes;
 
   /* Traverse the node tree in a post order depth first manner and compute the number of needed
    * buffers for each node. Post order traversal guarantee that all the node dependencies of each
@@ -301,23 +335,23 @@ static NeededBuffers compute_number_of_needed_buffers(DNode output_node)
  * doesn't always guarantee an optimal evaluation order, as the optimal evaluation order is very
  * difficult to compute, however, this method works well in most cases. Moreover it assumes that
  * all buffers will have roughly the same size, which may not always be the case. */
-Schedule compute_schedule(const DerivedNodeTree &tree)
+Schedule compute_schedule(const Context &context, const DerivedNodeTree &tree)
 {
   Schedule schedule;
 
-  /* Compute the output node whose result should be computed. */
-  const DNode output_node = compute_output_node(tree);
+  /* A stack of nodes used to traverse the node tree starting from the output nodes. */
+  Stack<DNode> node_stack;
 
-  /* No output node, the node tree has no effect, return an empty schedule. */
-  if (!output_node) {
+  /* Add the output nodes whose result should be computed to the stack. */
+  add_output_nodes(context, tree, node_stack);
+
+  /* No output nodes, the node tree has no effect, return an empty schedule. */
+  if (node_stack.is_empty()) {
     return schedule;
   }
 
-  /* Compute the number of buffers needed by each node connected to the output. */
-  const NeededBuffers needed_buffers = compute_number_of_needed_buffers(output_node);
-
-  /* A stack of nodes used to traverse the node tree starting from the output node. */
-  Stack<DNode> node_stack = {output_node};
+  /* Compute the number of buffers needed by each node connected to the outputs. */
+  const NeededBuffers needed_buffers = compute_number_of_needed_buffers(node_stack);
 
   /* Traverse the node tree in a post order depth first manner, scheduling the nodes in an order
    * informed by the number of buffers needed by each node. Post order traversal guarantee that all
@@ -360,7 +394,8 @@ Schedule compute_schedule(const DerivedNodeTree &tree)
       int insertion_position = 0;
       for (int i = 0; i < sorted_dependency_nodes.size(); i++) {
         if (needed_buffers.lookup(doutput.node()) >
-            needed_buffers.lookup(sorted_dependency_nodes[i])) {
+            needed_buffers.lookup(sorted_dependency_nodes[i]))
+        {
           insertion_position++;
         }
         else {
diff --git a/source/blender/draw/engines/compositor/compositor_engine.cc b/source/blender/draw/engines/compositor/compositor_engine.cc
index 79f8b46c60f..a7fc9fa3181 100644
--- a/source/blender/draw/engines/compositor/compositor_engine.cc
+++ b/source/blender/draw/engines/compositor/compositor_engine.cc
@@ -68,6 +68,14 @@ class Context : public realtime_compositor::Context {
     return false;
   }
 
+  /* The viewport compositor doesn't really support the composite output, it only displays the
+   * viewer output in the viewport. Settings this to false will make the compositor use the
+   * composite output as fallback viewer if no other viewer exists. */
+  bool use_composite_output() const override
+  {
+    return false;
+  }
+
   bool use_texture_color_management() const override
   {
     return BKE_scene_check_color_management_enabled(DRW_context_state_get()->scene);
@@ -145,6 +153,11 @@ class Context : public realtime_compositor::Context {
     return DRW_viewport_texture_list_get()->color;
   }
 
+  GPUTexture *get_viewer_output_texture() override
+  {
+    return DRW_viewport_texture_list_get()->color;
+  }
+
   GPUTexture *get_input_texture(int view_layer, const char *pass_name) override
   {
     if (view_layer == 0 && STREQ(pass_name, RE_PASSNAME_COMBINED)) {
diff --git a/source/blender/nodes/composite/nodes/node_composite_split_viewer.cc b/source/blender/nodes/composite/nodes/node_composite_split_viewer.cc
index f6980458f9a..7b492f474fd 100644
--- a/source/blender/nodes/composite/nodes/node_composite_split_viewer.cc
+++ b/source/blender/nodes/composite/nodes/node_composite_split_viewer.cc
@@ -77,7 +77,7 @@ class ViewerOperation : public NodeOperation {
     const Result &second_image = get_input("Image_001");
     second_image.bind_as_texture(shader, "second_image_tx");
 
-    GPUTexture *output_texture = context().get_output_texture();
+    GPUTexture *output_texture = context().get_viewer_output_texture();
     const int image_unit = GPU_shader_get_sampler_binding(shader, "output_img");
     GPU_texture_image_bind(output_texture, image_unit);
 
diff --git a/source/blender/nodes/composite/nodes/node_composite_viewer.cc b/source/blender/nodes/composite/nodes/node_composite_viewer.cc
index de0372a51d9..ded8c52bca5 100644
--- a/source/blender/nodes/composite/nodes/node_composite_viewer.cc
+++ b/source/blender/nodes/composite/nodes/node_composite_viewer.cc
@@ -105,7 +105,7 @@ class ViewerOperation : public NodeOperation {
       color.w = alpha.get_float_value();
     }
 
-    GPU_texture_clear(context().get_output_texture(), GPU_DATA_FLOAT, color);
+    GPU_texture_clear(context().get_viewer_output_texture(), GPU_DATA_FLOAT, color);
   }
 
   /* Executes when the alpha channel of the image is ignored. */
@@ -123,7 +123,7 @@ class ViewerOperation : public NodeOperation {
     const Result &image = get_input("Image");
     image.bind_as_texture(shader, "input_tx");
 
-    GPUTexture *output_texture = context().get_output_texture();
+    GPUTexture *output_texture = context().get_viewer_output_texture();
     const int image_unit = GPU_shader_get_sampler_binding(shader, "output_img");
     GPU_texture_image_bind(output_texture, image_unit);
 
@@ -151,7 +151,7 @@ class ViewerOperation : public NodeOperation {
     const Result &image = get_input("Image");
     image.bind_as_texture(shader, "input_tx");
 
-    GPUTexture *output_texture = context().get_output_texture();
+    GPUTexture *output_texture = context().get_viewer_output_texture();
     const int image_unit = GPU_shader_get_sampler_binding(shader, "output_img");
     GPU_texture_image_bind(output_texture, image_unit);
 
@@ -181,7 +181,7 @@ class ViewerOperation : public NodeOperation {
     const Result &alpha = get_input("Alpha");
     alpha.bind_as_texture(shader, "alpha_tx");
 
-    GPUTexture *output_texture = context().get_output_texture();
+    GPUTexture *output_texture = context().get_viewer_output_texture();
     const int image_unit = GPU_shader_get_sampler_binding(shader, "output_img");
     GPU_texture_image_bind(output_texture, image_unit);
 
diff --git a/source/blender/render/intern/compositor.cc b/source/blender/render/intern/compositor.cc
index 271ee17903b..5e121831e7c 100644
--- a/source/blender/render/intern/compositor.cc
+++ b/source/blender/render/intern/compositor.cc
@@ -2,9 +2,13 @@
  *
  * SPDX-License-Identifier: GPL-2.0-or-later */
 
+#include <cstring>
+
 #include "BLI_threads.h"
 #include "BLI_vector.hh"
 
+#include "MEM_guardedalloc.h"
+
 #include "BKE_global.h"
 #include "BKE_image.h"
 #include "BKE_node.hh"
@@ -12,6 +16,9 @@
 
 #include "DRW_engine.h"
 
+#include "IMB_colormanagement.h"
+#include "IMB_imbuf.h"
+
 #include "COM_context.hh"
 #include "COM_evaluator.hh"
 
@@ -63,6 +70,9 @@ class Context : public realtime_compositor::Context {
   /* Output combined texture. */
   GPUTexture *output_texture_ = nullptr;
 
+  /* Viewer output texture. */
+  GPUTexture *viewer_output_texture_ = nullptr;
+
  public:
   Context(const Scene &scene,
           const RenderData &render_data,
@@ -81,7 +91,8 @@ class Context : public realtime_compositor::Context {
 
   virtual ~Context()
   {
-    GPU_texture_free(output_texture_);
+    GPU_TEXTURE_FREE_SAFE(output_texture_);
+    GPU_TEXTURE_FREE_SAFE(viewer_output_texture_);
   }
 
   const bNodeTree &get_node_tree() const override
@@ -94,6 +105,11 @@ class Context : public realtime_compositor::Context {
     return use_file_output_;
   }
 
+  bool use_composite_output() const override
+  {
+    return true;
+  }
+
   bool use_texture_color_management() const override
   {
     return BKE_scene_check_color_management_enabled(&scene_);
@@ -121,7 +137,7 @@ class Context : public realtime_compositor::Context {
 
   GPUTexture *get_output_texture() override
   {
-    /* TODO: support outputting for viewers and previews.
+    /* TODO: support outputting for previews.
      * TODO: just a temporary hack, needs to get stored in RenderResult,
      * once that supports GPU buffers. */
     if (output_texture_ == nullptr) {
@@ -138,6 +154,25 @@ class Context : public realtime_compositor::Context {
     return output_texture_;
   }
 
+  GPUTexture *get_viewer_output_texture() override
+  {
+    /* TODO: support outputting previews.
+     * TODO: just a temporary hack, needs to get stored in RenderResult,
+     * once that supports GPU buffers. */
+    if (viewer_output_texture_ == nullptr) {
+      const int2 size = get_render_size();
+      viewer_output_texture_ = GPU_texture_create_2d("compositor_viewer_output_texture",
+                                                     size.x,
+                                                     size.y,
+                                                     1,
+                                                     GPU_RGBA16F,
+                                                     GPU_TEXTURE_USAGE_GENERAL,
+                                                     NULL);
+    }
+
+    return viewer_output_texture_;
+  }
+
   GPUTexture *get_input_texture(int view_layer_id, const char *pass_name) override
   {
     /* TODO: eventually this should get cached on the RenderResult itself when
@@ -224,6 +259,10 @@ class Context : public realtime_compositor::Context {
 
   void output_to_render_result()
   {
+    if (!output_texture_) {
+      return;
+    }
+
     Render *re = RE_GetSceneRender(&scene_);
     RenderResult *rr = RE_AcquireResultWrite(re);
 
@@ -253,6 +292,55 @@ class Context : public realtime_compositor::Context {
     BKE_image_signal(G.main, image, nullptr, IMA_SIGNAL_FREE);
     BLI_thread_unlock(LOCK_DRAW_IMAGE);
   }
+
+  void viewer_output_to_viewer_image()
+  {
+    if (!viewer_output_texture_) {
+      return;
+    }
+
+    Image *image = BKE_image_ensure_viewer(G.main, IMA_TYPE_COMPOSITE, "Viewer Node");
+
+    ImageUser image_user = {0};
+    image_user.multi_index = BKE_scene_multiview_view_id_get(&render_data_, view_name_);
+
+    if (BKE_scene_multiview_is_render_view_first(&render_data_, view_name_)) {
+      BKE_image_ensure_viewer_views(&render_data_, image, &image_user);
+    }
+
+    BLI_thread_lock(LOCK_DRAW_IMAGE);
+
+    void *lock;
+    ImBuf *image_buffer = BKE_image_acquire_ibuf(image, &image_user, &lock);
+
+    const int2 render_size = get_render_size();
+    if (image_buffer->x != render_size.x || image_buffer->y != render_size.y) {
+      imb_freerectImBuf(image_buffer);
+      imb_freerectfloatImBuf(image_buffer);
+      IMB_freezbuffloatImBuf(image_buffer);
+      image_buffer->x = render_size.x;
+      image_buffer->y = render_size.y;
+      imb_addrectfloatImBuf(image_buffer, 4);
+      image_buffer->userflags |= IB_DISPLAY_BUFFER_INVALID;
+    }
+
+    BKE_image_release_ibuf(image, image_buffer, lock);
+    BLI_thread_unlock(LOCK_DRAW_IMAGE);
+
+    GPU_memory_barrier(GPU_BARRIER_TEXTURE_UPDATE);
+    float *output_buffer = (float *)GPU_texture_read(viewer_output_texture_, GPU_DATA_FLOAT, 0);
+
+    std::memcpy(image_buffer->float_buffer.data,
+                output_buffer,
+                render_size.x * render_size.y * 4 * sizeof(float));
+
+    MEM_freeN(output_buffer);
+
+    BKE_image_partial_update_mark_full_update(image);
+    if (node_tree_.runtime->update_draw) {
+      node_tree_.runtime->update_draw(node_tree_.runtime->udh);
+    }
+  }
 };
 
 /* Render Realtime Compositor */
@@ -289,6 +377,7 @@ void RealtimeCompositor::execute()
   DRW_render_context_enable(&render_);
   evaluator_->evaluate();
   context_->output_to_render_result();
+  context_->viewer_output_to_viewer_image();
   DRW_render_context_disable(&render_);
 }