feat: added prediction argument (#334)

README.md

@@ -358,6 +358,7 @@ arguments:
   --rng {std_default, cuda}          RNG (default: cuda)
   -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)
   -b, --batch-count COUNT            number of images to generate
+  --prediction {eps, v, edm_v, sd3_flow, flux_flow} Prediction type override
   --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)
                                      <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   --vae-tiling                       process vae in tiles to reduce memory usage

examples/cli/main.cpp

@@ -84,6 +84,7 @@ struct SDParams {
 
     std::string prompt;
     std::string negative_prompt;
+
     int clip_skip   = -1;  // <= 0 represents unspecified
     int width       = 512;
     int height      = 512;
@@ -127,6 +128,8 @@ struct SDParams {
     int chroma_t5_mask_pad   = 1;
     float flow_shift         = INFINITY;
 
+    prediction_t prediction = DEFAULT_PRED;
+
     sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
 
     SDParams() {
@@ -188,6 +191,7 @@ void print_params(SDParams params) {
     printf("    sample_params:                     %s\n", SAFE_STR(sample_params_str));
     printf("    high_noise_sample_params:          %s\n", SAFE_STR(high_noise_sample_params_str));
     printf("    moe_boundary:                      %.3f\n", params.moe_boundary);
+    printf("    prediction:                        %s\n", sd_prediction_name(params.prediction));
     printf("    flow_shift:                        %.2f\n", params.flow_shift);
     printf("    strength(img2img):                 %.2f\n", params.strength);
     printf("    rng:                               %s\n", sd_rng_type_name(params.rng_type));
@@ -281,6 +285,7 @@ void print_usage(int argc, const char* argv[]) {
     printf("  --rng {std_default, cuda}          RNG (default: cuda)\n");
     printf("  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)\n");
     printf("  -b, --batch-count COUNT            number of images to generate\n");
+    printf("  --prediction {eps, v, edm_v, sd3_flow, flux_flow}        Prediction type override.\n");
     printf("  --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)\n");
     printf("                                     <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x\n");
     printf("  --vae-tiling                       process vae in tiles to reduce memory usage\n");
@@ -651,6 +656,20 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         return 1;
     };
 
+    auto on_prediction_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        const char* arg   = argv[index];
+        params.prediction = str_to_prediction(arg);
+        if (params.prediction == PREDICTION_COUNT) {
+            fprintf(stderr, "error: invalid prediction type %s\n",
+                    arg);
+            return -1;
+        }
+        return 1;
+    };
+
     auto on_sample_method_arg = [&](int argc, const char** argv, int index) {
         if (++index >= argc) {
             return -1;
@@ -807,6 +826,7 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         {"", "--rng", "", on_rng_arg},
         {"-s", "--seed", "", on_seed_arg},
         {"", "--sampling-method", "", on_sample_method_arg},
+        {"", "--prediction", "", on_prediction_arg},
         {"", "--scheduler", "", on_schedule_arg},
         {"", "--skip-layers", "", on_skip_layers_arg},
         {"", "--high-noise-sampling-method", "", on_high_noise_sample_method_arg},
@@ -1354,6 +1374,7 @@ int main(int argc, const char* argv[]) {
         params.n_threads,
         params.wtype,
         params.rng_type,
+        params.prediction,
         params.offload_params_to_cpu,
         params.clip_on_cpu,
         params.control_net_cpu,

stable-diffusion.cpp

@@ -700,64 +700,102 @@ public:
                 ggml_backend_is_cpu(clip_backend) ? "RAM" : "VRAM");
         }
 
-        // check is_using_v_parameterization_for_sd2
+        if (sd_ctx_params->prediction != DEFAULT_PRED) {
-        if (sd_version_is_sd2(version)) {
+            switch (sd_ctx_params->prediction) {
-            if (is_using_v_parameterization_for_sd2(ctx, sd_version_is_inpaint(version))) {
+                case EPS_PRED:
-                is_using_v_parameterization = true;
+                    LOG_INFO("running in eps-prediction mode");
-            }
+                    break;
-        } else if (sd_version_is_sdxl(version)) {
+                case V_PRED:
-            if (model_loader.tensor_storages_types.find("edm_vpred.sigma_max") != model_loader.tensor_storages_types.end()) {
+                    LOG_INFO("running in v-prediction mode");
-                // CosXL models
+                    denoiser = std::make_shared<CompVisVDenoiser>();
-                // TODO: get sigma_min and sigma_max values from file
+                    break;
-                is_using_edm_v_parameterization = true;
+                case EDM_V_PRED:
-            }
+                    LOG_INFO("running in v-prediction EDM mode");
-            if (model_loader.tensor_storages_types.find("v_pred") != model_loader.tensor_storages_types.end()) {
+                    denoiser = std::make_shared<EDMVDenoiser>();
-                is_using_v_parameterization = true;
+                    break;
-            }
+                case SD3_FLOW_PRED: {
-        } else if (version == VERSION_SVD) {
+                    LOG_INFO("running in FLOW mode");
-            // TODO: V_PREDICTION_EDM
+                    float shift = sd_ctx_params->flow_shift;
-            is_using_v_parameterization = true;
+                    if (shift == INFINITY) {
-        }
+                        shift = 3.0;
-
+                    }
-        if (sd_version_is_sd3(version)) {
+                    denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
-            LOG_INFO("running in FLOW mode");
-            float shift = sd_ctx_params->flow_shift;
-            if (shift == INFINITY) {
-                shift = 3.0;
-            }
-            denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
-        } else if (sd_version_is_flux(version)) {
-            LOG_INFO("running in Flux FLOW mode");
-            float shift = 1.0f;  // TODO: validate
-            for (auto pair : model_loader.tensor_storages_types) {
-                if (pair.first.find("model.diffusion_model.guidance_in.in_layer.weight") != std::string::npos) {
-                    shift = 1.15f;
                     break;
                 }
+                case FLUX_FLOW_PRED: {
+                    LOG_INFO("running in Flux FLOW mode");
+                    float shift = sd_ctx_params->flow_shift;
+                    if (shift == INFINITY) {
+                        shift = 3.0;
+                    }
+                    denoiser = std::make_shared<FluxFlowDenoiser>(shift);
+                    break;
+                }
+                default: {
+                    LOG_ERROR("Unknown parametrization %i", sd_ctx_params->prediction);
+                    return false;
+                }
             }
-            denoiser = std::make_shared<FluxFlowDenoiser>(shift);
-        } else if (sd_version_is_wan(version)) {
-            LOG_INFO("running in FLOW mode");
-            float shift = sd_ctx_params->flow_shift;
-            if (shift == INFINITY) {
-                shift = 5.0;
-            }
-            denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
-        } else if (sd_version_is_qwen_image(version)) {
-            LOG_INFO("running in FLOW mode");
-            float shift = sd_ctx_params->flow_shift;
-            if (shift == INFINITY) {
-                shift = 3.0;
-            }
-            denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
-        } else if (is_using_v_parameterization) {
-            LOG_INFO("running in v-prediction mode");
-            denoiser = std::make_shared<CompVisVDenoiser>();
-        } else if (is_using_edm_v_parameterization) {
-            LOG_INFO("running in v-prediction EDM mode");
-            denoiser = std::make_shared<EDMVDenoiser>();
         } else {
-            LOG_INFO("running in eps-prediction mode");
+            if (sd_version_is_sd2(version)) {
+                // check is_using_v_parameterization_for_sd2
+                if (is_using_v_parameterization_for_sd2(ctx, sd_version_is_inpaint(version))) {
+                    is_using_v_parameterization = true;
+                }
+            } else if (sd_version_is_sdxl(version)) {
+                if (model_loader.tensor_storages_types.find("edm_vpred.sigma_max") != model_loader.tensor_storages_types.end()) {
+                    // CosXL models
+                    // TODO: get sigma_min and sigma_max values from file
+                    is_using_edm_v_parameterization = true;
+                }
+                if (model_loader.tensor_storages_types.find("v_pred") != model_loader.tensor_storages_types.end()) {
+                    is_using_v_parameterization = true;
+                }
+            } else if (version == VERSION_SVD) {
+                // TODO: V_PREDICTION_EDM
+                is_using_v_parameterization = true;
+            }
+
+            if (sd_version_is_sd3(version)) {
+                LOG_INFO("running in FLOW mode");
+                float shift = sd_ctx_params->flow_shift;
+                if (shift == INFINITY) {
+                    shift = 3.0;
+                }
+                denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
+            } else if (sd_version_is_flux(version)) {
+                LOG_INFO("running in Flux FLOW mode");
+                float shift = 1.0f;  // TODO: validate
+                for (auto pair : model_loader.tensor_storages_types) {
+                    if (pair.first.find("model.diffusion_model.guidance_in.in_layer.weight") != std::string::npos) {
+                        shift = 1.15f;
+                        break;
+                    }
+                }
+                denoiser = std::make_shared<FluxFlowDenoiser>(shift);
+            } else if (sd_version_is_wan(version)) {
+                LOG_INFO("running in FLOW mode");
+                float shift = sd_ctx_params->flow_shift;
+                if (shift == INFINITY) {
+                    shift = 5.0;
+                }
+                denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
+            } else if (sd_version_is_qwen_image(version)) {
+                LOG_INFO("running in FLOW mode");
+                float shift = sd_ctx_params->flow_shift;
+                if (shift == INFINITY) {
+                    shift = 3.0;
+                }
+                denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
+            } else if (is_using_v_parameterization) {
+                LOG_INFO("running in v-prediction mode");
+                denoiser = std::make_shared<CompVisVDenoiser>();
+            } else if (is_using_edm_v_parameterization) {
+                LOG_INFO("running in v-prediction EDM mode");
+                denoiser = std::make_shared<EDMVDenoiser>();
+            } else {
+                LOG_INFO("running in eps-prediction mode");
+            }
         }
 
         auto comp_vis_denoiser = std::dynamic_pointer_cast<CompVisDenoiser>(denoiser);
@@ -1742,6 +1780,31 @@ enum scheduler_t str_to_schedule(const char* str) {
     return SCHEDULE_COUNT;
 }
 
+const char* prediction_to_str[] = {
+    "default",
+    "eps",
+    "v",
+    "edm_v",
+    "sd3_flow",
+    "flux_flow",
+};
+
+const char* sd_prediction_name(enum prediction_t prediction) {
+    if (prediction < PREDICTION_COUNT) {
+        return prediction_to_str[prediction];
+    }
+    return NONE_STR;
+}
+
+enum prediction_t str_to_prediction(const char* str) {
+    for (int i = 0; i < PREDICTION_COUNT; i++) {
+        if (!strcmp(str, prediction_to_str[i])) {
+            return (enum prediction_t)i;
+        }
+    }
+    return PREDICTION_COUNT;
+}
+
 void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     *sd_ctx_params                         = {};
     sd_ctx_params->vae_decode_only         = true;
@@ -1749,6 +1812,7 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     sd_ctx_params->n_threads               = get_num_physical_cores();
     sd_ctx_params->wtype                   = SD_TYPE_COUNT;
     sd_ctx_params->rng_type                = CUDA_RNG;
+    sd_ctx_params->prediction              = DEFAULT_PRED;
     sd_ctx_params->offload_params_to_cpu   = false;
     sd_ctx_params->keep_clip_on_cpu        = false;
     sd_ctx_params->keep_control_net_on_cpu = false;
@@ -1788,6 +1852,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              "n_threads: %d\n"
              "wtype: %s\n"
              "rng_type: %s\n"
+             "prediction: %s\n"
              "offload_params_to_cpu: %s\n"
              "keep_clip_on_cpu: %s\n"
              "keep_control_net_on_cpu: %s\n"
@@ -1816,6 +1881,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              sd_ctx_params->n_threads,
              sd_type_name(sd_ctx_params->wtype),
              sd_rng_type_name(sd_ctx_params->rng_type),
+             sd_prediction_name(sd_ctx_params->prediction),
              BOOL_STR(sd_ctx_params->offload_params_to_cpu),
              BOOL_STR(sd_ctx_params->keep_clip_on_cpu),
              BOOL_STR(sd_ctx_params->keep_control_net_on_cpu),

stable-diffusion.h

@@ -64,6 +64,16 @@ enum scheduler_t {
     SCHEDULE_COUNT
 };
 
+enum prediction_t {
+    DEFAULT_PRED,
+    EPS_PRED,
+    V_PRED,
+    EDM_V_PRED,
+    SD3_FLOW_PRED,
+    FLUX_FLOW_PRED,
+    PREDICTION_COUNT
+};
+
 // same as enum ggml_type
 enum sd_type_t {
     SD_TYPE_F32  = 0,
@@ -146,6 +156,7 @@ typedef struct {
     int n_threads;
     enum sd_type_t wtype;
     enum rng_type_t rng_type;
+    enum prediction_t prediction;
     bool offload_params_to_cpu;
     bool keep_clip_on_cpu;
     bool keep_control_net_on_cpu;
@@ -255,6 +266,8 @@ SD_API const char* sd_sample_method_name(enum sample_method_t sample_method);
 SD_API enum sample_method_t str_to_sample_method(const char* str);
 SD_API const char* sd_schedule_name(enum scheduler_t scheduler);
 SD_API enum scheduler_t str_to_schedule(const char* str);
+SD_API const char* sd_prediction_name(enum prediction_t prediction);
+SD_API enum prediction_t str_to_prediction(const char* str);
 
 SD_API void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params);
 SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);