feat: add SD2.x support (#40)

stable-diffusion.cpp

@@ -48,6 +48,16 @@ static SDLogLevel log_level = SDLogLevel::INFO;
 
 #define TIMESTEPS 1000
 
+enum ModelType {
+    SD1 = 0,
+    SD2 = 1,
+    MODEL_TYPE_COUNT,
+};
+
+const char* model_type_to_str[] = {
+    "SD1.x",
+    "SD2.x"};
+
 /*================================================== Helper Functions ================================================*/
 
 void set_sd_log_level(SDLogLevel level) {
@@ -257,8 +267,8 @@ void image_vec_to_ggml(const std::vector<uint8_t>& vec,
     }
 }
 
-struct ggml_tensor * ggml_group_norm_32(struct ggml_context * ctx,
-                                        struct ggml_tensor * a) {
+struct ggml_tensor* ggml_group_norm_32(struct ggml_context* ctx,
+                                       struct ggml_tensor* a) {
     return ggml_group_norm(ctx, a, 32);
 }
 
@@ -278,6 +288,7 @@ const int PAD_TOKEN_ID = 49407;
 // TODO: implement bpe
 class CLIPTokenizer {
    private:
+    ModelType model_type = SD1;
     std::map<std::string, int32_t> encoder;
     std::regex pat;
 
@@ -300,7 +311,8 @@ class CLIPTokenizer {
     }
 
    public:
-    CLIPTokenizer() = default;
+    CLIPTokenizer(ModelType model_type = SD1)
+        : model_type(model_type){};
     std::string bpe(std::string token) {
         std::string word = token + "</w>";
         if (encoder.find(word) != encoder.end()) {
@@ -321,13 +333,18 @@ class CLIPTokenizer {
         if (max_length > 0) {
             if (tokens.size() > max_length - 1) {
                 tokens.resize(max_length - 1);
+                tokens.push_back(EOS_TOKEN_ID);
             } else {
+                tokens.push_back(EOS_TOKEN_ID);
                 if (padding) {
-                    tokens.insert(tokens.end(), max_length - 1 - tokens.size(), PAD_TOKEN_ID);
+                    int pad_token_id = PAD_TOKEN_ID;
+                    if (model_type == SD2) {
+                        pad_token_id = 0;
+                    }
+                    tokens.insert(tokens.end(), max_length - tokens.size(), pad_token_id);
                 }
             }
         }
-        tokens.push_back(EOS_TOKEN_ID);
         return tokens;
     }
 
@@ -635,7 +652,11 @@ struct ResidualAttentionBlock {
         x = ggml_mul_mat(ctx, fc1_w, x);
         x = ggml_add(ctx, ggml_repeat(ctx, fc1_b, x), x);
 
-        x = ggml_gelu_quick_inplace(ctx, x);
+        if (hidden_size == 1024) {  // SD 2.x
+            x = ggml_gelu_inplace(ctx, x);
+        } else {  // SD 1.x
+            x = ggml_gelu_quick_inplace(ctx, x);
+        }
 
         x = ggml_mul_mat(ctx, fc2_w, x);
         x = ggml_add(ctx, ggml_repeat(ctx, fc2_b, x), x);
@@ -647,26 +668,40 @@ struct ResidualAttentionBlock {
     }
 };
 
+// SD1.x: https://huggingface.co/openai/clip-vit-large-patch14/blob/main/config.json
+// SD2.x: https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/blob/main/config.json
 struct CLIPTextModel {
+    ModelType model_type = SD1;
     // network hparams
     int32_t vocab_size = 49408;
     int32_t max_position_embeddings = 77;
-    int32_t hidden_size = 768;
-    int32_t intermediate_size = 3072;
-    int32_t projection_dim = 768;
-    int32_t n_head = 12;  // num_attention_heads
-    int32_t num_hidden_layers = 12;
+    int32_t hidden_size = 768;         // 1024 for SD 2.x
+    int32_t intermediate_size = 3072;  // 4096 for SD 2.x
+    int32_t n_head = 12;               // num_attention_heads, 16 for SD 2.x
+    int32_t num_hidden_layers = 12;    // 24 for SD 2.x
 
     // embeddings
     struct ggml_tensor* position_ids;
     struct ggml_tensor* token_embed_weight;
     struct ggml_tensor* position_embed_weight;
     // transformer
-    ResidualAttentionBlock resblocks[12];
+    std::vector<ResidualAttentionBlock> resblocks;
     struct ggml_tensor* final_ln_w;
     struct ggml_tensor* final_ln_b;
 
-    CLIPTextModel() {
+    CLIPTextModel(ModelType model_type = SD1)
+        : model_type(model_type) {
+        if (model_type == SD2) {
+            hidden_size = 1024;
+            intermediate_size = 4096;
+            n_head = 16;
+            num_hidden_layers = 24;
+        }
+        resblocks.resize(num_hidden_layers);
+        set_resblocks_hp_params();
+    }
+
+    void set_resblocks_hp_params() {
         int d_model = hidden_size / n_head;  // 64
         for (int i = 0; i < num_hidden_layers; i++) {
             resblocks[i].d_model = d_model;
@@ -729,6 +764,9 @@ struct CLIPTextModel {
 
         // transformer
         for (int i = 0; i < num_hidden_layers; i++) {
+            if (model_type == SD2 && i == num_hidden_layers - 1) {  // layer: "penultimate"
+                break;
+            }
             x = resblocks[i].forward(ctx, x);  // [N, n_token, hidden_size]
         }
 
@@ -759,9 +797,13 @@ struct FrozenCLIPEmbedder {
 
 // Ref: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/cad87bf4e3e0b0a759afa94e933527c3123d59bc/modules/sd_hijack_clip.py#L283
 struct FrozenCLIPEmbedderWithCustomWords {
+    ModelType model_type = SD1;
     CLIPTokenizer tokenizer;
     CLIPTextModel text_model;
 
+    FrozenCLIPEmbedderWithCustomWords(ModelType model_type = SD1)
+        : model_type(model_type), tokenizer(model_type), text_model(model_type) {}
+
     std::pair<std::vector<int>, std::vector<float>> tokenize(std::string text,
                                                              size_t max_length = 0,
                                                              bool padding = false) {
@@ -793,15 +835,21 @@ struct FrozenCLIPEmbedderWithCustomWords {
             if (tokens.size() > max_length - 1) {
                 tokens.resize(max_length - 1);
                 weights.resize(max_length - 1);
+                tokens.push_back(EOS_TOKEN_ID);
+                weights.push_back(1.0);
             } else {
+                tokens.push_back(EOS_TOKEN_ID);
+                weights.push_back(1.0);
                 if (padding) {
-                    tokens.insert(tokens.end(), max_length - 1 - tokens.size(), PAD_TOKEN_ID);
-                    weights.insert(weights.end(), max_length - 1 - weights.size(), 1.0);
+                    int pad_token_id = PAD_TOKEN_ID;
+                    if (model_type == SD2) {
+                        pad_token_id = 0;
+                    }
+                    tokens.insert(tokens.end(), max_length - tokens.size(), pad_token_id);
+                    weights.insert(weights.end(), max_length - weights.size(), 1.0);
                 }
             }
         }
-        tokens.push_back(EOS_TOKEN_ID);
-        weights.push_back(1.0);
 
         // for (int i = 0; i < tokens.size(); i++) {
         //     std::cout << tokens[i] << ":" << weights[i] << ", ";
@@ -974,7 +1022,7 @@ struct SpatialTransformer {
     int n_head;             // num_heads
     int d_head;             // in_channels // n_heads
     int depth = 1;          // 1
-    int context_dim = 768;  // hidden_size
+    int context_dim = 768;  // hidden_size, 1024 for SD2.x
 
     // group norm
     struct ggml_tensor* norm_w;  // [in_channels,]
@@ -1459,6 +1507,7 @@ struct UNetModel {
     int time_embed_dim = 1280;  // model_channels*4
     int num_heads = 8;
     int num_head_channels = -1;  // channels // num_heads
+    int context_dim = 768;       // 1024 for SD2.x
 
     // network params
     struct ggml_tensor* time_embed_0_w;  // [time_embed_dim, model_channels]
@@ -1493,7 +1542,12 @@ struct UNetModel {
     struct ggml_tensor* out_2_w;  // [out_channels, model_channels, 3, 3]
     struct ggml_tensor* out_2_b;  // [out_channels, ]
 
-    UNetModel() {
+    UNetModel(ModelType model_type = SD1) {
+        if (model_type == SD2) {
+            context_dim = 1024;
+            num_head_channels = 64;
+            num_heads = -1;
+        }
         // set up hparams of blocks
 
         // input_blocks
@@ -1513,9 +1567,16 @@ struct UNetModel {
                 ch = mult * model_channels;
 
                 if (ds == attention_resolutions[0] || ds == attention_resolutions[1] || ds == attention_resolutions[2]) {
+                    int n_head = num_heads;
+                    int d_head = ch / num_heads;
+                    if (num_head_channels != -1) {
+                        d_head = num_head_channels;
+                        n_head = ch / d_head;
+                    }
                     input_transformers[i][j].in_channels = ch;
-                    input_transformers[i][j].n_head = num_heads;
-                    input_transformers[i][j].d_head = ch / num_heads;
+                    input_transformers[i][j].n_head = n_head;
+                    input_transformers[i][j].d_head = d_head;
+                    input_transformers[i][j].context_dim = context_dim;
                 }
                 input_block_chans.push_back(ch);
             }
@@ -1533,9 +1594,16 @@ struct UNetModel {
         middle_block_0.emb_channels = time_embed_dim;
         middle_block_0.out_channels = ch;
 
+        int n_head = num_heads;
+        int d_head = ch / num_heads;
+        if (num_head_channels != -1) {
+            d_head = num_head_channels;
+            n_head = ch / d_head;
+        }
         middle_block_1.in_channels = ch;
-        middle_block_1.n_head = num_heads;
-        middle_block_1.d_head = ch / num_heads;
+        middle_block_1.n_head = n_head;
+        middle_block_1.d_head = d_head;
+        middle_block_1.context_dim = context_dim;
 
         middle_block_2.channels = ch;
         middle_block_2.emb_channels = time_embed_dim;
@@ -1555,9 +1623,16 @@ struct UNetModel {
                 ch = mult * model_channels;
 
                 if (ds == attention_resolutions[0] || ds == attention_resolutions[1] || ds == attention_resolutions[2]) {
+                    int n_head = num_heads;
+                    int d_head = ch / num_heads;
+                    if (num_head_channels != -1) {
+                        d_head = num_head_channels;
+                        n_head = ch / d_head;
+                    }
                     output_transformers[i][j].in_channels = ch;
-                    output_transformers[i][j].n_head = num_heads;
-                    output_transformers[i][j].d_head = ch / num_heads;
+                    output_transformers[i][j].n_head = n_head;
+                    output_transformers[i][j].d_head = d_head;
+                    output_transformers[i][j].context_dim = context_dim;
                 }
 
                 if (i > 0 && j == num_res_blocks) {
@@ -2584,7 +2659,8 @@ struct AutoEncoderKL {
 /*================================================= CompVisDenoiser ==================================================*/
 
 // Ref: https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/external.py
-struct CompVisDenoiser {
+
+struct DiscreteSchedule {
     float alphas_cumprod[TIMESTEPS];
     float sigmas[TIMESTEPS];
     float log_sigmas[TIMESTEPS];
@@ -2602,12 +2678,6 @@ struct CompVisDenoiser {
         return result;
     }
 
-    std::pair<float, float> get_scalings(float sigma) {
-        float c_out = -sigma;
-        float c_in = 1.0f / std::sqrt(sigma * sigma + 1);
-        return std::pair<float, float>(c_in, c_out);
-    }
-
     float sigma_to_t(float sigma) {
         float log_sigma = std::log(sigma);
         std::vector<float> dists;
@@ -2641,6 +2711,29 @@ struct CompVisDenoiser {
         float log_sigma = (1.0f - w) * log_sigmas[low_idx] + w * log_sigmas[high_idx];
         return std::exp(log_sigma);
     }
+
+    virtual std::vector<float> get_scalings(float sigma) = 0;
+};
+
+struct CompVisDenoiser : public DiscreteSchedule {
+    float sigma_data = 1.0f;
+
+    std::vector<float> get_scalings(float sigma) {
+        float c_out = -sigma;
+        float c_in = 1.0f / std::sqrt(sigma * sigma + sigma_data * sigma_data);
+        return {c_out, c_in};
+    }
+};
+
+struct CompVisVDenoiser : public DiscreteSchedule {
+    float sigma_data = 1.0f;
+
+    std::vector<float> get_scalings(float sigma) {
+        float c_skip = sigma_data * sigma_data / (sigma * sigma + sigma_data * sigma_data);
+        float c_out = -sigma * sigma_data / std::sqrt(sigma * sigma + sigma_data * sigma_data);
+        float c_in = 1.0f / std::sqrt(sigma * sigma + sigma_data * sigma_data);
+        return {c_skip, c_out, c_in};
+    }
 };
 
 /*=============================================== StableDiffusionGGML ================================================*/
@@ -2666,7 +2759,7 @@ class StableDiffusionGGML {
     UNetModel diffusion_model;
     AutoEncoderKL first_stage_model;
 
-    CompVisDenoiser denoiser;
+    std::shared_ptr<DiscreteSchedule> denoiser = std::make_shared<CompVisDenoiser>();
 
     StableDiffusionGGML() = default;
 
@@ -2717,9 +2810,20 @@ class StableDiffusionGGML {
         LOG_DEBUG("loading hparams");
         // load hparams
         file.read(reinterpret_cast<char*>(&ftype), sizeof(ftype));
-        // for the big tensors, we have the option to store the data in 16-bit floats or quantized
-        // in order to save memory and also to speed up the computation
-        ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype)(ftype));
+
+        int model_type = (ftype >> 16) & 0xFFFF;
+        if (model_type >= MODEL_TYPE_COUNT) {
+            LOG_ERROR("invalid model file '%s' (bad model type value %d)", file_path.c_str(), ftype);
+            return false;
+        }
+        LOG_INFO("model type: %s", model_type_to_str[model_type]);
+
+        if (model_type == SD2) {
+            cond_stage_model = FrozenCLIPEmbedderWithCustomWords((ModelType)model_type);
+            diffusion_model = UNetModel((ModelType)model_type);
+        }
+
+        ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype)(ftype & 0xFFFF));
         LOG_INFO("ftype: %s", ggml_type_name(wtype));
         if (wtype == GGML_TYPE_COUNT) {
             LOG_ERROR("invalid model file '%s' (bad ftype value %d)", file_path.c_str(), ftype);
@@ -2840,6 +2944,7 @@ class StableDiffusionGGML {
         std::set<std::string> tensor_names_in_file;
         int64_t t0 = ggml_time_ms();
         // load weights
+        float alphas_cumprod[TIMESTEPS];
         {
             int n_tensors = 0;
             size_t total_size = 0;
@@ -2872,12 +2977,7 @@ class StableDiffusionGGML {
                 tensor_names_in_file.insert(std::string(name.data()));
 
                 if (std::string(name.data()) == "alphas_cumprod") {
-                    file.read(reinterpret_cast<char*>(denoiser.alphas_cumprod),
-                              nelements * ggml_type_size((ggml_type)ttype));
-                    for (int i = 0; i < 1000; i++) {
-                        denoiser.sigmas[i] = std::sqrt((1 - denoiser.alphas_cumprod[i]) / denoiser.alphas_cumprod[i]);
-                        denoiser.log_sigmas[i] = std::log(denoiser.sigmas[i]);
-                    }
+                    file.read(reinterpret_cast<char*>(alphas_cumprod), nelements * ggml_type_size((ggml_type)ttype));
                     continue;
                 }
 
@@ -2953,9 +3053,143 @@ class StableDiffusionGGML {
         int64_t t1 = ggml_time_ms();
         LOG_INFO("loading model from '%s' completed, taking %.2fs", file_path.c_str(), (t1 - t0) * 1.0f / 1000);
         file.close();
+
+        // check is_using_v_parameterization_for_sd2
+        bool is_using_v_parameterization = false;
+        if (model_type == SD2) {
+            struct ggml_init_params params;
+            params.mem_size = static_cast<size_t>(10 * 1024) * 1024;  // 10M
+            params.mem_buffer = NULL;
+            params.no_alloc = false;
+            params.dynamic = false;
+            struct ggml_context* ctx = ggml_init(params);
+            if (!ctx) {
+                LOG_ERROR("ggml_init() failed");
+                return false;
+            }
+            if (is_using_v_parameterization_for_sd2(ctx)) {
+                is_using_v_parameterization = true;
+            }
+        }
+
+        if (is_using_v_parameterization) {
+            denoiser = std::make_shared<CompVisVDenoiser>();
+            LOG_INFO("running in v-prediction mode");
+        } else {
+            LOG_INFO("running in eps-prediction mode");
+        }
+
+        for (int i = 0; i < TIMESTEPS; i++) {
+            denoiser->alphas_cumprod[i] = alphas_cumprod[i];
+            denoiser->sigmas[i] = std::sqrt((1 - denoiser->alphas_cumprod[i]) / denoiser->alphas_cumprod[i]);
+            denoiser->log_sigmas[i] = std::log(denoiser->sigmas[i]);
+        }
+
         return true;
     }
 
+    bool is_using_v_parameterization_for_sd2(ggml_context* res_ctx) {
+        struct ggml_tensor* x_t = ggml_new_tensor_4d(res_ctx, GGML_TYPE_F32, 8, 8, 4, 1);
+        ggml_set_f32(x_t, 0.5);
+        struct ggml_tensor* c = ggml_new_tensor_4d(res_ctx, GGML_TYPE_F32, 1024, 2, 1, 1);
+        ggml_set_f32(c, 0.5);
+
+        size_t ctx_size = 1 * 1024 * 1024;  // 1MB
+        // calculate the amount of memory required
+        {
+            struct ggml_init_params params;
+            params.mem_size = ctx_size;
+            params.mem_buffer = NULL;
+            params.no_alloc = true;
+            params.dynamic = dynamic;
+
+            struct ggml_context* ctx = ggml_init(params);
+            if (!ctx) {
+                LOG_ERROR("ggml_init() failed");
+                return false;
+            }
+
+            ggml_set_dynamic(ctx, false);
+            struct ggml_tensor* timesteps = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);                           // [N, ]
+            struct ggml_tensor* t_emb = new_timestep_embedding(ctx, timesteps, diffusion_model.model_channels);  // [N, model_channels]
+            ggml_set_dynamic(ctx, params.dynamic);
+
+            struct ggml_tensor* out = diffusion_model.forward(ctx, x_t, NULL, c, t_emb);
+            ctx_size += ggml_used_mem(ctx) + ggml_used_mem_of_data(ctx);
+
+            struct ggml_cgraph diffusion_graph = ggml_build_forward(out);
+            struct ggml_cplan cplan = ggml_graph_plan(&diffusion_graph, n_threads);
+
+            ctx_size += cplan.work_size;
+            LOG_DEBUG("diffusion context need %.2fMB static memory, with work_size needing %.2fMB",
+                      ctx_size * 1.0f / 1024 / 1024,
+                      cplan.work_size * 1.0f / 1024 / 1024);
+
+            ggml_free(ctx);
+        }
+
+        struct ggml_init_params params;
+        params.mem_size = ctx_size;
+        params.mem_buffer = NULL;
+        params.no_alloc = false;
+        params.dynamic = dynamic;
+
+        struct ggml_context* ctx = ggml_init(params);
+        if (!ctx) {
+            LOG_ERROR("ggml_init() failed");
+            return false;
+        }
+
+        ggml_set_dynamic(ctx, false);
+        struct ggml_tensor* timesteps = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);                           // [N, ]
+        struct ggml_tensor* t_emb = new_timestep_embedding(ctx, timesteps, diffusion_model.model_channels);  // [N, model_channels]
+        ggml_set_dynamic(ctx, params.dynamic);
+        ggml_set_f32(timesteps, 999);
+        set_timestep_embedding(timesteps, t_emb, diffusion_model.model_channels);
+
+        struct ggml_tensor* out = diffusion_model.forward(ctx, x_t, NULL, c, t_emb);
+        ggml_hold_dynamic_tensor(out);
+
+        struct ggml_cgraph diffusion_graph = ggml_build_forward(out);
+        struct ggml_cplan cplan = ggml_graph_plan(&diffusion_graph, n_threads);
+
+        ggml_set_dynamic(ctx, false);
+        struct ggml_tensor* buf = ggml_new_tensor_1d(ctx, GGML_TYPE_I8, cplan.work_size);
+        ggml_set_dynamic(ctx, params.dynamic);
+
+        cplan.work_data = (uint8_t*)buf->data;
+
+        int64_t t0 = ggml_time_ms();
+        ggml_graph_compute(&diffusion_graph, &cplan);
+
+        double result = 0.f;
+
+        {
+            float* vec_x = (float*)x_t->data;
+            float* vec_out = (float*)out->data;
+
+            int64_t n = ggml_nelements(out);
+
+            for (int i = 0; i < n; i++) {
+                result += ((double)vec_out[i] - (double)vec_x[i]);
+            }
+            result /= n;
+        }
+
+#ifdef GGML_PERF
+        ggml_graph_print(&diffusion_graph);
+#endif
+        int64_t t1 = ggml_time_ms();
+        LOG_INFO("check is_using_v_parameterization_for_sd2 completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+        LOG_DEBUG("diffusion graph use %.2fMB runtime memory: static %.2fMB, dynamic %.2fMB",
+                  (ctx_size + ggml_curr_max_dynamic_size()) * 1.0f / 1024 / 1024,
+                  ctx_size * 1.0f / 1024 / 1024,
+                  ggml_curr_max_dynamic_size() * 1.0f / 1024 / 1024);
+        LOG_DEBUG("%zu bytes of dynamic memory has not been released yet", ggml_dynamic_size());
+
+        return result < -1;
+    }
+
     ggml_tensor* get_learned_condition(ggml_context* res_ctx, const std::string& text) {
         auto tokens_and_weights = cond_stage_model.tokenize(text,
                                                             cond_stage_model.text_model.max_position_embeddings,
@@ -3093,8 +3327,8 @@ class StableDiffusionGGML {
         size_t steps = sigmas.size() - 1;
         // x_t = load_tensor_from_file(res_ctx, "./rand0.bin");
         // print_ggml_tensor(x_t);
-        struct ggml_tensor* x_out = ggml_dup_tensor(res_ctx, x_t);
-        copy_ggml_tensor(x_out, x_t);
+        struct ggml_tensor* x = ggml_dup_tensor(res_ctx, x_t);
+        copy_ggml_tensor(x, x_t);
 
         size_t ctx_size = 1 * 1024 * 1024;  // 1MB
         // calculate the amount of memory required
@@ -3112,16 +3346,16 @@ class StableDiffusionGGML {
             }
 
             ggml_set_dynamic(ctx, false);
-            struct ggml_tensor* x = ggml_dup_tensor(ctx, x_t);
+            struct ggml_tensor* noised_input = ggml_dup_tensor(ctx, x_t);
             struct ggml_tensor* context = ggml_dup_tensor(ctx, c);
             struct ggml_tensor* timesteps = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);                           // [N, ]
             struct ggml_tensor* t_emb = new_timestep_embedding(ctx, timesteps, diffusion_model.model_channels);  // [N, model_channels]
             ggml_set_dynamic(ctx, params.dynamic);
 
-            struct ggml_tensor* eps = diffusion_model.forward(ctx, x, NULL, context, t_emb);
+            struct ggml_tensor* out = diffusion_model.forward(ctx, noised_input, NULL, context, t_emb);
             ctx_size += ggml_used_mem(ctx) + ggml_used_mem_of_data(ctx);
 
-            struct ggml_cgraph diffusion_graph = ggml_build_forward(eps);
+            struct ggml_cgraph diffusion_graph = ggml_build_forward(out);
             struct ggml_cplan cplan = ggml_graph_plan(&diffusion_graph, n_threads);
 
             ctx_size += cplan.work_size;
@@ -3145,16 +3379,16 @@ class StableDiffusionGGML {
         }
 
         ggml_set_dynamic(ctx, false);
-        struct ggml_tensor* x = ggml_dup_tensor(ctx, x_t);
+        struct ggml_tensor* noised_input = ggml_dup_tensor(ctx, x_t);
         struct ggml_tensor* context = ggml_dup_tensor(ctx, c);
         struct ggml_tensor* timesteps = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);                           // [N, ]
         struct ggml_tensor* t_emb = new_timestep_embedding(ctx, timesteps, diffusion_model.model_channels);  // [N, model_channels]
         ggml_set_dynamic(ctx, params.dynamic);
 
-        struct ggml_tensor* eps = diffusion_model.forward(ctx, x, NULL, context, t_emb);
-        ggml_hold_dynamic_tensor(eps);
+        struct ggml_tensor* out = diffusion_model.forward(ctx, noised_input, NULL, context, t_emb);
+        ggml_hold_dynamic_tensor(out);
 
-        struct ggml_cgraph diffusion_graph = ggml_build_forward(eps);
+        struct ggml_cgraph diffusion_graph = ggml_build_forward(out);
         struct ggml_cplan cplan = ggml_graph_plan(&diffusion_graph, n_threads);
 
         ggml_set_dynamic(ctx, false);
@@ -3163,80 +3397,129 @@ class StableDiffusionGGML {
 
         cplan.work_data = (uint8_t*)buf->data;
 
+        // x = x * sigmas[0]
+        {
+            float* vec = (float*)x->data;
+            for (int i = 0; i < ggml_nelements(x); i++) {
+                vec[i] = vec[i] * sigmas[0];
+            }
+        }
+
+        // denoise wrapper
+        ggml_set_dynamic(ctx, false);
+        struct ggml_tensor* out_cond = NULL;
+        struct ggml_tensor* out_uncond = NULL;
+        if (cfg_scale != 1.0f && uc != NULL) {
+            out_uncond = ggml_dup_tensor(ctx, x);
+        }
+        struct ggml_tensor* denoised = ggml_dup_tensor(ctx, x);
+        ggml_set_dynamic(ctx, params.dynamic);
+
+        auto denoise = [&](ggml_tensor* input, float sigma, int step) {
+            int64_t t0 = ggml_time_ms();
+
+            float c_skip = 1.0f;
+            float c_out = 1.0f;
+            float c_in = 1.0f;
+            std::vector<float> scaling = denoiser->get_scalings(sigma);
+            if (scaling.size() == 3) {  // CompVisVDenoiser
+                c_skip = scaling[0];
+                c_out = scaling[1];
+                c_in = scaling[2];
+            } else {  // CompVisDenoiser
+                c_out = scaling[0];
+                c_in = scaling[1];
+            }
+
+            float t = denoiser->sigma_to_t(sigma);
+            ggml_set_f32(timesteps, t);
+            set_timestep_embedding(timesteps, t_emb, diffusion_model.model_channels);
+
+            copy_ggml_tensor(noised_input, input);
+            // noised_input = noised_input * c_in
+            {
+                float* vec = (float*)noised_input->data;
+                for (int i = 0; i < ggml_nelements(noised_input); i++) {
+                    vec[i] = vec[i] * c_in;
+                }
+            }
+
+            if (cfg_scale != 1.0 && uc != NULL) {
+                // uncond
+                copy_ggml_tensor(context, uc);
+                ggml_graph_compute(&diffusion_graph, &cplan);
+                copy_ggml_tensor(out_uncond, out);
+
+                // cond
+                copy_ggml_tensor(context, c);
+                ggml_graph_compute(&diffusion_graph, &cplan);
+
+                out_cond = out;
+
+                // out_uncond + cfg_scale * (out_cond - out_uncond)
+                {
+                    float* vec_out = (float*)out->data;
+                    float* vec_out_uncond = (float*)out_uncond->data;
+                    float* vec_out_cond = (float*)out_cond->data;
+
+                    for (int i = 0; i < ggml_nelements(out); i++) {
+                        vec_out[i] = vec_out_uncond[i] + cfg_scale * (vec_out_cond[i] - vec_out_uncond[i]);
+                    }
+                }
+            } else {
+                // cond
+                copy_ggml_tensor(context, c);
+                ggml_graph_compute(&diffusion_graph, &cplan);
+            }
+
+            // v = out, eps = out
+            // denoised = (v * c_out + input * c_skip) or (input + eps * c_out)
+            {
+                float* vec_denoised = (float*)denoised->data;
+                float* vec_input = (float*)input->data;
+                float* vec_out = (float*)out->data;
+
+                for (int i = 0; i < ggml_nelements(denoised); i++) {
+                    vec_denoised[i] = vec_out[i] * c_out + vec_input[i] * c_skip;
+                }
+            }
+
+#ifdef GGML_PERF
+            ggml_graph_print(&diffusion_graph);
+#endif
+            int64_t t1 = ggml_time_ms();
+            LOG_INFO("step %d sampling completed, taking %.2fs", step, (t1 - t0) * 1.0f / 1000);
+            LOG_DEBUG("diffusion graph use %.2fMB runtime memory: static %.2fMB, dynamic %.2fMB",
+                      (ctx_size + ggml_curr_max_dynamic_size()) * 1.0f / 1024 / 1024,
+                      ctx_size * 1.0f / 1024 / 1024,
+                      ggml_curr_max_dynamic_size() * 1.0f / 1024 / 1024);
+            LOG_DEBUG("%zu bytes of dynamic memory has not been released yet", ggml_dynamic_size());
+        };
+
         // sample_euler_ancestral
         {
             ggml_set_dynamic(ctx, false);
-            struct ggml_tensor* eps_cond = NULL;
-            struct ggml_tensor* eps_uncond = NULL;
-            struct ggml_tensor* noise = ggml_dup_tensor(ctx, x_out);
-            if (cfg_scale != 1.0f && uc != NULL) {
-                eps_uncond = ggml_dup_tensor(ctx, x_out);
-            }
-            struct ggml_tensor* d = ggml_dup_tensor(ctx, x_out);
+            struct ggml_tensor* noise = ggml_dup_tensor(ctx, x);
+            struct ggml_tensor* d = ggml_dup_tensor(ctx, x);
             ggml_set_dynamic(ctx, params.dynamic);
 
-            // x_out = x_out * sigmas[0]
-            {
-                float* vec = (float*)x_out->data;
-                for (int i = 0; i < ggml_nelements(x_out); i++) {
-                    vec[i] = vec[i] * sigmas[0];
-                }
-            }
-
             for (int i = 0; i < steps; i++) {
-                int64_t t0 = ggml_time_ms();
+                float sigma = sigmas[i];
 
-                copy_ggml_tensor(x, x_out);
+                // denoise
+                denoise(x, sigma, i + 1);
 
-                std::pair<float, float> scaling = denoiser.get_scalings(sigmas[i]);
-                float c_in = scaling.first;
-                float c_out = scaling.second;
-                float t = denoiser.sigma_to_t(sigmas[i]);
-                ggml_set_f32(timesteps, t);
-                set_timestep_embedding(timesteps, t_emb, diffusion_model.model_channels);
-
-                // x = x * c_in
+                // d = (x - denoised) / sigma
                 {
-                    float* vec = (float*)x->data;
-                    for (int i = 0; i < ggml_nelements(x); i++) {
-                        vec[i] = vec[i] * c_in;
+                    float* vec_d = (float*)d->data;
+                    float* vec_x = (float*)x->data;
+                    float* vec_denoised = (float*)denoised->data;
+
+                    for (int i = 0; i < ggml_nelements(d); i++) {
+                        vec_d[i] = (vec_x[i] - vec_denoised[i]) / sigma;
                     }
                 }
 
-                /*d = (x - denoised) / sigma
-                        = (-eps_uncond * c_out - cfg_scale * (eps_cond * c_out - eps_uncond * c_out)) / sigma
-                        = eps_uncond + cfg_scale * (eps_cond - eps_uncond)*/
-                if (cfg_scale != 1.0 && uc != NULL) {
-                    // uncond
-                    copy_ggml_tensor(context, uc);
-                    ggml_graph_compute(&diffusion_graph, &cplan);
-                    copy_ggml_tensor(eps_uncond, eps);
-
-                    // cond
-                    copy_ggml_tensor(context, c);
-                    ggml_graph_compute(&diffusion_graph, &cplan);
-
-                    eps_cond = eps;
-
-                    /*d = (x - denoised) / sigma
-                        = (-eps_uncond * c_out - cfg_scale * (eps_cond * c_out - eps_uncond * c_out)) / sigma
-                        = eps_uncond + cfg_scale * (eps_cond - eps_uncond)*/
-                    {
-                        float* vec_d = (float*)d->data;
-                        float* vec_eps_uncond = (float*)eps_uncond->data;
-                        float* vec_eps_cond = (float*)eps_cond->data;
-
-                        for (int i = 0; i < ggml_nelements(d); i++) {
-                            vec_d[i] = vec_eps_uncond[i] + cfg_scale * (vec_eps_cond[i] - vec_eps_uncond[i]);
-                        }
-                    }
-                } else {
-                    // cond
-                    copy_ggml_tensor(context, c);
-                    ggml_graph_compute(&diffusion_graph, &cplan);
-                    copy_ggml_tensor(d, eps);
-                }
-
                 // get_ancestral_step
                 float sigma_up = std::min(sigmas[i + 1],
                                           std::sqrt(sigmas[i + 1] * sigmas[i + 1] * (sigmas[i] * sigmas[i] - sigmas[i + 1] * sigmas[i + 1]) / (sigmas[i] * sigmas[i])));
@@ -3247,9 +3530,9 @@ class StableDiffusionGGML {
                 // x = x + d * dt
                 {
                     float* vec_d = (float*)d->data;
-                    float* vec_x = (float*)x_out->data;
+                    float* vec_x = (float*)x->data;
 
-                    for (int i = 0; i < ggml_nelements(x_out); i++) {
+                    for (int i = 0; i < ggml_nelements(x); i++) {
                         vec_x[i] = vec_x[i] + vec_d[i] * dt;
                     }
                 }
@@ -3259,25 +3542,14 @@ class StableDiffusionGGML {
                     ggml_tensor_set_f32_randn(noise);
                     // noise = load_tensor_from_file(res_ctx, "./rand" + std::to_string(i+1) + ".bin");
                     {
-                        float* vec_x = (float*)x_out->data;
+                        float* vec_x = (float*)x->data;
                         float* vec_noise = (float*)noise->data;
 
-                        for (int i = 0; i < ggml_nelements(x_out); i++) {
+                        for (int i = 0; i < ggml_nelements(x); i++) {
                             vec_x[i] = vec_x[i] + vec_noise[i] * sigma_up;
                         }
                     }
                 }
-
-#ifdef GGML_PERF
-                ggml_graph_print(&diffusion_graph);
-#endif
-                int64_t t1 = ggml_time_ms();
-                LOG_INFO("step %d sampling completed, taking %.2fs", i + 1, (t1 - t0) * 1.0f / 1000);
-                LOG_DEBUG("diffusion graph use %.2fMB runtime memory: static %.2fMB, dynamic %.2fMB",
-                          (ctx_size + ggml_curr_max_dynamic_size()) * 1.0f / 1024 / 1024,
-                          ctx_size * 1.0f / 1024 / 1024,
-                          ggml_curr_max_dynamic_size() * 1.0f / 1024 / 1024);
-                LOG_DEBUG("%zu bytes of dynamic memory has not been released yet", ggml_dynamic_size());
             }
         }
 
@@ -3304,7 +3576,7 @@ class StableDiffusionGGML {
 
         ggml_free(ctx);
 
-        return x_out;
+        return x;
     }
 
     ggml_tensor* encode_first_stage(ggml_context* res_ctx, ggml_tensor* x) {
@@ -3586,7 +3858,7 @@ std::vector<uint8_t> StableDiffusion::txt2img(const std::string& prompt,
     struct ggml_tensor* x_t = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, W, H, C, 1);
     ggml_tensor_set_f32_randn(x_t);
 
-    std::vector<float> sigmas = sd->denoiser.get_sigmas(sample_steps);
+    std::vector<float> sigmas = sd->denoiser->get_sigmas(sample_steps);
 
     LOG_INFO("start sampling");
     struct ggml_tensor* x_0 = sd->sample(ctx, x_t, c, uc, cfg_scale, sample_method, sigmas);
@@ -3642,7 +3914,7 @@ std::vector<uint8_t> StableDiffusion::img2img(const std::vector<uint8_t>& init_i
     }
     LOG_INFO("img2img %dx%d", width, height);
 
-    std::vector<float> sigmas = sd->denoiser.get_sigmas(sample_steps);
+    std::vector<float> sigmas = sd->denoiser->get_sigmas(sample_steps);
     size_t t_enc = static_cast<size_t>(sample_steps * strength);
     LOG_INFO("target t_enc is %zu steps", t_enc);
     std::vector<float> sigma_sched;