diff --git a/source/blender/draw/CMakeLists.txt b/source/blender/draw/CMakeLists.txt
index e1aad7d4280..dc16818d22b 100644
--- a/source/blender/draw/CMakeLists.txt
+++ b/source/blender/draw/CMakeLists.txt
@@ -429,6 +429,7 @@ set(GLSL_SRC
   engines/eevee_next/shaders/eevee_transparency_lib.glsl
   engines/eevee_next/shaders/eevee_debug_surfels_vert.glsl
   engines/eevee_next/shaders/eevee_debug_surfels_frag.glsl
+  engines/eevee_next/shaders/eevee_deferred_light_frag.glsl
   engines/eevee_next/shaders/eevee_depth_of_field_accumulator_lib.glsl
   engines/eevee_next/shaders/eevee_depth_of_field_bokeh_lut_comp.glsl
   engines/eevee_next/shaders/eevee_depth_of_field_downsample_comp.glsl
@@ -448,6 +449,7 @@ set(GLSL_SRC
   engines/eevee_next/shaders/eevee_film_cryptomatte_post_comp.glsl
   engines/eevee_next/shaders/eevee_film_frag.glsl
   engines/eevee_next/shaders/eevee_film_lib.glsl
+  engines/eevee_next/shaders/eevee_gbuffer_lib.glsl
   engines/eevee_next/shaders/eevee_geom_curves_vert.glsl
   engines/eevee_next/shaders/eevee_geom_gpencil_vert.glsl
   engines/eevee_next/shaders/eevee_geom_mesh_vert.glsl
diff --git a/source/blender/draw/engines/eevee_next/eevee_defines.hh b/source/blender/draw/engines/eevee_next/eevee_defines.hh
index bc778dc874d..c54d537db61 100644
--- a/source/blender/draw/engines/eevee_next/eevee_defines.hh
+++ b/source/blender/draw/engines/eevee_next/eevee_defines.hh
@@ -105,6 +105,9 @@
 #define RBUFS_AOV_COLOR_SLOT 5
 #define RBUFS_AOV_VALUE_SLOT 6
 #define RBUFS_CRYPTOMATTE_SLOT 7
+/* G-buffer reuses render passes slots. */
+#define GBUF_CLOSURE_SLOT RBUFS_LIGHT_SLOT
+#define GBUF_COLOR_SLOT RBUFS_DIFF_COLOR_SLOT
 
 /* Uniform Buffers. */
 /* Only during prepass. */
diff --git a/source/blender/draw/engines/eevee_next/eevee_film.cc b/source/blender/draw/engines/eevee_next/eevee_film.cc
index beab3bdb196..d49ddd7866e 100644
--- a/source/blender/draw/engines/eevee_next/eevee_film.cc
+++ b/source/blender/draw/engines/eevee_next/eevee_film.cc
@@ -649,7 +649,7 @@ void Film::accumulate(const DRWView *view, GPUTexture *combined_final_tx)
 
   draw::View drw_view("MainView", view);
 
-  DRW_manager_get()->submit(accumulate_ps_, drw_view);
+  inst_.manager->submit(accumulate_ps_, drw_view);
 
   combined_tx_.swap();
   weight_tx_.swap();
diff --git a/source/blender/draw/engines/eevee_next/eevee_gbuffer.hh b/source/blender/draw/engines/eevee_next/eevee_gbuffer.hh
new file mode 100644
index 00000000000..74a17e317f2
--- /dev/null
+++ b/source/blender/draw/engines/eevee_next/eevee_gbuffer.hh
@@ -0,0 +1,89 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+
+/** \file
+ * \ingroup eevee
+ *
+ * Gbuffer layout used for deferred shading pipeline.
+ */
+
+#pragma once
+
+#include "DRW_render.h"
+
+#include "eevee_material.hh"
+#include "eevee_shader_shared.hh"
+
+namespace blender::eevee {
+
+class Instance;
+
+/**
+ * Fullscreen textures containing geometric and surface data.
+ * Used by deferred shading passes. Only one gbuffer is allocated per view
+ * and is reused for each deferred layer. This is why there can only be temporary
+ * texture inside it.
+ *
+ * Everything is stored inside two array texture, one for each format. This is to fit the
+ * limitation of the number of images we can bind on a single shader.
+ *
+ * First layer is always for reflection. All parameters to shoot a reflection ray are inside
+ * this layer.
+ *
+ * - Layer 1 : Reflection
+ *   - R : Normal packed X
+ *   - G : Normal packed Y
+ *   - B : Roughness
+ *   - A : Unused (Could be used for anisotropic roughness)
+ *
+ * Second layer is either for diffuse or transmission. Material mixing both are not
+ * physically based and are uncommon. So in order to save bandwidth and texture memory, we only
+ * store one. We use random sampling to mix between both. All parameters to shoot a refraction
+ * ray are inside this layer.
+ *
+ * - Layer 2 : Refraction
+ *   - R : Normal packed X
+ *   - G : Normal packed Y
+ *   - B : Roughness (isotropic)
+ *   - A : IOR
+ *
+ * - Layer 2 : Diffuse / Sub-Surface Scattering
+ *   - R : Normal packed X
+ *   - G : Normal packed Y
+ *   - B : Thickness
+ *   - A : Unused (Could be used for diffuse roughness)
+ *
+ * Layer 3 is only allocated if Sub-Surface Scattering is needed. All parameters for
+ * screen-space scattering are inside this layer.
+ *
+ * - Layer 3 : Sub-Surface Scattering
+ *   - R : Scattering radius R
+ *   - G : Scattering radius G
+ *   - B : Scattering radius B
+ *   - A : Object ID
+ *
+ * For each output closure, we also output the color to apply after the lighting computation.
+ * The color is stored with a 2 exponent that allows input color with component higher than 1.
+ * Color degradation is expected to happen in this case.
+ */
+struct GBuffer {
+  /* TODO(fclem): Use texture from pool once they support texture array. */
+  Texture closure_tx = {"GbufferClosure"};
+  Texture color_tx = {"GbufferColor"};
+
+  void acquire(int2 extent, eClosureBits closure_bits_)
+  {
+    const bool use_sss = (closure_bits_ & CLOSURE_SSS) != 0;
+    eGPUTextureUsage usage = GPU_TEXTURE_USAGE_SHADER_READ | GPU_TEXTURE_USAGE_SHADER_WRITE;
+    closure_tx.ensure_2d_array(GPU_RGBA16, extent, use_sss ? 3 : 2, usage);
+    color_tx.ensure_2d_array(GPU_RGB10_A2, extent, 2, usage);
+  }
+
+  void release()
+  {
+    /* TODO(fclem): Use texture from pool once they support texture array. */
+    // closure_tx.release();
+    // color_tx.release();
+  }
+};
+
+}  // namespace blender::eevee
\ No newline at end of file
diff --git a/source/blender/draw/engines/eevee_next/eevee_instance.cc b/source/blender/draw/engines/eevee_next/eevee_instance.cc
index 491190334e1..b0a40d27982 100644
--- a/source/blender/draw/engines/eevee_next/eevee_instance.cc
+++ b/source/blender/draw/engines/eevee_next/eevee_instance.cc
@@ -105,6 +105,7 @@ void Instance::begin_sync()
   velocity.begin_sync(); /* NOTE: Also syncs camera. */
   lights.begin_sync();
   shadows.begin_sync();
+  pipelines.begin_sync();
   cryptomatte.begin_sync();
 
   gpencil_engine_enabled = false;
@@ -114,7 +115,6 @@ void Instance::begin_sync()
   depth_of_field.sync();
   motion_blur.sync();
   hiz_buffer.sync();
-  pipelines.sync();
   main_view.sync();
   world.sync();
   film.sync();
@@ -206,6 +206,7 @@ void Instance::end_sync()
   sampling.end_sync();
   film.end_sync();
   cryptomatte.end_sync();
+  pipelines.end_sync();
 }
 
 void Instance::render_sync()
diff --git a/source/blender/draw/engines/eevee_next/eevee_instance.hh b/source/blender/draw/engines/eevee_next/eevee_instance.hh
index 6bb54a37a2c..8e7855961e4 100644
--- a/source/blender/draw/engines/eevee_next/eevee_instance.hh
+++ b/source/blender/draw/engines/eevee_next/eevee_instance.hh
@@ -19,6 +19,7 @@
 #include "eevee_cryptomatte.hh"
 #include "eevee_depth_of_field.hh"
 #include "eevee_film.hh"
+#include "eevee_gbuffer.hh"
 #include "eevee_hizbuffer.hh"
 #include "eevee_irradiance_cache.hh"
 #include "eevee_light.hh"
@@ -54,6 +55,7 @@ class Instance {
   MotionBlurModule motion_blur;
   DepthOfField depth_of_field;
   Cryptomatte cryptomatte;
+  GBuffer gbuffer;
   HiZBuffer hiz_buffer;
   Sampling sampling;
   Camera camera;
diff --git a/source/blender/draw/engines/eevee_next/eevee_material.cc b/source/blender/draw/engines/eevee_next/eevee_material.cc
index 43c91dc32d5..a210681517a 100644
--- a/source/blender/draw/engines/eevee_next/eevee_material.cc
+++ b/source/blender/draw/engines/eevee_next/eevee_material.cc
@@ -193,11 +193,6 @@ MaterialPass MaterialModule::material_pass_get(Object *ob,
     inst_.sampling.reset();
   }
 
-  if ((pipeline_type == MAT_PIPE_DEFERRED) &&
-      GPU_material_flag_get(matpass.gpumat, GPU_MATFLAG_SHADER_TO_RGBA)) {
-    pipeline_type = MAT_PIPE_FORWARD;
-  }
-
   if (ELEM(pipeline_type,
            MAT_PIPE_FORWARD,
            MAT_PIPE_FORWARD_PREPASS,
@@ -240,10 +235,6 @@ Material &MaterialModule::material_sync(Object *ob,
                                        (has_motion ? MAT_PIPE_DEFERRED_PREPASS_VELOCITY :
                                                      MAT_PIPE_DEFERRED_PREPASS);
 
-  /* TEST until we have deferred pipeline up and running. */
-  surface_pipe = MAT_PIPE_FORWARD;
-  prepass_pipe = has_motion ? MAT_PIPE_FORWARD_PREPASS_VELOCITY : MAT_PIPE_FORWARD_PREPASS;
-
   MaterialKey material_key(blender_mat, geometry_type, surface_pipe);
 
   Material &mat = material_map_.lookup_or_add_cb(material_key, [&]() {
diff --git a/source/blender/draw/engines/eevee_next/eevee_pipeline.cc b/source/blender/draw/engines/eevee_next/eevee_pipeline.cc
index 94ce44522ba..8b6f0ccfdce 100644
--- a/source/blender/draw/engines/eevee_next/eevee_pipeline.cc
+++ b/source/blender/draw/engines/eevee_next/eevee_pipeline.cc
@@ -278,4 +278,229 @@ void ForwardPipeline::render(View &view,
 
 /** \} */
 
+/* -------------------------------------------------------------------- */
+/** \name Deferred Layer
+ * \{ */
+
+void DeferredLayer::begin_sync()
+{
+  {
+    prepass_ps_.init();
+    {
+      /* Common resources. */
+
+      /* Textures. */
+      prepass_ps_.bind_texture(RBUFS_UTILITY_TEX_SLOT, inst_.pipelines.utility_tx);
+      /* Uniform Buf. */
+      prepass_ps_.bind_ubo(CAMERA_BUF_SLOT, inst_.camera.ubo_get());
+
+      inst_.velocity.bind_resources(&prepass_ps_);
+      inst_.sampling.bind_resources(&prepass_ps_);
+    }
+
+    DRWState state_depth_only = DRW_STATE_WRITE_DEPTH | DRW_STATE_DEPTH_LESS;
+    DRWState state_depth_color = DRW_STATE_WRITE_DEPTH | DRW_STATE_DEPTH_LESS |
+                                 DRW_STATE_WRITE_COLOR;
+
+    prepass_double_sided_static_ps_ = &prepass_ps_.sub("DoubleSided.Static");
+    prepass_double_sided_static_ps_->state_set(state_depth_only);
+
+    prepass_single_sided_static_ps_ = &prepass_ps_.sub("SingleSided.Static");
+    prepass_single_sided_static_ps_->state_set(state_depth_only | DRW_STATE_CULL_BACK);
+
+    prepass_double_sided_moving_ps_ = &prepass_ps_.sub("DoubleSided.Moving");
+    prepass_double_sided_moving_ps_->state_set(state_depth_color);
+
+    prepass_single_sided_moving_ps_ = &prepass_ps_.sub("SingleSided.Moving");
+    prepass_single_sided_moving_ps_->state_set(state_depth_color | DRW_STATE_CULL_BACK);
+  }
+  {
+    gbuffer_ps_.init();
+    gbuffer_ps_.clear_stencil(0x00u);
+    gbuffer_ps_.state_stencil(0x01u, 0x01u, 0x01u);
+
+    {
+      /* Common resources. */
+
+      /* G-buffer. */
+      gbuffer_ps_.bind_image(GBUF_CLOSURE_SLOT, &inst_.gbuffer.closure_tx);
+      gbuffer_ps_.bind_image(GBUF_COLOR_SLOT, &inst_.gbuffer.color_tx);
+      /* RenderPasses. */
+      gbuffer_ps_.bind_image(RBUFS_NORMAL_SLOT, &inst_.render_buffers.normal_tx);
+      /* TODO(fclem): Pack all render pass into the same texture. */
+      // gbuffer_ps_.bind_image(RBUFS_DIFF_COLOR_SLOT, &inst_.render_buffers.diffuse_color_tx);
+      gbuffer_ps_.bind_image(RBUFS_SPEC_COLOR_SLOT, &inst_.render_buffers.specular_color_tx);
+      gbuffer_ps_.bind_image(RBUFS_EMISSION_SLOT, &inst_.render_buffers.emission_tx);
+      /* AOVs. */
+      gbuffer_ps_.bind_image(RBUFS_AOV_COLOR_SLOT, &inst_.render_buffers.aov_color_tx);
+      gbuffer_ps_.bind_image(RBUFS_AOV_VALUE_SLOT, &inst_.render_buffers.aov_value_tx);
+      /* Cryptomatte. */
+      gbuffer_ps_.bind_image(RBUFS_CRYPTOMATTE_SLOT, &inst_.render_buffers.cryptomatte_tx);
+      /* Storage Buf. */
+      gbuffer_ps_.bind_ssbo(RBUFS_AOV_BUF_SLOT, &inst_.film.aovs_info);
+      /* Textures. */
+      gbuffer_ps_.bind_texture(RBUFS_UTILITY_TEX_SLOT, inst_.pipelines.utility_tx);
+      /* Uniform Buf. */
+      gbuffer_ps_.bind_ubo(CAMERA_BUF_SLOT, inst_.camera.ubo_get());
+
+      inst_.sampling.bind_resources(&gbuffer_ps_);
+      inst_.cryptomatte.bind_resources(&gbuffer_ps_);
+    }
+
+    DRWState state = DRW_STATE_WRITE_COLOR | DRW_STATE_BLEND_CUSTOM | DRW_STATE_DEPTH_EQUAL |
+                     DRW_STATE_WRITE_STENCIL | DRW_STATE_STENCIL_ALWAYS;
+
+    gbuffer_double_sided_ps_ = &gbuffer_ps_.sub("DoubleSided");
+    gbuffer_double_sided_ps_->state_set(state);
+
+    gbuffer_single_sided_ps_ = &gbuffer_ps_.sub("SingleSided");
+    gbuffer_single_sided_ps_->state_set(state | DRW_STATE_CULL_BACK);
+  }
+}
+
+void DeferredLayer::end_sync()
+{
+  /* Use stencil test to reject pixel not written by this layer. */
+  /* WORKAROUND: Stencil write is only here to avoid rasterizer discard. */
+  DRWState state = DRW_STATE_WRITE_STENCIL | DRW_STATE_STENCIL_EQUAL;
+  /* Allow output to combined pass for the last pass. */
+  DRWState state_write_color = state | DRW_STATE_WRITE_COLOR | DRW_STATE_BLEND_CUSTOM;
+
+  if (closure_bits_ & (CLOSURE_DIFFUSE | CLOSURE_REFLECTION)) {
+    const bool is_last_eval_pass = true;
+
+    eval_light_ps_.init();
+    eval_light_ps_.state_set(is_last_eval_pass ? state_write_color : state);
+    eval_light_ps_.state_stencil(0x00u, 0x01u, 0xFFu);
+    eval_light_ps_.shader_set(inst_.shaders.static_shader_get(DEFERRED_LIGHT));
+    eval_light_ps_.bind_image("out_diffuse_light_img", &diffuse_light_tx_);
+    eval_light_ps_.bind_image("out_specular_light_img", &specular_light_tx_);
+    eval_light_ps_.bind_texture("gbuffer_closure_tx", &inst_.gbuffer.closure_tx);
+    eval_light_ps_.bind_texture("gbuffer_color_tx", &inst_.gbuffer.color_tx);
+    eval_light_ps_.push_constant("is_last_eval_pass", is_last_eval_pass);
+    eval_light_ps_.bind_image(RBUFS_LIGHT_SLOT, &inst_.render_buffers.light_tx);
+    eval_light_ps_.bind_texture(RBUFS_UTILITY_TEX_SLOT, inst_.pipelines.utility_tx);
+
+    inst_.lights.bind_resources(&eval_light_ps_);
+    inst_.shadows.bind_resources(&eval_light_ps_);
+    inst_.sampling.bind_resources(&eval_light_ps_);
+    inst_.hiz_buffer.bind_resources(&eval_light_ps_);
+
+    eval_light_ps_.barrier(GPU_BARRIER_TEXTURE_FETCH | GPU_BARRIER_SHADER_IMAGE_ACCESS);
+    eval_light_ps_.draw_procedural(GPU_PRIM_TRIS, 1, 3);
+  }
+}
+
+PassMain::Sub *DeferredLayer::prepass_add(::Material *blender_mat,
+                                          GPUMaterial *gpumat,
+                                          bool has_motion)
+{
+  PassMain::Sub *pass = (blender_mat->blend_flag & MA_BL_CULL_BACKFACE) ?
+                            (has_motion ? prepass_single_sided_moving_ps_ :
+                                          prepass_single_sided_static_ps_) :
+                            (has_motion ? prepass_double_sided_moving_ps_ :
+                                          prepass_double_sided_static_ps_);
+
+  return &pass->sub(GPU_material_get_name(gpumat));
+}
+
+PassMain::Sub *DeferredLayer::material_add(::Material *blender_mat, GPUMaterial *gpumat)
+{
+  closure_bits_ |= shader_closure_bits_from_flag(gpumat);
+
+  PassMain::Sub *pass = (blender_mat->blend_flag & MA_BL_CULL_BACKFACE) ?
+                            gbuffer_single_sided_ps_ :
+                            gbuffer_double_sided_ps_;
+  return &pass->sub(GPU_material_get_name(gpumat));
+}
+
+void DeferredLayer::render(View &view,
+                           Framebuffer &prepass_fb,
+                           Framebuffer &combined_fb,
+                           int2 extent)
+{
+
+  GPU_framebuffer_bind(prepass_fb);
+  inst_.manager->submit(prepass_ps_, view);
+
+  inst_.hiz_buffer.set_dirty();
+  inst_.shadows.set_view(view);
+
+  inst_.gbuffer.acquire(extent, closure_bits_);
+
+  GPU_framebuffer_bind(combined_fb);
+  inst_.manager->submit(gbuffer_ps_, view);
+
+  eGPUTextureUsage usage = GPU_TEXTURE_USAGE_SHADER_READ | GPU_TEXTURE_USAGE_SHADER_WRITE;
+  diffuse_light_tx_.acquire(extent, GPU_RGBA16F, usage);
+  specular_light_tx_.acquire(extent, GPU_RGBA16F, usage);
+  diffuse_light_tx_.clear(float4(0.0f));
+  specular_light_tx_.clear(float4(0.0f));
+
+  inst_.manager->submit(eval_light_ps_, view);
+
+  diffuse_light_tx_.release();
+  specular_light_tx_.release();
+
+  inst_.gbuffer.release();
+}
+
+/** \} */
+
+/* -------------------------------------------------------------------- */
+/** \name Deferred Pipeline
+ *
+ * Closure data are written to intermediate buffer allowing screen space processing.
+ * \{ */
+
+void DeferredPipeline::begin_sync()
+{
+  opaque_layer_.begin_sync();
+  refraction_layer_.begin_sync();
+}
+
+void DeferredPipeline::end_sync()
+{
+  opaque_layer_.end_sync();
+  refraction_layer_.end_sync();
+}
+
+PassMain::Sub *DeferredPipeline::prepass_add(::Material *blender_mat,
+                                             GPUMaterial *gpumat,
+                                             bool has_motion)
+{
+  if (blender_mat->blend_flag & MA_BL_SS_REFRACTION) {
+    return refraction_layer_.prepass_add(blender_mat, gpumat, has_motion);
+  }
+  else {
+    return opaque_layer_.prepass_add(blender_mat, gpumat, has_motion);
+  }
+}
+
+PassMain::Sub *DeferredPipeline::material_add(::Material *blender_mat, GPUMaterial *gpumat)
+{
+  if (blender_mat->blend_flag & MA_BL_SS_REFRACTION) {
+    return refraction_layer_.material_add(blender_mat, gpumat);
+  }
+  else {
+    return opaque_layer_.material_add(blender_mat, gpumat);
+  }
+}
+
+void DeferredPipeline::render(View &view,
+                              Framebuffer &prepass_fb,
+                              Framebuffer &combined_fb,
+                              int2 extent)
+{
+  DRW_stats_group_start("Deferred.Opaque");
+  opaque_layer_.render(view, prepass_fb, combined_fb, extent);
+  DRW_stats_group_end();
+
+  DRW_stats_group_start("Deferred.Refract");
+  refraction_layer_.render(view, prepass_fb, combined_fb, extent);
+  DRW_stats_group_end();
+}
+
+/** \} */
+
 }  // namespace blender::eevee
diff --git a/source/blender/draw/engines/eevee_next/eevee_pipeline.hh b/source/blender/draw/engines/eevee_next/eevee_pipeline.hh
index a437ca8c522..d390fad4775 100644
--- a/source/blender/draw/engines/eevee_next/eevee_pipeline.hh
+++ b/source/blender/draw/engines/eevee_next/eevee_pipeline.hh
@@ -113,6 +113,77 @@ class ForwardPipeline {
 
 /** \} */
 
+/* -------------------------------------------------------------------- */
+/** \name Deferred lighting.
+ * \{ */
+
+class DeferredLayer {
+ private:
+  Instance &inst_;
+
+  PassMain prepass_ps_ = {"Prepass"};
+  PassMain::Sub *prepass_single_sided_static_ps_ = nullptr;
+  PassMain::Sub *prepass_single_sided_moving_ps_ = nullptr;
+  PassMain::Sub *prepass_double_sided_static_ps_ = nullptr;
+  PassMain::Sub *prepass_double_sided_moving_ps_ = nullptr;
+
+  PassMain gbuffer_ps_ = {"Shading"};
+  PassMain::Sub *gbuffer_single_sided_ps_ = nullptr;
+  PassMain::Sub *gbuffer_double_sided_ps_ = nullptr;
+
+  PassSimple eval_light_ps_ = {"EvalLights"};
+
+  /* Closures bits from the materials in this pass. */
+  eClosureBits closure_bits_;
+
+  /**
+   * Accumulation textures for all stages of lighting evaluation (Light, SSR, SSSS, SSGI ...).
+   * These are split and separate from the main radiance buffer in order to accumulate light for
+   * the render passes and avoid too much bandwidth waste. Otherwise, we would have to load the
+   * BSDF color and do additive blending for each of the lighting step.
+   *
+   * NOTE: Not to be confused with the render passes.
+   */
+  TextureFromPool diffuse_light_tx_ = {"diffuse_light_accum_tx"};
+  TextureFromPool specular_light_tx_ = {"specular_light_accum_tx"};
+
+ public:
+  DeferredLayer(Instance &inst) : inst_(inst){};
+
+  void begin_sync();
+  void end_sync();
+
+  PassMain::Sub *prepass_add(::Material *blender_mat, GPUMaterial *gpumat, bool has_motion);
+  PassMain::Sub *material_add(::Material *blender_mat, GPUMaterial *gpumat);
+
+  void render(View &view, Framebuffer &prepass_fb, Framebuffer &combined_fb, int2 extent);
+};
+
+class DeferredPipeline {
+ private:
+  Instance &inst_;
+
+  /* Gbuffer filling passes. We could have an arbitrary number of them but for now we just have
+   * a hardcoded number of them. */
+  DeferredLayer opaque_layer_;
+  DeferredLayer refraction_layer_;
+  DeferredLayer volumetric_layer_;
+
+ public:
+  DeferredPipeline(Instance &inst)
+      : inst_(inst), opaque_layer_(inst), refraction_layer_(inst), volumetric_layer_(inst){};
+
+  void begin_sync();
+  void end_sync();
+
+  PassMain::Sub *prepass_add(::Material *material, GPUMaterial *gpumat, bool has_motion);
+  PassMain::Sub *material_add(::Material *material, GPUMaterial *gpumat);
+
+  void render(View &view, Framebuffer &prepass_fb, Framebuffer &combined_fb, int2 extent);
+};
+
+/** \} */
+
 /* -------------------------------------------------------------------- */
 /** \name Utility texture
  *
@@ -197,22 +268,25 @@ class UtilityTexture : public Texture {
 class PipelineModule {
  public:
   WorldPipeline world;
-  // DeferredPipeline deferred;
+  DeferredPipeline deferred;
   ForwardPipeline forward;
   ShadowPipeline shadow;
-  // VelocityPipeline velocity;
 
   UtilityTexture utility_tx;
 
  public:
-  PipelineModule(Instance &inst) : world(inst), forward(inst), shadow(inst){};
+  PipelineModule(Instance &inst) : world(inst), deferred(inst), forward(inst), shadow(inst){};
 
-  void sync()
+  void begin_sync()
   {
-    // deferred.sync();
+    deferred.begin_sync();
     forward.sync();
     shadow.sync();
-    // velocity.sync();
+  }
+
+  void end_sync()
+  {
+    deferred.end_sync();
   }
 
   PassMain::Sub *material_add(Object *ob,
@@ -222,7 +296,7 @@ class PipelineModule {
   {
     switch (pipeline_type) {
       case MAT_PIPE_DEFERRED_PREPASS:
-        // return deferred.prepass_add(blender_mat, gpumat, false);
+        return deferred.prepass_add(blender_mat, gpumat, false);
       case MAT_PIPE_FORWARD_PREPASS:
         if (GPU_material_flag_get(gpumat, GPU_MATFLAG_TRANSPARENT)) {
           return forward.prepass_transparent_add(ob, blender_mat, gpumat);
@@ -230,7 +304,7 @@ class PipelineModule {
         return forward.prepass_opaque_add(blender_mat, gpumat, false);
 
       case MAT_PIPE_DEFERRED_PREPASS_VELOCITY:
-        // return deferred.prepass_add(blender_mat, gpumat, true);
+        return deferred.prepass_add(blender_mat, gpumat, true);
       case MAT_PIPE_FORWARD_PREPASS_VELOCITY:
         if (GPU_material_flag_get(gpumat, GPU_MATFLAG_TRANSPARENT)) {
           return forward.prepass_transparent_add(ob, blender_mat, gpumat);
@@ -238,7 +312,7 @@ class PipelineModule {
         return forward.prepass_opaque_add(blender_mat, gpumat, true);
 
       case MAT_PIPE_DEFERRED:
-        // return deferred.material_add(blender_mat, gpumat);
+        return deferred.material_add(blender_mat, gpumat);
       case MAT_PIPE_FORWARD:
         if (GPU_material_flag_get(gpumat, GPU_MATFLAG_TRANSPARENT)) {
           return forward.material_transparent_add(ob, blender_mat, gpumat);
diff --git a/source/blender/draw/engines/eevee_next/eevee_shader.cc b/source/blender/draw/engines/eevee_next/eevee_shader.cc
index 4d9ac75687c..116d6004642 100644
--- a/source/blender/draw/engines/eevee_next/eevee_shader.cc
+++ b/source/blender/draw/engines/eevee_next/eevee_shader.cc
@@ -86,6 +86,8 @@ const char *ShaderModule::static_shader_create_info_name_get(eShaderType shader_
       return "eevee_film_comp";
     case FILM_CRYPTOMATTE_POST:
       return "eevee_film_cryptomatte_post";
+    case DEFERRED_LIGHT:
+      return "eevee_deferred_light";
     case HIZ_DEBUG:
       return "eevee_hiz_debug";
     case HIZ_UPDATE:
@@ -241,6 +243,11 @@ void ShaderModule::material_create_info_ammend(GPUMaterial *gpumat, GPUCodegenOu
     }
   }
 
+  /* WORKAROUND: Needed because node_tree isn't present in test shaders. */
+  if (pipeline_type == MAT_PIPE_DEFERRED) {
+    info.define("MAT_RENDER_PASS_SUPPORT");
+  }
+
   if (GPU_material_flag_get(gpumat, GPU_MATFLAG_TRANSPARENT)) {
     info.define("MAT_TRANSPARENT");
     /* Transparent material do not have any velocity specific pipeline. */
diff --git a/source/blender/draw/engines/eevee_next/eevee_shader.hh b/source/blender/draw/engines/eevee_next/eevee_shader.hh
index 173dc7bfa86..d987105e3fc 100644
--- a/source/blender/draw/engines/eevee_next/eevee_shader.hh
+++ b/source/blender/draw/engines/eevee_next/eevee_shader.hh
@@ -30,6 +30,8 @@ enum eShaderType {
   FILM_COMP,
   FILM_CRYPTOMATTE_POST,
 
+  DEFERRED_LIGHT,
+
   DEBUG_SURFELS,
 
   DOF_BOKEH_LUT,
diff --git a/source/blender/draw/engines/eevee_next/eevee_view.cc b/source/blender/draw/engines/eevee_next/eevee_view.cc
index 8037af78db1..d2d8f80feb7 100644
--- a/source/blender/draw/engines/eevee_next/eevee_view.cc
+++ b/source/blender/draw/engines/eevee_next/eevee_view.cc
@@ -123,8 +123,7 @@ void ShadingView::render()
   /* TODO(fclem): Move it after the first prepass (and hiz update) once pipeline is stabilized. */
   inst_.lights.set_view(render_view_new_, extent_);
 
-  // inst_.pipelines.deferred.render(
-  //     render_view_, rt_buffer_opaque_, rt_buffer_refract_, depth_tx_, combined_tx_);
+  inst_.pipelines.deferred.render(render_view_new_, prepass_fb_, combined_fb_, extent_);
 
   // inst_.lightprobes.draw_cache_display();
 
diff --git a/source/blender/draw/engines/eevee_next/shaders/eevee_deferred_light_frag.glsl b/source/blender/draw/engines/eevee_next/shaders/eevee_deferred_light_frag.glsl
new file mode 100644
index 00000000000..80dcaaf67cb
--- /dev/null
+++ b/source/blender/draw/engines/eevee_next/shaders/eevee_deferred_light_frag.glsl
@@ -0,0 +1,97 @@
+
+/**
+ * Compute light objects lighting contribution using Gbuffer data.
+ *
+ * Output light either directly to the radiance buffers or to temporary radiance accumulation
+ * buffer that will be processed by other deferred lighting passes.
+ */
+
+#pragma BLENDER_REQUIRE(eevee_gbuffer_lib.glsl)
+#pragma BLENDER_REQUIRE(common_view_lib.glsl)
+#pragma BLENDER_REQUIRE(eevee_light_eval_lib.glsl)
+
+void main()
+{
+  ivec2 texel = ivec2(gl_FragCoord.xy);
+
+  float depth = texelFetch(hiz_tx, ivec2(gl_FragCoord.xy), 0).r;
+  vec3 P = get_world_space_from_depth(uvcoordsvar.xy, depth);
+
+  /* TODO(fclem): High precision derivative. */
+  vec3 Ng = safe_normalize(cross(dFdx(P), dFdy(P)));
+  vec3 V = cameraVec(P);
+  float vP_z = dot(cameraForward, P) - dot(cameraForward, cameraPos);
+
+  vec4 gbuffer_0_packed = texelFetch(gbuffer_closure_tx, ivec3(texel, 0), 0);
+  vec4 gbuffer_1_packed = texelFetch(gbuffer_closure_tx, ivec3(texel, 1), 0);
+
+  ClosureReflection reflection_data;
+  reflection_data.N = gbuffer_normal_unpack(gbuffer_0_packed.xy);
+  reflection_data.roughness = gbuffer_0_packed.z;
+
+  ClosureDiffuse diffuse_data;
+  diffuse_data.N = gbuffer_normal_unpack(gbuffer_1_packed.xy);
+  /* These are only set for SSS case. */
+  diffuse_data.sss_radius = vec3(0.0);
+  diffuse_data.sss_id = 0u;
+  float thickness = 0.0;
+
+  bool is_refraction = gbuffer_is_refraction(gbuffer_1_packed);
+  if (is_refraction) {
+    /* Still evaluate the diffuse light so that dithered SSS / Refraction combination still
+     * produces a complete diffuse light buffer that will be correctly convolved by the SSSS.
+     * The refraction pixels will just set the diffuse radiance to 0. */
+  }
+  else if (false /* TODO */) {
+    vec4 gbuffer_2_packed = texelFetch(gbuffer_closure_tx, ivec3(texel, 2), 0);
+    diffuse_data.sss_radius = gbuffer_sss_radii_unpack(gbuffer_2_packed.xyz);
+    diffuse_data.sss_id = gbuffer_object_id_unorm16_unpack(gbuffer_2_packed.w);
+    thickness = gbuffer_thickness_pack(gbuffer_1_packed.z);
+  }
+
+  vec3 diffuse_light = vec3(0.0);
+  vec3 reflection_light = vec3(0.0);
+
+  light_eval(
+      diffuse_data, reflection_data, P, Ng, V, vP_z, thickness, diffuse_light, reflection_light);
+
+  if (is_last_eval_pass) {
+    /* Apply color and output lighting to render-passes. */
+    vec4 color_0_packed = texelFetch(gbuffer_color_tx, ivec3(texel, 0), 0);
+    vec4 color_1_packed = texelFetch(gbuffer_color_tx, ivec3(texel, 1), 0);
+
+    reflection_data.color = gbuffer_color_unpack(color_0_packed);
+    diffuse_data.color = gbuffer_color_unpack(color_1_packed);
+
+    if (is_refraction) {
+      diffuse_data.color = vec3(0.0);
+    }
+
+    reflection_light *= reflection_data.color;
+    diffuse_light *= diffuse_data.color;
+    /* Add radiance to light pass. */
+    imageStore(
+        rp_light_img, ivec3(texel, RENDER_PASS_LAYER_DIFFUSE_LIGHT), vec4(diffuse_light, 1.0));
+    imageStore(
+        rp_light_img, ivec3(texel, RENDER_PASS_LAYER_SPECULAR_LIGHT), vec4(reflection_light, 1.0));
+    /* Add radiance to combined pass. */
+    out_radiance = vec4(diffuse_light + reflection_light, 0.0);
+    out_transmittance = vec4(1.0);
+  }
+  else {
+    /* Store lighting for next deferred pass. */
+
+    /* Output diffuse light along with object ID for sub-surface screen space processing. */
+    vec4 diffuse_radiance;
+    diffuse_radiance.xyz = diffuse_light;
+    diffuse_radiance.w = gbuffer_object_id_f16_pack(diffuse_data.sss_id);
+    imageStore(out_diffuse_light_img, texel, diffuse_radiance);
+
+    imageStore(out_specular_light_img, texel, vec4(reflection_light, 0.0));
+
+    /* Final radiance will be amended by the last pass.
+     * This should do nothing as color write should be disabled in this case. */
+    out_radiance = vec4(0.0);
+    out_transmittance = vec4(0.0);
+  }
+}
diff --git a/source/blender/draw/engines/eevee_next/shaders/eevee_gbuffer_lib.glsl b/source/blender/draw/engines/eevee_next/shaders/eevee_gbuffer_lib.glsl
new file mode 100644
index 00000000000..8156166f42f
--- /dev/null
+++ b/source/blender/draw/engines/eevee_next/shaders/eevee_gbuffer_lib.glsl
@@ -0,0 +1,106 @@
+
+/**
+ * G-buffer: Packing and upacking of G-buffer data.
+ *
+ * See #GBuffer for a breakdown of the G-buffer layout.
+ */
+
+#pragma BLENDER_REQUIRE(gpu_shader_math_vector_lib.glsl)
+
+vec2 gbuffer_normal_pack(vec3 N)
+{
+  N /= length_manhattan(N);
+  N.xy = (N.z >= 0.0) ? N.xy : ((1.0 - abs(N.yx)) * sign(N.xy));
+  N.xy = N.xy * 0.5 + 0.5;
+  return N.xy;
+}
+
+vec3 gbuffer_normal_unpack(vec2 N_packed)
+{
+  N_packed = N_packed * 2.0 - 1.0;
+  vec3 N = vec3(N_packed.x, N_packed.y, 1.0 - abs(N_packed.x) - abs(N_packed.y));
+  float t = clamp(-N.z, 0.0, 1.0);
+  N.x += (N.x >= 0.0) ? -t : t;
+  N.y += (N.y >= 0.0) ? -t : t;
+  return normalize(N);
+}
+
+float gbuffer_ior_pack(float ior)
+{
+  return (ior > 1.0) ? (1.0 - 0.5 / ior) : (0.5 * ior);
+}
+
+float gbuffer_ior_unpack(float ior_packed)
+{
+  return (ior_packed > 0.5) ? (-1.0 / (ior_packed * 2.0 + 2.0)) : (2.0 * ior_packed);
+}
+
+float gbuffer_thickness_pack(float thickness)
+{
+  /* TODO(fclem): Something better. */
+  return gbuffer_ior_pack(thickness);
+}
+
+float gbuffer_thickness_unpack(float thickness_packed)
+{
+  /* TODO(fclem): Something better. */
+  return gbuffer_ior_unpack(thickness_packed);
+}
+
+vec3 gbuffer_sss_radii_pack(vec3 sss_radii)
+{
+  /* TODO(fclem): Something better. */
+  return vec3(
+      gbuffer_ior_pack(sss_radii.x), gbuffer_ior_pack(sss_radii.y), gbuffer_ior_pack(sss_radii.z));
+}
+
+vec3 gbuffer_sss_radii_unpack(vec3 sss_radii_packed)
+{
+  /* TODO(fclem): Something better. */
+  return vec3(gbuffer_ior_unpack(sss_radii_packed.x),
+              gbuffer_ior_unpack(sss_radii_packed.y),
+              gbuffer_ior_unpack(sss_radii_packed.z));
+}
+
+vec4 gbuffer_color_pack(vec3 color)
+{
+  float max_comp = max(color.x, max(color.y, color.z));
+  /* Store 2bit exponent inside Alpha. Allows values up to 8 with some color degradation.
+   * Above 8, the result will be clampped when writing the data to the output buffer. */
+  float exponent = (max_comp > 1) ? ((max_comp > 2) ? ((max_comp > 4) ? 3.0 : 2.0) : 1.0) : 0.0;
+  /* TODO(fclem): Could try dithering to avoid banding artifacts on higher exponents. */
+  return vec4(color / exp2(exponent), exponent / 3.0);
+}
+
+vec3 gbuffer_color_unpack(vec4 color_packed)
+{
+  float exponent = color_packed.a * 3.0;
+  return color_packed.rgb * exp2(exponent);
+}
+
+float gbuffer_object_id_unorm16_pack(uint object_id)
+{
+  return float(object_id & 0xFFFFu) / float(0xFFFF);
+}
+
+uint gbuffer_object_id_unorm16_unpack(float object_id_packed)
+{
+  return uint(object_id_packed * float(0xFFFF));
+}
+
+float gbuffer_object_id_f16_pack(uint object_id)
+{
+  /* TODO(fclem): Make use of all the 16 bits in a half float.
+   * This here only correctly represent values up to 1024. */
+  return float(object_id);
+}
+
+uint gbuffer_object_id_f16_unpack(float object_id_packed)
+{
+  return uint(object_id_packed);
+}
+
+bool gbuffer_is_refraction(vec4 gbuffer)
+{
+  return gbuffer.w < 1.0;
+}
\ No newline at end of file
diff --git a/source/blender/draw/engines/eevee_next/shaders/eevee_nodetree_lib.glsl b/source/blender/draw/engines/eevee_next/shaders/eevee_nodetree_lib.glsl
index 965780d9bcf..db38baab6a4 100644
--- a/source/blender/draw/engines/eevee_next/shaders/eevee_nodetree_lib.glsl
+++ b/source/blender/draw/engines/eevee_next/shaders/eevee_nodetree_lib.glsl
@@ -29,14 +29,16 @@ bool closure_select(float weight, inout float total_weight, inout float r)
   float x = weight / total_weight;
   bool chosen = (r < x);
   /* Assuming that if r is in the interval [0,x] or [x,1], it's still uniformly distributed within
-   * that interval, so you remapping to [0,1] again to explore this space of probability. */
+   * that interval, so remapping to [0,1] again to explore this space of probability. */
   r = (chosen) ? (r / x) : ((r - x) / (1.0 - x));
   return chosen;
 }
 
 #define SELECT_CLOSURE(destination, random, candidate) \
   if (closure_select(candidate.weight, destination.weight, random)) { \
+    float tmp = destination.weight; \
     destination = candidate; \
+    destination.weight = tmp; \
   }
 
 float g_closure_rand;
diff --git a/source/blender/draw/engines/eevee_next/shaders/eevee_shadow_tag_usage_frag.glsl b/source/blender/draw/engines/eevee_next/shaders/eevee_shadow_tag_usage_frag.glsl
index dd946ffb637..7bc51513da1 100644
--- a/source/blender/draw/engines/eevee_next/shaders/eevee_shadow_tag_usage_frag.glsl
+++ b/source/blender/draw/engines/eevee_next/shaders/eevee_shadow_tag_usage_frag.glsl
@@ -82,7 +82,7 @@ void main()
 {
   vec2 screen_uv = gl_FragCoord.xy / vec2(fb_resolution);
 
-  float opaque_depth = texelFetch(hiz_tx, int2(gl_FragCoord.xy), fb_lod).r;
+  float opaque_depth = texelFetch(hiz_tx, ivec2(gl_FragCoord.xy), fb_lod).r;
   vec3 ws_opaque = get_world_space_from_depth(screen_uv, opaque_depth);
 
   vec3 ws_near_plane = get_world_space_from_depth(screen_uv, 0);
diff --git a/source/blender/draw/engines/eevee_next/shaders/eevee_surf_deferred_frag.glsl b/source/blender/draw/engines/eevee_next/shaders/eevee_surf_deferred_frag.glsl
index 7848c4a0611..d84d0a030e2 100644
--- a/source/blender/draw/engines/eevee_next/shaders/eevee_surf_deferred_frag.glsl
+++ b/source/blender/draw/engines/eevee_next/shaders/eevee_surf_deferred_frag.glsl
@@ -3,17 +3,136 @@
  * Deferred lighting evaluation: Lighting is evaluated in a separate pass.
  *
  * Outputs shading parameter per pixel using a randomized set of BSDFs.
- **/
+ * Some render-pass are written during this pass.
+ */
 
+#pragma BLENDER_REQUIRE(eevee_gbuffer_lib.glsl)
 #pragma BLENDER_REQUIRE(common_view_lib.glsl)
 #pragma BLENDER_REQUIRE(common_math_lib.glsl)
 #pragma BLENDER_REQUIRE(common_hair_lib.glsl)
 #pragma BLENDER_REQUIRE(eevee_surf_lib.glsl)
 #pragma BLENDER_REQUIRE(eevee_nodetree_lib.glsl)
+#pragma BLENDER_REQUIRE(eevee_sampling_lib.glsl)
+
+vec4 closure_to_rgba(Closure cl)
+{
+  vec4 out_color;
+  out_color.rgb = g_emission;
+  out_color.a = saturate(1.0 - avg(g_transmittance));
+
+  /* Reset for the next closure tree. */
+  closure_weights_reset();
+
+  return out_color;
+}
 
 void main()
 {
   init_globals();
 
+  float noise = utility_tx_fetch(utility_tx, gl_FragCoord.xy, UTIL_BLUE_NOISE_LAYER).r;
+  g_closure_rand = fract(noise + sampling_rng_1D_get(SAMPLING_CLOSURE));
+
+  fragment_displacement();
+
   nodetree_surface();
+
+  g_holdout = saturate(g_holdout);
+
+  out_transmittance = vec4(1.0 - g_holdout);
+  float transmittance_mono = saturate(avg(g_transmittance));
+
+  float thickness = nodetree_thickness();
+
+  g_diffuse_data.color *= g_diffuse_data.weight;
+  g_reflection_data.color *= g_reflection_data.weight;
+  g_refraction_data.color *= g_refraction_data.weight;
+
+  /* TODO(fclem): This feels way too complex for what is it. */
+  bool has_any_bsdf_weight = g_diffuse_data.weight != 0.0 || g_reflection_data.weight != 0.0 ||
+                             g_refraction_data.weight != 0.0;
+  vec3 out_normal = has_any_bsdf_weight ? vec3(0.0) : g_data.N;
+  out_normal += g_diffuse_data.N * g_diffuse_data.weight;
+  out_normal += g_reflection_data.N * g_reflection_data.weight;
+  out_normal += g_refraction_data.N * g_refraction_data.weight;
+  out_normal = safe_normalize(out_normal);
+
+  vec3 specular_color = g_reflection_data.color + g_refraction_data.color;
+
+  /* ----- Render Passes output ----- */
+
+  ivec2 out_texel = ivec2(gl_FragCoord.xy);
+#ifdef MAT_RENDER_PASS_SUPPORT /* Needed because node_tree isn't present in test shaders. */
+  /* Some render pass can be written during the gbuffer pass. Light passes are written later. */
+  vec4 cryptomatte_output = vec4(cryptomatte_object_buf[resource_id], node_tree.crypto_hash, 0.0);
+  imageStore(rp_cryptomatte_img, out_texel, cryptomatte_output);
+  imageStore(rp_normal_img, out_texel, vec4(out_normal, 1.0));
+  /* TODO(fclem): For now, just don't do anything. In the future all render passes should be in an
+   * array texture and have a UBO with indirection to the correct layer. */
+  // imageStore(rp_diffuse_color_img, out_texel, vec4(g_diffuse_data.color, 1.0));
+  imageStore(rp_specular_color_img, out_texel, vec4(specular_color, 1.0));
+  imageStore(rp_emission_img, out_texel, vec4(g_emission, 1.0));
+#endif
+
+  /* ----- GBuffer output ----- */
+
+  if (true) {
+    /* Reflection. */
+    vec4 out_reflect = vec4(gbuffer_normal_pack(g_reflection_data.N),
+                            g_reflection_data.roughness,
+                            g_reflection_data.roughness);
+    imageStore(out_gbuff_closure_img, ivec3(out_texel, 0), out_reflect);
+
+    vec4 color = gbuffer_color_pack(g_reflection_data.color);
+    imageStore(out_gbuff_color_img, ivec3(out_texel, 0), color);
+  }
+
+  /* TODO(fclem) other RNG. */
+  float refract_rand = fract(g_closure_rand * 6.1803398875);
+  float combined_weight = g_refraction_data.weight + g_diffuse_data.weight;
+  bool output_refraction = combined_weight > 0.0 &&
+                           (refract_rand * combined_weight) < g_refraction_data.weight;
+  if (output_refraction) {
+    /* Refraction. */
+    vec4 closure;
+    closure.xy = gbuffer_normal_pack(g_refraction_data.N);
+    closure.z = g_refraction_data.roughness;
+    closure.w = gbuffer_ior_pack(g_refraction_data.ior);
+    /* Clamp to just bellow 1 to be able to distinguish between refraction and diffuse.
+     * Ceiling value is chosen by the storage format (16bit UNORM). */
+    closure.w = min(closure.w, float(0xFFFFu - 1u) / float(0xFFFFu));
+    imageStore(out_gbuff_closure_img, ivec3(out_texel, 1), closure);
+
+    vec4 color = gbuffer_color_pack(g_refraction_data.color);
+    imageStore(out_gbuff_color_img, ivec3(out_texel, 1), color);
+  }
+  else {
+    /* Diffuse. */
+    vec4 closure;
+    closure.xy = gbuffer_normal_pack(g_diffuse_data.N);
+    closure.z = gbuffer_thickness_pack(thickness);
+    /* Used to detect the refraction case. Could be used for roughness. */
+    closure.w = 1.0;
+    imageStore(out_gbuff_closure_img, ivec3(out_texel, 1), closure);
+
+    vec4 color = gbuffer_color_pack(g_diffuse_data.color);
+    imageStore(out_gbuff_color_img, ivec3(out_texel, 1), color);
+  }
+
+  if (true) {
+    /* SubSurface Scattering. */
+    vec4 closure;
+    closure.xyz = gbuffer_sss_radii_pack(g_diffuse_data.sss_radius);
+    closure.w = gbuffer_object_id_unorm16_pack(g_diffuse_data.sss_id);
+    imageStore(out_gbuff_closure_img, ivec3(out_texel, 2), closure);
+  }
+
+  /* ----- Radiance output ----- */
+
+  /* Only output emission during the gbuffer pass. */
+  out_radiance = vec4(g_emission, 0.0);
+  out_radiance.rgb *= 1.0 - g_holdout;
+
+  out_transmittance.rgb = g_transmittance;
+  out_transmittance.a = saturate(avg(g_transmittance));
 }
diff --git a/source/blender/draw/engines/eevee_next/shaders/infos/eevee_deferred_info.hh b/source/blender/draw/engines/eevee_next/shaders/infos/eevee_deferred_info.hh
new file mode 100644
index 00000000000..f5bc967e5b6
--- /dev/null
+++ b/source/blender/draw/engines/eevee_next/shaders/infos/eevee_deferred_info.hh
@@ -0,0 +1,47 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+
+#include "eevee_defines.hh"
+#include "gpu_shader_create_info.hh"
+
+#define image_out(slot, qualifier, format, name) \
+  image(slot, format, qualifier, ImageType::FLOAT_2D, name, Frequency::PASS)
+#define image_array_out(slot, qualifier, format, name) \
+  image(slot, format, qualifier, ImageType::FLOAT_2D_ARRAY, name, Frequency::PASS)
+
+/**
+ * Specific deferred pass accumulate the computed lighting to either:
+ * - a split diffuse / specular temporary light buffer.
+ * or to
+ * - the combined pass & the light render-pass (if needed).
+ *
+ * This is in order to minimize the number of blending step.
+ */
+GPU_SHADER_CREATE_INFO(eevee_deferred_base)
+    /* Early fragment test is needed to avoid processing fragments without correct GBuffer data. */
+    .early_fragment_test(true)
+    /* Select which output to write to. */
+    .push_constant(Type::BOOL, "is_last_eval_pass")
+    /* Combined pass output. */
+    .fragment_out(0, Type::VEC4, "out_radiance", DualBlend::SRC_0)
+    .fragment_out(0, Type::VEC4, "out_transmittance", DualBlend::SRC_1)
+    /* Light pass output. */
+    .image_array_out(RBUFS_LIGHT_SLOT, Qualifier::READ_WRITE, GPU_RGBA16F, "rp_light_img")
+    /* Chaining to next pass. */
+    .image_out(2, Qualifier::READ_WRITE, GPU_RGBA16F, "out_diffuse_light_img")
+    .image_out(3, Qualifier::READ_WRITE, GPU_RGBA16F, "out_specular_light_img");
+
+GPU_SHADER_CREATE_INFO(eevee_deferred_light)
+    .fragment_source("eevee_deferred_light_frag.glsl")
+    .sampler(0, ImageType::FLOAT_2D_ARRAY, "gbuffer_closure_tx")
+    .sampler(1, ImageType::FLOAT_2D_ARRAY, "gbuffer_color_tx")
+    .additional_info("eevee_shared",
+                     "eevee_utility_texture",
+                     "eevee_light_data",
+                     "eevee_shadow_data",
+                     "eevee_deferred_base",
+                     "eevee_hiz_data",
+                     "draw_view",
+                     "draw_fullscreen")
+    .do_static_compilation(true);
+
+#undef image_array_out
diff --git a/source/blender/draw/engines/eevee_next/shaders/infos/eevee_material_info.hh b/source/blender/draw/engines/eevee_next/shaders/infos/eevee_material_info.hh
index 0ef16bf3522..f16113a5efc 100644
--- a/source/blender/draw/engines/eevee_next/shaders/infos/eevee_material_info.hh
+++ b/source/blender/draw/engines/eevee_next/shaders/infos/eevee_material_info.hh
@@ -88,11 +88,11 @@ GPU_SHADER_CREATE_INFO(eevee_aov_out)
 
 GPU_SHADER_CREATE_INFO(eevee_render_pass_out)
     .define("MAT_RENDER_PASS_SUPPORT")
-    .image_out(RBUFS_NORMAL_SLOT, Qualifier::READ_WRITE, GPU_RGBA16F, "rp_normal_img")
-    .image_array_out(RBUFS_LIGHT_SLOT, Qualifier::READ_WRITE, GPU_RGBA16F, "rp_light_img")
-    .image_out(RBUFS_DIFF_COLOR_SLOT, Qualifier::READ_WRITE, GPU_RGBA16F, "rp_diffuse_color_img")
-    .image_out(RBUFS_SPEC_COLOR_SLOT, Qualifier::READ_WRITE, GPU_RGBA16F, "rp_specular_color_img")
-    .image_out(RBUFS_EMISSION_SLOT, Qualifier::READ_WRITE, GPU_RGBA16F, "rp_emission_img");
+    .image_out(RBUFS_NORMAL_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_normal_img")
+    .image_array_out(RBUFS_LIGHT_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_light_img")
+    .image_out(RBUFS_DIFF_COLOR_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_diffuse_color_img")
+    .image_out(RBUFS_SPEC_COLOR_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_specular_color_img")
+    .image_out(RBUFS_EMISSION_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_emission_img");
 
 GPU_SHADER_CREATE_INFO(eevee_cryptomatte_out)
     .storage_buf(CRYPTOMATTE_BUF_SLOT, Qualifier::READ, "vec2", "cryptomatte_object_buf[]")
@@ -101,23 +101,29 @@ GPU_SHADER_CREATE_INFO(eevee_cryptomatte_out)
 GPU_SHADER_CREATE_INFO(eevee_surf_deferred)
     .vertex_out(eevee_surf_iface)
     /* NOTE: This removes the possibility of using gl_FragDepth. */
-    // .early_fragment_test(true)
-    /* Direct output. */
+    .early_fragment_test(true)
+    /* Direct output. (Emissive, Holdout) */
     .fragment_out(0, Type::VEC4, "out_radiance", DualBlend::SRC_0)
     .fragment_out(0, Type::VEC4, "out_transmittance", DualBlend::SRC_1)
-    /* Gbuffer. */
-    // .image_out(0, Qualifier::WRITE, GPU_R11F_G11F_B10F, "gbuff_transmit_color")
-    // .image_out(1, Qualifier::WRITE, GPU_R11F_G11F_B10F, "gbuff_transmit_data")
-    // .image_out(2, Qualifier::WRITE, GPU_RGBA16F, "gbuff_transmit_normal")
-    // .image_out(3, Qualifier::WRITE, GPU_R11F_G11F_B10F, "gbuff_reflection_color")
-    // .image_out(4, Qualifier::WRITE, GPU_RGBA16F, "gbuff_reflection_normal")
-    // .image_out(5, Qualifier::WRITE, GPU_R11F_G11F_B10F, "gbuff_emission")
-    /* Render-passes. */
-    // .image_out(6, Qualifier::READ_WRITE, GPU_RGBA16F, "rpass_volume_light")
+    /* Everything is stored inside a two layered target, one for each format. This is to fit the
+     * limitation of the number of images we can bind on a single shader. */
+    .image_array_out(GBUF_CLOSURE_SLOT, Qualifier::WRITE, GPU_RGBA16, "out_gbuff_closure_img")
+    .image_array_out(GBUF_COLOR_SLOT, Qualifier::WRITE, GPU_RGB10_A2, "out_gbuff_color_img")
+    /* Render-passes need to be declared manually to avoid overlap with the G-buffer which reuse
+     * some of binding points. */
+    .image_out(RBUFS_NORMAL_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_normal_img")
+    // .image_array_out(RBUFS_LIGHT_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_light_img")
+    /* TODO(fclem): Merge all render-pass into the same texture array. */
+    // .image_out(RBUFS_DIFF_COLOR_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_diffuse_color_img")
+    .image_out(RBUFS_SPEC_COLOR_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_specular_color_img")
+    .image_out(RBUFS_EMISSION_SLOT, Qualifier::WRITE, GPU_RGBA16F, "rp_emission_img")
     .fragment_source("eevee_surf_deferred_frag.glsl")
     .additional_info("eevee_camera",
                      "eevee_utility_texture",
                      "eevee_sampling_data",
+                     /* Added manually to avoid overlap. */
+                     // "eevee_render_pass_out",
+                     "eevee_cryptomatte_out",
                      "eevee_aov_out");
 
 GPU_SHADER_CREATE_INFO(eevee_surf_forward)
diff --git a/source/blender/draw/intern/draw_texture_pool.cc b/source/blender/draw/intern/draw_texture_pool.cc
index dc155399425..d572cc75b27 100644
--- a/source/blender/draw/intern/draw_texture_pool.cc
+++ b/source/blender/draw/intern/draw_texture_pool.cc
@@ -206,7 +206,8 @@ void DRW_texture_pool_reset(DRWTexturePool *pool)
     }
   }
 
-  BLI_assert(pool->tmp_tex_acquired.is_empty());
+  BLI_assert_msg(pool->tmp_tex_acquired.is_empty(),
+                 "Missing a TextureFromPool.release() before end of draw.");
   for (GPUTexture *tmp_tex : pool->tmp_tex_pruned) {
     GPU_texture_free(tmp_tex);
   }
diff --git a/source/blender/gpu/CMakeLists.txt b/source/blender/gpu/CMakeLists.txt
index bfc0a1b3c7a..329cb211c3f 100644
--- a/source/blender/gpu/CMakeLists.txt
+++ b/source/blender/gpu/CMakeLists.txt
@@ -603,6 +603,7 @@ list(APPEND INC ${CMAKE_CURRENT_BINARY_DIR})
 
 set(SRC_SHADER_CREATE_INFOS
   ../draw/engines/basic/shaders/infos/basic_depth_info.hh
+  ../draw/engines/eevee_next/shaders/infos/eevee_deferred_info.hh
   ../draw/engines/eevee_next/shaders/infos/eevee_depth_of_field_info.hh
   ../draw/engines/eevee_next/shaders/infos/eevee_film_info.hh
   ../draw/engines/eevee_next/shaders/infos/eevee_hiz_info.hh
diff --git a/source/blender/gpu/shaders/common/gpu_shader_common_math_utils.glsl b/source/blender/gpu/shaders/common/gpu_shader_common_math_utils.glsl
index ca9c42f2027..ddb014cc045 100644
--- a/source/blender/gpu/shaders/common/gpu_shader_common_math_utils.glsl
+++ b/source/blender/gpu/shaders/common/gpu_shader_common_math_utils.glsl
@@ -1,18 +1,23 @@
 /* Float Math */
 
+/* WORKAROUND: To be removed once we port all code to use gpu_shader_math_base_lib.glsl. */
+#ifndef GPU_SHADER_MATH_BASE_LIB_GLSL
+
 float safe_divide(float a, float b)
 {
   return (b != 0.0) ? a / b : 0.0;
 }
 
+#endif
+
 /* fmod function compatible with OSL (copy from OSL/dual.h) */
 float compatible_fmod(float a, float b)
 {
-  if (b != 0.0f) {
+  if (b != 0.0) {
     int N = int(a / b);
     return a - N * b;
   }
-  return 0.0f;
+  return 0.0;
 }
 
 float compatible_pow(float x, float y)
@@ -59,11 +64,16 @@ vec3 wrap(vec3 a, vec3 b, vec3 c)
   return vec3(wrap(a.x, b.x, c.x), wrap(a.y, b.y, c.y), wrap(a.z, b.z, c.z));
 }
 
+/* WORKAROUND: To be removed once we port all code to use gpu_shader_math_base_lib.glsl. */
+#ifndef GPU_SHADER_MATH_BASE_LIB_GLSL
+
 float hypot(float x, float y)
 {
   return sqrt(x * x + y * y);
 }
 
+#endif
+
 int floor_to_int(float x)
 {
   return int(floor(x));
@@ -76,6 +86,9 @@ int quick_floor(float x)
 
 /* Vector Math */
 
+/* WORKAROUND: To be removed once we port all code to use gpu_shader_math_base_lib.glsl. */
+#ifndef GPU_SHADER_MATH_BASE_LIB_GLSL
+
 vec2 safe_divide(vec2 a, vec2 b)
 {
   return vec2(safe_divide(a.x, b.x), safe_divide(a.y, b.y));
@@ -107,6 +120,8 @@ vec4 safe_divide(vec4 a, float b)
   return (b != 0.0) ? a / b : vec4(0.0);
 }
 
+#endif
+
 vec3 compatible_fmod(vec3 a, vec3 b)
 {
   return vec3(compatible_fmod(a.x, b.x), compatible_fmod(a.y, b.y), compatible_fmod(a.z, b.z));