feat: update submodule

Author: william-murray1204

stable_diffusion_cpp/_internals.py

@@ -58,7 +58,6 @@ def __init__(
         chroma_use_t5_mask: bool,
         chroma_t5_mask_pad: int,
         qwen_image_zero_cond_t: bool,
-        flow_shift: int,
         verbose: bool,
     ):
         self._exit_stack = ExitStack()
@@ -105,7 +104,6 @@ def __init__(
             chroma_use_t5_mask=chroma_use_t5_mask,
             chroma_t5_mask_pad=chroma_t5_mask_pad,
             qwen_image_zero_cond_t=qwen_image_zero_cond_t,
-            flow_shift=flow_shift,
         )
 
         # Load the free_sd_ctx function

stable_diffusion_cpp/stable_diffusion.py

@@ -73,7 +73,6 @@ def __init__(
         chroma_use_t5_mask: bool = False,
         chroma_t5_mask_pad: int = 1,
         qwen_image_zero_cond_t: bool = False,
-        flow_shift: float = float("inf"),
         image_resize_method: str = "crop",
         verbose: bool = True,
     ):
@@ -132,7 +131,6 @@ def __init__(
             chroma_use_t5_mask: Use T5 mask for Chroma.
             chroma_t5_mask_pad: T5 mask padding size of Chroma.
             qwen_image_zero_cond_t: Enable zero_cond_t for Qwen image.
-            flow_shift: Shift value for Flow models like SD3.x or WAN (default: auto).
             image_resize_method: Method to resize images for init, mask, control and reference images ("crop" or "resize").
             verbose: Print verbose output.
 
@@ -187,7 +185,6 @@ def __init__(
         self.chroma_use_t5_mask = chroma_use_t5_mask
         self.chroma_t5_mask_pad = chroma_t5_mask_pad
         self.qwen_image_zero_cond_t = qwen_image_zero_cond_t
-        self.flow_shift = flow_shift
         self.image_resize_method = image_resize_method
         self._stack = contextlib.ExitStack()
 
@@ -284,7 +281,6 @@ def __init__(
                     chroma_use_t5_mask=self.chroma_use_t5_mask,
                     chroma_t5_mask_pad=self.chroma_t5_mask_pad,
                     qwen_image_zero_cond_t=self.qwen_image_zero_cond_t,
-                    flow_shift=self.flow_shift,
                     verbose=self.verbose,
                 )
             )
@@ -346,6 +342,7 @@ def generate_image(
         eta: float = 0.0,
         timestep_shift: int = 0,
         sigmas: Optional[str] = None,
+        flow_shift: float = float("inf"),
         # slg_params
         skip_layers: List[int] = [7, 8, 9],
         skip_layer_start: float = 0.01,
@@ -379,6 +376,13 @@ def generate_image(
         cache_max_continuous_cached_steps: int = -1,
         cache_taylorseer_n_derivatives: int = 1,
         cache_taylorseer_skip_interval: int = 1,
+        cache_spectrum_w: float = 0.40,
+        cache_spectrum_m: int = 3,
+        cache_spectrum_lam: float = 1.0,
+        cache_spectrum_window_size: int = 2,
+        cache_spectrum_flex_window: float = 0.5,
+        cache_spectrum_warmup_steps: int = 4,
+        cache_spectrum_stop_percent: float = 0.9,
         scm_mask: str = "",
         scm_policy: Literal["dynamic", "static"] = "dynamic",
         # ---
@@ -412,6 +416,7 @@ def generate_image(
             eta: Eta in DDIM, only for DDIM and TCD.
             timestep_shift: Shift timestep for NitroFusion models, default: 0, recommended N for NitroSD-Realism around 250 and 500 for NitroSD-Vibrant.
             sigmas: Custom sigma values for the sampler, comma-separated (e.g. "14.61,7.8,3.5,0.0").
+            flow_shift: Shift value for Flow models like SD3.x or WAN (default: auto).
             skip_layers: Layers to skip for SLG steps (SLG will be enabled at step int([STEPS]x[START]) and disabled at int([STEPS]x[END])).
             skip_layer_start: SLG enabling point.
             skip_layer_end: SLG disabling point.
@@ -552,17 +557,17 @@ def sd_preview_callback(
         rel_size_x, rel_size_y = self._parse_tile_size(vae_relative_tile_size, as_float=True)
 
         # -------------------------------------------
-        # Scheduler/Sample Method
+        # Sample Method/Scheduler
         # -------------------------------------------
 
-        scheduler = self._validate_and_set_input(scheduler, SCHEDULER_MAP, "scheduler", allow_none=True)
-        if scheduler is None:
-            scheduler = sd_cpp.sd_get_default_scheduler(self.model)
-
         sample_method = self._validate_and_set_input(sample_method, SAMPLE_METHOD_MAP, "sample_method", allow_none=True)
         if sample_method is None:
             sample_method = sd_cpp.sd_get_default_sample_method(self.model)
 
+        scheduler = self._validate_and_set_input(scheduler, SCHEDULER_MAP, "scheduler", allow_none=True)
+        if scheduler is None:
+            scheduler = sd_cpp.sd_get_default_scheduler(self.model, sample_method)
+
         # -------------------------------------------
         # Sigmas
         # -------------------------------------------
@@ -590,6 +595,7 @@ def sd_preview_callback(
         # -------------------------------------------
 
         _cache_params = sd_cpp.sd_cache_params_t(
+            # General cache params
             mode=cache_mode,
             reuse_threshold=cache_reuse_threshold,
             start_percent=cache_start_percent,
@@ -602,8 +608,18 @@ def sd_preview_callback(
             residual_diff_threshold=cache_residual_diff_threshold,
             max_warmup_steps=cache_max_warmup_steps,
             max_continuous_cached_steps=cache_max_continuous_cached_steps,
+            # Taylorseer cache params
             taylorseer_n_derivatives=cache_taylorseer_n_derivatives,
             taylorseer_skip_interval=cache_taylorseer_skip_interval,
+            # Spectrum cache params
+            spectrum_w=cache_spectrum_w,
+            spectrum_m=cache_spectrum_m,
+            spectrum_lam=cache_spectrum_lam,
+            spectrum_window_size=cache_spectrum_window_size,
+            spectrum_flex_window=cache_spectrum_flex_window,
+            spectrum_warmup_steps=cache_spectrum_warmup_steps,
+            spectrum_stop_percent=cache_spectrum_stop_percent,
+            # General SCM params
             scm_mask=scm_mask.encode("utf-8"),
             scm_policy_dynamic=scm_policy,
         )
@@ -646,6 +662,7 @@ def sd_preview_callback(
             shifted_timestep=timestep_shift,
             custom_sigmas=_custom_sigmas,
             custom_sigmas_count=_custom_sigmas_count,
+            flow_shift=flow_shift,
         )
 
         _params = sd_cpp.sd_img_gen_params_t(
@@ -738,6 +755,7 @@ def generate_video(
         eta: float = 0.0,
         timestep_shift: int = 0,
         sigmas: Optional[str] = None,
+        flow_shift: float = float("inf"),
         # slg_params
         skip_layers: List[int] = [7, 8, 9],
         skip_layer_start: float = 0.01,
@@ -813,6 +831,7 @@ def generate_video(
             eta: Eta in DDIM, only for DDIM and TCD.
             timestep_shift: Shift timestep for NitroFusion models, default: 0, recommended N for NitroSD-Realism around 250 and 500 for NitroSD-Vibrant.
             sigmas: Custom sigma values for the sampler, comma-separated (e.g. "14.61,7.8,3.5,0.0").
+            flow_shift: Shift value for Flow models like SD3.x or WAN (default: auto).
             skip_layers: Layers to skip for SLG steps (SLG will be enabled at step int([STEPS]x[START]) and disabled at int([STEPS]x[END])).
             skip_layer_start: SLG enabling point.
             skip_layer_end: SLG disabling point.
@@ -965,13 +984,9 @@ def sd_preview_callback(
         rel_size_x, rel_size_y = self._parse_tile_size(vae_relative_tile_size, as_float=True)
 
         # -------------------------------------------
-        # Scheduler/Sample Method
+        # Sample Method/Scheduler
         # -------------------------------------------
 
-        scheduler = self._validate_and_set_input(scheduler, SCHEDULER_MAP, "scheduler", allow_none=True)
-        if scheduler is None:
-            scheduler = sd_cpp.sd_get_default_scheduler(self.model)
-
         # "sample_method_count" is not valid here (it will crash)
         sample_method = self._validate_and_set_input(
             sample_method,
@@ -982,19 +997,23 @@ def sd_preview_callback(
         if sample_method is None:
             sample_method = sd_cpp.sd_get_default_sample_method(self.model)
 
-        # High Noise
-        high_noise_scheduler = self._validate_and_set_input(
-            high_noise_scheduler, SCHEDULER_MAP, "high_noise_scheduler", allow_none=True
-        )
-        if high_noise_scheduler is None:
-            high_noise_scheduler = scheduler
+        scheduler = self._validate_and_set_input(scheduler, SCHEDULER_MAP, "scheduler", allow_none=True)
+        if scheduler is None:
+            scheduler = sd_cpp.sd_get_default_scheduler(self.model, sample_method)
 
+        # High Noise
         high_noise_sample_method = self._validate_and_set_input(
             high_noise_sample_method, SAMPLE_METHOD_MAP, "high_noise_sample_method", allow_none=True
         )
         if high_noise_sample_method is None:
             high_noise_sample_method = sample_method
 
+        high_noise_scheduler = self._validate_and_set_input(
+            high_noise_scheduler, SCHEDULER_MAP, "high_noise_scheduler", allow_none=True
+        )
+        if high_noise_scheduler is None:
+            high_noise_scheduler = scheduler
+
         # -------------------------------------------
         # Sigmas
         # -------------------------------------------
@@ -1043,6 +1062,7 @@ def sd_preview_callback(
             shifted_timestep=timestep_shift,
             custom_sigmas=_custom_sigmas,
             custom_sigmas_count=_custom_sigmas_count,
+            flow_shift=flow_shift,
         )
 
         # -------------------------------------------
@@ -1099,6 +1119,7 @@ def sd_preview_callback(
             shifted_timestep=timestep_shift,
             custom_sigmas=_custom_sigmas,
             custom_sigmas_count=_custom_sigmas_count,
+            flow_shift=flow_shift,
         )
 
         _params = sd_cpp.sd_vid_gen_params_t(
@@ -1978,4 +1999,5 @@ def __del__(self) -> None:
     "dbcache": SDCacheMode.SD_CACHE_DBCACHE,
     "taylorseer": SDCacheMode.SD_CACHE_TAYLORSEER,
     "cachedit": SDCacheMode.SD_CACHE_CACHE_DIT,
+    "spectrum": SDCacheMode.SD_CACHE_SPECTRUM,
 }

stable_diffusion_cpp/stable_diffusion_cpp.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import os
+import re
 import sys
 import ctypes
 import pathlib
@@ -40,6 +41,15 @@ def _load_shared_library(lib_base_name: str):
             _base_path / f"lib{lib_base_name}.dylib",
         ]
     elif sys.platform == "win32":
+        # Load the HIP runtime if present in the lib directory
+        def extract_version(p: pathlib.Path):
+            m = re.search(r"amdhip64[_\-]?(\d+)", p.name)
+            return int(m.group(1)) if m else 0
+
+        hip_dlls = sorted(_base_path.glob("amdhip64*.dll"), key=extract_version, reverse=True)
+        if hip_dlls:
+            ctypes.CDLL(str(hip_dlls[0]), winmode=0x8)
+
         _lib_paths += [
             _base_path / f"{lib_base_name}.dll",
             _base_path / f"lib{lib_base_name}.dll",
@@ -144,9 +154,9 @@ def byref(obj: CtypesCData, offset: Optional[int] = None) -> CtypesRef[CtypesCDa
 ggml_abort_callback = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_void_p)
 
 
-################################################
-# stable-diffusion.h bindings
-################################################
+# ===========================================
+# include/stable-diffusion.h bindings
+# ===========================================
 
 
 # enum rng_type_t {
@@ -369,6 +379,7 @@ class LoraApplyMode(IntEnum):
 #     SD_CACHE_DBCACHE,
 #     SD_CACHE_TAYLORSEER,
 #     SD_CACHE_CACHE_DIT,
+#     SD_CACHE_SPECTRUM,
 # };
 class SDCacheMode(IntEnum):
     SD_CACHE_DISABLED = 0
@@ -377,6 +388,7 @@ class SDCacheMode(IntEnum):
     SD_CACHE_DBCACHE = 3
     SD_CACHE_TAYLORSEER = 4
     SD_CACHE_CACHE_DIT = 5
+    SD_CACHE_SPECTRUM = 6
 
 
 # ===========================================
@@ -402,7 +414,7 @@ class sd_embedding_t(ctypes.Structure):
 # -------------------------------------------
 
 
-# typedef struct { const char* model_path; const char* clip_l_path; const char* clip_g_path; const char* clip_vision_path; const char* t5xxl_path; const char* llm_path; const char* llm_vision_path; const char* diffusion_model_path; const char* high_noise_diffusion_model_path; const char* vae_path; const char* taesd_path; const char* control_net_path; const sd_embedding_t* embeddings; uint32_t embedding_count; const char* photo_maker_path; const char* tensor_type_rules; bool vae_decode_only; bool free_params_immediately; int n_threads; enum sd_type_t wtype; enum rng_type_t rng_type; enum rng_type_t sampler_rng_type; enum prediction_t prediction; enum lora_apply_mode_t lora_apply_mode; bool offload_params_to_cpu; bool enable_mmap; bool keep_clip_on_cpu; bool keep_control_net_on_cpu; bool keep_vae_on_cpu; bool flash_attn; bool diffusion_flash_attn; bool tae_preview_only; bool diffusion_conv_direct; bool vae_conv_direct; bool circular_x; bool circular_y; bool force_sdxl_vae_conv_scale; bool chroma_use_dit_mask; bool chroma_use_t5_mask; int chroma_t5_mask_pad; bool qwen_image_zero_cond_t; float flow_shift; } sd_ctx_params_t;
+# typedef struct { const char* model_path; const char* clip_l_path; const char* clip_g_path; const char* clip_vision_path; const char* t5xxl_path; const char* llm_path; const char* llm_vision_path; const char* diffusion_model_path; const char* high_noise_diffusion_model_path; const char* vae_path; const char* taesd_path; const char* control_net_path; const sd_embedding_t* embeddings; uint32_t embedding_count; const char* photo_maker_path; const char* tensor_type_rules; bool vae_decode_only; bool free_params_immediately; int n_threads; enum sd_type_t wtype; enum rng_type_t rng_type; enum rng_type_t sampler_rng_type; enum prediction_t prediction; enum lora_apply_mode_t lora_apply_mode; bool offload_params_to_cpu; bool enable_mmap; bool keep_clip_on_cpu; bool keep_control_net_on_cpu; bool keep_vae_on_cpu; bool flash_attn; bool diffusion_flash_attn; bool tae_preview_only; bool diffusion_conv_direct; bool vae_conv_direct; bool circular_x; bool circular_y; bool force_sdxl_vae_conv_scale; bool chroma_use_dit_mask; bool chroma_use_t5_mask; int chroma_t5_mask_pad; bool qwen_image_zero_cond_t; } sd_ctx_params_t;
 class sd_ctx_params_t(ctypes.Structure):
     _fields_ = [
         ("model_path", ctypes.c_char_p),
@@ -446,7 +458,6 @@ class sd_ctx_params_t(ctypes.Structure):
         ("chroma_use_t5_mask", ctypes.c_bool),
         ("chroma_t5_mask_pad", ctypes.c_int),
         ("qwen_image_zero_cond_t", ctypes.c_bool),
-        ("flow_shift", ctypes.c_float),
     ]
 
 
@@ -586,7 +597,7 @@ class sd_guidance_params_t(ctypes.Structure):
 # -------------------------------------------
 
 
-# typedef struct { sd_guidance_params_t guidance; enum scheduler_t scheduler; enum sample_method_t sample_method; int sample_steps; float eta; int shifted_timestep; float* custom_sigmas; int custom_sigmas_count; } sd_sample_params_t;
+# typedef struct { sd_guidance_params_t guidance; enum scheduler_t scheduler; enum sample_method_t sample_method; int sample_steps; float eta; int shifted_timestep; float* custom_sigmas; int custom_sigmas_count; float flow_shift; } sd_sample_params_t;
 class sd_sample_params_t(ctypes.Structure):
     _fields_ = [
         ("guidance", sd_guidance_params_t),
@@ -597,6 +608,7 @@ class sd_sample_params_t(ctypes.Structure):
         ("shifted_timestep", ctypes.c_int),
         ("custom_sigmas", ctypes.POINTER(ctypes.c_float)),
         ("custom_sigmas_count", ctypes.c_int),
+        ("flow_shift", ctypes.c_float),
     ]
 
 
@@ -605,7 +617,7 @@ class sd_sample_params_t(ctypes.Structure):
 # -------------------------------------------
 
 
-# typedef struct { enum sd_cache_mode_t mode; float reuse_threshold; float start_percent; float end_percent; float error_decay_rate; bool use_relative_threshold; bool reset_error_on_compute; int Fn_compute_blocks; int Bn_compute_blocks; float residual_diff_threshold; int max_warmup_steps; int max_cached_steps; int max_continuous_cached_steps; int taylorseer_n_derivatives; int taylorseer_skip_interval; const char* scm_mask; bool scm_policy_dynamic; } sd_cache_params_t;
+# typedef struct { enum sd_cache_mode_t mode; float reuse_threshold; float start_percent; float end_percent; float error_decay_rate; bool use_relative_threshold; bool reset_error_on_compute; int Fn_compute_blocks; int Bn_compute_blocks; float residual_diff_threshold; int max_warmup_steps; int max_cached_steps; int max_continuous_cached_steps; int taylorseer_n_derivatives; int taylorseer_skip_interval; const char* scm_mask; bool scm_policy_dynamic; float spectrum_w; int spectrum_m; float spectrum_lam; int spectrum_window_size; float spectrum_flex_window; int spectrum_warmup_steps; float spectrum_stop_percent; } sd_cache_params_t;
 class sd_cache_params_t(ctypes.Structure):
     _fields_ = [
         ("mode", ctypes.c_int),  # SDCacheMode
@@ -625,6 +637,13 @@ class sd_cache_params_t(ctypes.Structure):
         ("taylorseer_skip_interval", ctypes.c_int),
         ("scm_mask", ctypes.c_char_p),
         ("scm_policy_dynamic", ctypes.c_bool),
+        ("spectrum_w", ctypes.c_float),
+        ("spectrum_m", ctypes.c_int),
+        ("spectrum_lam", ctypes.c_float),
+        ("spectrum_window_size", ctypes.c_int),
+        ("spectrum_flex_window", ctypes.c_float),
+        ("spectrum_warmup_steps", ctypes.c_int),
+        ("spectrum_stop_percent", ctypes.c_float),
     ]
 
 
@@ -777,16 +796,18 @@ def sd_get_default_sample_method(
 # -------------------------------------------
 
 
-# SD_API enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx);
+# SD_API enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx, enum sample_method_t sample_method);
 @ctypes_function(
     "sd_get_default_scheduler",
     [
         sd_ctx_t_p_ctypes,  # sd_ctx
+        ctypes.c_int,  # sample_method
     ],
     ctypes.c_int,  # Scheduler
 )
 def sd_get_default_scheduler(
     sd_ctx: sd_ctx_t_p,
+    sample_method: SampleMethod,
     /,
 ) -> Optional[Scheduler]: ...
 
@@ -1032,7 +1053,7 @@ def sd_version() -> bytes:
 @ctypes_function(
     "sd_set_progress_callback",
     [
-        ctypes.c_void_p,  # sd_progress_cb_t
+        ctypes.c_void_p,  # cb
         ctypes.c_void_p,  # data
     ],
     None,
@@ -1060,7 +1081,7 @@ def sd_set_progress_callback(
 @ctypes_function(
     "sd_set_preview_callback",
     [
-        ctypes.c_void_p,  # sd_preview_cb_t
+        ctypes.c_void_p,  # cb
         ctypes.c_int,  # mode
         ctypes.c_int,  # interval
         ctypes.c_bool,  # denoised
@@ -1092,7 +1113,10 @@ def sd_set_preview_callback(
 # SD_API void sd_set_log_callback(sd_log_cb_t sd_log_cb, void* data);
 @ctypes_function(
     "sd_set_log_callback",
-    [ctypes.c_void_p, ctypes.c_void_p],
+    [
+        ctypes.c_void_p,  # sd_log_cb
+        ctypes.c_void_p,  # data
+    ],
     None,
 )
 def sd_set_log_callback(