From 8af140d58b9d5f65fb7069b72c028602d8b40e73 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?St=C3=A9phane=20du=20Hamel?= <stephduh@live.fr>
Date: Wed, 23 Sep 2026 15:30:15 +0200
Subject: [PATCH 1/3] feat: optimize cfg special cases with guidance schdeule

---
 src/pipeline/diffusion_engine.cpp | 33 ++++++++++++++++++++++++++++---
 1 file changed, 30 insertions(+), 3 deletions(-)

diff --git a/src/pipeline/diffusion_engine.cpp b/src/pipeline/diffusion_engine.cpp
index 022093957..0f9bf78bc 100644
--- a/src/pipeline/diffusion_engine.cpp
+++ b/src/pipeline/diffusion_engine.cpp
@@ -2569,12 +2569,38 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
             }
         }
 
+        float effective_guidance_scale = guidance_schedule.empty() 
+            ? cfg_scale
+            : guidance_schedule[guidance_schedule.size() - 1 - step];
+
+        float image_guidance_scale = img_cfg_scale; 
+
+        constexpr float kEpsilon = 1e-5f;
+
+        bool skip_uncond = false;
+        if (!uncond.empty() && !needs_uncond_denoised) {
+            if (!img_uncond.empty()) {
+                skip_uncond = std::abs(image_guidance_scale - effective_guidance_scale) < kEpsilon;
+            } else {
+                skip_uncond = std::abs(effective_guidance_scale - 1.0f) < kEpsilon;
+            }
+        }
+
+        bool skip_img_uncond = false;
+        if (!img_uncond.empty() && !needs_uncond_denoised) {
+            if (!uncond.empty()) {
+                skip_img_uncond = std::abs(image_guidance_scale - 1.0f) < kEpsilon;
+            } else {
+                skip_img_uncond = std::abs(effective_guidance_scale - 1.0f) < kEpsilon;
+            }
+        }
+
         cond_out = run_condition(*positive_condition, c_concat_override);
         if (cond_out.empty()) {
             return {};
         }
 
-        if (!uncond.empty()) {
+        if (!uncond.empty() && !skip_uncond) {
             if (!step_cache.is_step_skipped()) {
                 compute_sample_controls(control_image,
                                         noised_input,
@@ -2596,7 +2622,8 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
                 return {};
             }
         }
-        if (!img_uncond.empty()) {
+
+        if (!img_uncond.empty() && !skip_img_uncond) {
             img_uncond_out = run_condition(img_uncond,
                                            img_uncond.c_concat.empty() ? nullptr : &img_uncond.c_concat,
                                            nullptr,
@@ -2613,7 +2640,7 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
         guidance_input.pred_uncond     = uncond_out.empty() ? nullptr : &uncond_out;
         guidance_input.pred_img_uncond = img_uncond_out.empty() ? nullptr : &img_uncond_out;
 
-        sd::guidance::GuiderOutput guided = guidance_schedule.empty() ? primary_guidance.forward(guidance_input, {}) : primary_guidance.forward(guidance_input, {}, guidance_schedule[guidance_schedule.size() - 1 - step]);
+        sd::guidance::GuiderOutput guided = primary_guidance.forward(guidance_input, {}, effective_guidance_scale);
         if (guided.pred.empty()) {
             return {};
         }

From 75e1c85560fe9952ec69f13dc7729560e4c3d646 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?St=C3=A9phane=20du=20Hamel?= <stephduh@live.fr>
Date: Wed, 23 Sep 2026 21:04:07 +0200
Subject: [PATCH 2/3] requested changes

---
 src/pipeline/diffusion_engine.cpp | 85 ++++++++++++++++++++-----------
 1 file changed, 55 insertions(+), 30 deletions(-)

diff --git a/src/pipeline/diffusion_engine.cpp b/src/pipeline/diffusion_engine.cpp
index 0f9bf78bc..f8cb9d9fb 100644
--- a/src/pipeline/diffusion_engine.cpp
+++ b/src/pipeline/diffusion_engine.cpp
@@ -2578,7 +2578,7 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
         constexpr float kEpsilon = 1e-5f;
 
         bool skip_uncond = false;
-        if (!uncond.empty() && !needs_uncond_denoised) {
+        if (!uncond.empty() && !needs_uncond_denoised && !use_apg_guidance) {
             if (!img_uncond.empty()) {
                 skip_uncond = std::abs(image_guidance_scale - effective_guidance_scale) < kEpsilon;
             } else {
@@ -2587,7 +2587,7 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
         }
 
         bool skip_img_uncond = false;
-        if (!img_uncond.empty() && !needs_uncond_denoised) {
+        if (!img_uncond.empty() && !needs_uncond_denoised && !use_apg_guidance) {
             if (!uncond.empty()) {
                 skip_img_uncond = std::abs(image_guidance_scale - 1.0f) < kEpsilon;
             } else {
@@ -2600,37 +2600,62 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
             return {};
         }
 
-        if (!uncond.empty() && !skip_uncond) {
-            if (!step_cache.is_step_skipped()) {
-                compute_sample_controls(control_image,
-                                        noised_input,
-                                        timesteps_tensor,
-                                        uncond,
-                                        &controls);
-            }
-            const std::vector<int>* uncond_skip_layers = nullptr;
-            if (is_skiplayer_step && slg_uncond) {
-                LOG_VERBOSE("Skipping layers at uncond step %d\n", step);
-                uncond_skip_layers = &skip_layer_guidance.layers();
-            }
-            uncond_out = run_condition(uncond,
-                                       uncond.c_concat.empty() ? nullptr : &uncond.c_concat,
-                                       uncond_skip_layers,
-                                       nullptr,
-                                       true);
-            if (uncond_out.empty()) {
-                return {};
+        if (!uncond.empty()) {
+            if (!skip_uncond) {
+                if (!step_cache.is_step_skipped()) {
+                    compute_sample_controls(control_image,
+                                            noised_input,
+                                            timesteps_tensor,
+                                            uncond,
+                                            &controls);
+                }
+                const std::vector<int>* uncond_skip_layers = nullptr;
+                if (is_skiplayer_step && slg_uncond) {
+                    LOG_VERBOSE("Skipping layers at uncond step %d\n", step);
+                    uncond_skip_layers = &skip_layer_guidance.layers();
+                }
+                uncond_out = run_condition(uncond,
+                                           uncond.c_concat.empty() ? nullptr : &uncond.c_concat,
+                                           uncond_skip_layers,
+                                           nullptr,
+                                           true);
+                if (uncond_out.empty()) {
+                    return {};
+                }
+            } else {
+                if (step_cache.runtime.easycache_enabled())
+                    step_cache.runtime.easycache.reset_runtime();
+                if (step_cache.runtime.ucache_enabled())
+                    step_cache.runtime.ucache.reset_runtime();
+                if (step_cache.runtime.cachedit_enabled())
+                    step_cache.runtime.cachedit.reset_runtime();
+                if (!img_uncond.empty() && !skip_img_uncond && !step_cache.is_step_skipped()) {
+                    compute_sample_controls(control_image,
+                                            noised_input,
+                                            timesteps_tensor,
+                                            uncond,
+                                            &controls);
+                }
             }
         }
 
-        if (!img_uncond.empty() && !skip_img_uncond) {
-            img_uncond_out = run_condition(img_uncond,
-                                           img_uncond.c_concat.empty() ? nullptr : &img_uncond.c_concat,
-                                           nullptr,
-                                           uncond_without_ref_latents ? &empty_ref_latents : nullptr,
-                                           true);
-            if (img_uncond_out.empty()) {
-                return {};
+        if (!img_uncond.empty()) {
+            if (!skip_img_uncond) {
+                img_uncond_out = run_condition(img_uncond,
+                                               img_uncond.c_concat.empty() ? nullptr : &img_uncond.c_concat,
+                                               nullptr,
+                                               uncond_without_ref_latents ? &empty_ref_latents : nullptr,
+                                               true);
+                if (img_uncond_out.empty()) {
+                    return {};
+                }
+            } else {
+                if (step_cache.runtime.easycache_enabled())
+                    step_cache.runtime.easycache.reset_runtime();
+                if (step_cache.runtime.ucache_enabled())
+                    step_cache.runtime.ucache.reset_runtime();
+                if (step_cache.runtime.cachedit_enabled())
+                    step_cache.runtime.cachedit.reset_runtime();
             }
         }
         sd::guidance::GuidanceInput guidance_input;

From 1fea1b9729d9ec2eecca07accf0458ed8e510d24 Mon Sep 17 00:00:00 2001
From: leejet <leejet714@gmail.com>
Date: Fri, 25 Sep 2026 01:05:10 +0800
Subject: [PATCH 3/3] fix: preserve cache state when skipping CFG conditions

---
 src/pipeline/diffusion_engine.cpp | 48 ++++++++++++-------------------
 src/runtime/sample-cache.cpp      | 20 +++++++++++++
 src/runtime/sample-cache.h        |  1 +
 3 files changed, 39 insertions(+), 30 deletions(-)

diff --git a/src/pipeline/diffusion_engine.cpp b/src/pipeline/diffusion_engine.cpp
index f8cb9d9fb..7d380fa5d 100644
--- a/src/pipeline/diffusion_engine.cpp
+++ b/src/pipeline/diffusion_engine.cpp
@@ -2441,6 +2441,7 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
                                 timesteps_tensor,
                                 cond,
                                 &controls);
+        bool uncond_controls_ready = false;
 
         static const std::vector<sd::Tensor<float>> empty_ref_latents;
         bool uncond_without_ref_latents = !img_uncond.empty() &&
@@ -2530,6 +2531,17 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
                 return std::move(cached_output);
             }
 
+            // A re-enabled condition can miss the cache even when the positive pass was reused.
+            if (!uncond_controls_ready && !uncond.empty() &&
+                (&condition == &uncond || &condition == &img_uncond)) {
+                compute_sample_controls(control_image,
+                                        noised_input,
+                                        timesteps_tensor,
+                                        uncond,
+                                        &controls);
+                uncond_controls_ready = true;
+            }
+
             for (const auto& extension : generation_extensions) {
                 extension->before_diffusion(diffusion_params, step);
             }
@@ -2569,11 +2581,11 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
             }
         }
 
-        float effective_guidance_scale = guidance_schedule.empty() 
-            ? cfg_scale
-            : guidance_schedule[guidance_schedule.size() - 1 - step];
+        float effective_guidance_scale = guidance_schedule.empty()
+                                             ? cfg_scale
+                                             : guidance_schedule[guidance_schedule.size() - 1 - step];
 
-        float image_guidance_scale = img_cfg_scale; 
+        float image_guidance_scale = img_cfg_scale;
 
         constexpr float kEpsilon = 1e-5f;
 
@@ -2602,13 +2614,6 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
 
         if (!uncond.empty()) {
             if (!skip_uncond) {
-                if (!step_cache.is_step_skipped()) {
-                    compute_sample_controls(control_image,
-                                            noised_input,
-                                            timesteps_tensor,
-                                            uncond,
-                                            &controls);
-                }
                 const std::vector<int>* uncond_skip_layers = nullptr;
                 if (is_skiplayer_step && slg_uncond) {
                     LOG_VERBOSE("Skipping layers at uncond step %d\n", step);
@@ -2623,19 +2628,7 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
                     return {};
                 }
             } else {
-                if (step_cache.runtime.easycache_enabled())
-                    step_cache.runtime.easycache.reset_runtime();
-                if (step_cache.runtime.ucache_enabled())
-                    step_cache.runtime.ucache.reset_runtime();
-                if (step_cache.runtime.cachedit_enabled())
-                    step_cache.runtime.cachedit.reset_runtime();
-                if (!img_uncond.empty() && !skip_img_uncond && !step_cache.is_step_skipped()) {
-                    compute_sample_controls(control_image,
-                                            noised_input,
-                                            timesteps_tensor,
-                                            uncond,
-                                            &controls);
-                }
+                step_cache.invalidate_condition(&uncond);
             }
         }
 
@@ -2650,12 +2643,7 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
                     return {};
                 }
             } else {
-                if (step_cache.runtime.easycache_enabled())
-                    step_cache.runtime.easycache.reset_runtime();
-                if (step_cache.runtime.ucache_enabled())
-                    step_cache.runtime.ucache.reset_runtime();
-                if (step_cache.runtime.cachedit_enabled())
-                    step_cache.runtime.cachedit.reset_runtime();
+                step_cache.invalidate_condition(&img_uncond);
             }
         }
         sd::guidance::GuidanceInput guidance_input;
diff --git a/src/runtime/sample-cache.cpp b/src/runtime/sample-cache.cpp
index bfa3d4487..0b92c7335 100644
--- a/src/runtime/sample-cache.cpp
+++ b/src/runtime/sample-cache.cpp
@@ -275,6 +275,26 @@ namespace sd_sample {
         }
     }
 
+    void SampleStepCacheDispatcher::invalidate_condition(const void* condition) {
+        if (condition == nullptr) {
+            return;
+        }
+
+        switch (runtime.mode) {
+            case SampleCacheMode::EASYCACHE:
+                runtime.easycache.cache_diffs.erase(condition);
+                break;
+            case SampleCacheMode::UCACHE:
+                runtime.ucache.cache_diffs.erase(condition);
+                break;
+            case SampleCacheMode::CACHEDIT:
+                runtime.cachedit.cache_diffs.erase(condition);
+                break;
+            case SampleCacheMode::NONE:
+                break;
+        }
+    }
+
     bool SampleStepCacheDispatcher::is_step_skipped() const {
         switch (runtime.mode) {
             case SampleCacheMode::EASYCACHE:
diff --git a/src/runtime/sample-cache.h b/src/runtime/sample-cache.h
index cb0fe7bc5..30ff66e5f 100644
--- a/src/runtime/sample-cache.h
+++ b/src/runtime/sample-cache.h
@@ -46,6 +46,7 @@ namespace sd_sample {
 
         bool before_condition(const void* condition, const sd::Tensor<float>& input, sd::Tensor<float>* output);
         void after_condition(const void* condition, const sd::Tensor<float>& input, const sd::Tensor<float>& output);
+        void invalidate_condition(const void* condition);
         bool is_step_skipped() const;
     };