fix: hires now uses more optimal steps for second pass; feat: hires on layer diffuse; feat: comfyui 2dc84d14

hordelib/comfy_horde.py

@@ -213,14 +213,7 @@ def do_comfy_import(
 
         # comfy.model_management.unet_offload_device = _unet_offload_device_hijack
 
-        total_vram = get_torch_total_vram_mb()
-        total_ram = psutil.virtual_memory().total / (1024 * 1024)
-        free_ram = psutil.virtual_memory().available / (1024 * 1024)
-
-        free_vram = get_torch_free_vram_mb()
-
-        logger.debug(f"Total VRAM {total_vram:0.0f} MB, Total System RAM {total_ram:0.0f} MB")
-        logger.debug(f"Free VRAM {free_vram:0.0f} MB, Free System RAM {free_ram:0.0f} MB")
+    log_free_ram()
     output_collector.replay()
 
 
@@ -308,6 +301,11 @@ def get_torch_free_vram_mb():
     return round(_comfy_get_free_memory() / (1024 * 1024))
 
 
+def log_free_ram():
+    logger.debug(f"Free VRAM: {get_torch_free_vram_mb():0.0f} MB")
+    logger.debug(f"Free RAM: {psutil.virtual_memory().available / (1024 * 1024):0.0f} MB")
+
+
 class Comfy_Horde:
     """Handles horde-specific behavior against ComfyUI."""
 
@@ -446,7 +444,7 @@ def _this_dir(self, filename: str, subdir="") -> str:
 
     def _load_custom_nodes(self) -> None:
         """Force ComfyUI to load its normal custom nodes and the horde custom nodes."""
-        _comfy_nodes.init_custom_nodes()
+        _comfy_nodes.init_extra_nodes(init_custom_nodes=True)
 
     def _get_executor(self):
         """Return the ComfyUI PromptExecutor object."""
@@ -744,7 +742,7 @@ def _run_pipeline(
         # if time.time() - self._gc_timer > Comfy_Horde.GC_TIME:
         #     self._gc_timer = time.time()
         #     garbage_collect()
-
+        log_free_ram()
         return self.images
 
     # Run a pipeline that returns an image in pixel space

hordelib/consts.py

@@ -6,7 +6,7 @@
 
 from hordelib.config_path import get_hordelib_path
 
-COMFYUI_VERSION = "16eabdf70dbdb64dc4822908f0fe455c56d11ec3"
+COMFYUI_VERSION = "2dc84d14447782683862616eaf8c19c0c1feacf3"
 """The exact version of ComfyUI version to load."""
 
 REMOTE_PROXY = ""

hordelib/horde.py

@@ -11,6 +11,7 @@
 from collections.abc import Callable
 from copy import deepcopy
 from enum import Enum, auto
+from types import FunctionType
 
 from horde_sdk.ai_horde_api.apimodels import ImageGenerateJobPopResponse
 from horde_sdk.ai_horde_api.apimodels.base import (
@@ -77,6 +78,17 @@ def __init__(
         self.faults = faults
 
 
+def _calc_upscale_sampler_steps(payload):
+    """Calculates the amount of hires_fix upscaler steps based on the denoising used and the steps used for the
+    primary image"""
+    upscale_steps = round(payload["ddim_steps"] * (0.9 - payload["hires_fix_denoising_strength"]))
+    if upscale_steps < 3:
+        upscale_steps = 3
+
+    logger.debug(f"Upscale steps calculated as {upscale_steps}")
+    return upscale_steps
+
+
 class HordeLib:
     _instance: HordeLib | None = None
     _initialised = False
@@ -227,7 +239,7 @@ class HordeLib:
         "upscale_sampler.denoise": "hires_fix_denoising_strength",
         "upscale_sampler.seed": "seed",
         "upscale_sampler.cfg": "cfg_scale",
-        "upscale_sampler.steps": "ddim_steps",
+        "upscale_sampler.steps": _calc_upscale_sampler_steps,
         "upscale_sampler.sampler_name": "sampler_name",
         "controlnet_apply.strength": "control_strength",
         "controlnet_model_loader.control_net_name": "control_type",
@@ -243,6 +255,8 @@ class HordeLib:
         "sampler_stage_c.denoise": "denoising_strength",
         "sampler_stage_b.seed": "seed",
         "sampler_stage_c.seed": "seed",
+        "sampler_stage_b.steps": "ddim_steps*0.33",
+        "sampler_stage_c.steps": "ddim_steps*0.67",
         "model_loader_stage_c.ckpt_name": "stable_cascade_stage_c",
         "model_loader_stage_c.model_name": "stable_cascade_stage_c",
         "model_loader_stage_c.horde_model_name": "model_name",
@@ -251,8 +265,10 @@ class HordeLib:
         "model_loader_stage_b.horde_model_name": "model_name",
         # Stable Cascade 2pass
         "2pass_sampler_stage_c.sampler_name": "sampler_name",
+        "2pass_sampler_stage_c.steps": "ddim_steps*0.67",
         "2pass_sampler_stage_c.denoise": "hires_fix_denoising_strength",
         "2pass_sampler_stage_b.sampler_name": "sampler_name",
+        "2pass_sampler_stage_b.steps": "ddim_steps*0.33",
         # QR Codes
         "sampler_bg.sampler_name": "sampler_name",
         "sampler_bg.cfg": "cfg_scale",
@@ -519,8 +535,15 @@ def _apply_aihorde_compatibility_hacks(self, payload: dict) -> tuple[dict, list[
 
         # Turn off hires fix if we're not generating a hires image, or if the params are just confused
         try:
-            if "hires_fix" in payload and (payload["width"] <= 512 or payload["height"] <= 512):
-                payload["hires_fix"] = False
+            if "hires_fix" in payload:
+                if SharedModelManager.manager.compvis.model_reference[model].get(
+                    "baseline",
+                ) == "stable diffusion 1" and (payload["width"] <= 512 or payload["height"] <= 512):
+                    payload["hires_fix"] = False
+                elif SharedModelManager.manager.compvis.model_reference[model].get(
+                    "baseline",
+                ) == "stable_diffusion_xl" and (payload["width"] <= 1024 or payload["height"] <= 1024):
+                    payload["hires_fix"] = False
         except (TypeError, KeyError):
             payload["hires_fix"] = False
 
@@ -792,8 +815,18 @@ def _final_pipeline_adjustments(self, payload, pipeline_data) -> tuple[dict, lis
         # Translate the payload parameters into pipeline parameters
         pipeline_params = {}
         for newkey, key in HordeLib.PAYLOAD_TO_PIPELINE_PARAMETER_MAPPING.items():
-            if key in payload:
-                pipeline_params[newkey] = payload.get(key)
+            multiplier = None
+            # We allow a multiplier in the param, so that I can adjust easily the
+            # values for steps on things like stable cascade
+            if isinstance(key, FunctionType):
+                pipeline_params[newkey] = key(payload)
+            elif "*" in key:
+                key, multiplier = key.split("*", 1)
+            elif key in payload:
+                if multiplier:
+                    pipeline_params[newkey] = round(payload.get(key) * float(multiplier))
+                else:
+                    pipeline_params[newkey] = payload.get(key)
             else:
                 logger.error(f"Parameter {key} not found")
         # We inject these parameters to ensure the HordeCheckpointLoader knows what file to load, if necessary
@@ -827,8 +860,12 @@ def _final_pipeline_adjustments(self, payload, pipeline_data) -> tuple[dict, lis
             original_height = pipeline_params.get("empty_latent_image.height")
 
             if original_width is None or original_height is None:
-                logger.error("empty_latent_image.width or empty_latent_image.height not found. Using 512x512.")
-                original_width, original_height = (512, 512)
+                if model_details and model_details.get("baseline") == "stable diffusion 1":
+                    logger.error("empty_latent_image.width or empty_latent_image.height not found. Using 512x512.")
+                    original_width, original_height = (512, 512)
+                else:
+                    logger.error("empty_latent_image.width or empty_latent_image.height not found. Using 1024x1024.")
+                    original_width, original_height = (1024, 1024)
 
             new_width, new_height = (None, None)
 
@@ -1041,6 +1078,8 @@ def _final_pipeline_adjustments(self, payload, pipeline_data) -> tuple[dict, lis
                     self.generator.reconnect_input(pipeline_data, "layer_diffuse_apply.model", "model_loader")
                     self.generator.reconnect_input(pipeline_data, "output_image.images", "layer_diffuse_decode_rgba")
                     self.generator.reconnect_input(pipeline_data, "layer_diffuse_decode_rgba.images", "vae_decode")
+                    if payload.get("hires_fix") is True:
+                        self.generator.reconnect_input(pipeline_data, "upscale_sampler.model", "layer_diffuse_apply")
                     if model_details.get("baseline") == "stable diffusion 1":
                         pipeline_params["layer_diffuse_apply.config"] = "SD15, Attention Injection, attn_sharing"
                         pipeline_params["layer_diffuse_decode_rgba.sd_version"] = "SD15"
@@ -1489,6 +1528,11 @@ def basic_inference_rawpng(self, payload: dict) -> list[io.BytesIO]:
 
     def image_upscale(self, payload) -> ResultingImageReturn:
         logger.debug("image_upscale called")
+
+        from hordelib.comfy_horde import log_free_ram
+
+        log_free_ram()
+
         # AIHorde hacks to payload
         payload, compatibility_faults = self._apply_aihorde_compatibility_hacks(payload)
         # Remember if we were passed width and height, we wouldn't normally be passed width and height
@@ -1522,10 +1566,16 @@ def image_upscale(self, payload) -> ResultingImageReturn:
         if not isinstance(image, Image.Image):
             raise RuntimeError(f"Expected a PIL.Image.Image but got {type(image)}")
 
+        log_free_ram()
         return ResultingImageReturn(image=image, rawpng=rawpng, faults=compatibility_faults + final_adjustment_faults)
 
     def image_facefix(self, payload) -> ResultingImageReturn:
         logger.debug("image_facefix called")
+
+        from hordelib.comfy_horde import log_free_ram
+
+        log_free_ram()
+
         # AIHorde hacks to payload
         payload, compatibility_faults = self._apply_aihorde_compatibility_hacks(payload)
         # Check payload types/values and normalise it's format
@@ -1547,4 +1597,6 @@ def image_facefix(self, payload) -> ResultingImageReturn:
         if not isinstance(image, Image.Image):
             raise RuntimeError(f"Expected a PIL.Image.Image but got {type(image)}")
 
+        log_free_ram()
+
         return ResultingImageReturn(image=image, rawpng=rawpng, faults=compatibility_faults + final_adjustment_faults)

hordelib/nodes/node_controlnet_model_loader.py

@@ -18,16 +18,21 @@ def INPUT_TYPES(s):
     CATEGORY = "loaders"
 
     def load_controlnet(self, model, control_net_name, model_manager):
+        from hordelib.comfy_horde import log_free_ram
+
         logger.debug(f"Loading controlnet {control_net_name} through our custom node")
+        log_free_ram()
 
         if not model_manager or not model_manager.manager or not model_manager.manager.controlnet:
             logger.error("controlnet model_manager appears to be missing!")
             raise RuntimeError  # XXX better guarantees need to be made
 
-        return model_manager.manager.controlnet.merge_controlnet(
+        merge_result = model_manager.manager.controlnet.merge_controlnet(
             control_net_name,
             model,
         )
+        log_free_ram()
+        return merge_result
 
 
 NODE_CLASS_MAPPINGS = {"HordeDiffControlNetLoader": HordeDiffControlNetLoader}

hordelib/nodes/node_lora_loader.py

@@ -27,11 +27,18 @@ def INPUT_TYPES(s):
     CATEGORY = "loaders"
 
     def load_lora(self, model, clip, lora_name, strength_model, strength_clip):
+        from hordelib.comfy_horde import log_free_ram
+
+        log_free_ram()
+
         _test_exception = os.getenv("FAILURE_TEST", False)
         if _test_exception:
             raise Exception("This tests exceptions being thrown from within the pipeline")
 
+        logger.debug(f"Loading lora {lora_name} through our custom node")
+
         if strength_model == 0 and strength_clip == 0:
+            logger.debug("Strengths are 0, skipping lora loading")
             return (model, clip)
 
         if lora_name is None or lora_name == "" or lora_name == "None":
@@ -67,6 +74,7 @@ def load_lora(self, model, clip, lora_name, strength_model, strength_clip):
             self.loaded_lora = (lora_path, lora)
 
         model_lora, clip_lora = comfy.sd.load_lora_for_models(model, clip, lora, strength_model, strength_clip)
+        log_free_ram()
         return (model_lora, clip_lora)
 
 

hordelib/nodes/node_model_loader.py

@@ -10,6 +10,7 @@
 from loguru import logger
 
 from hordelib.shared_model_manager import SharedModelManager
+from hordelib.comfy_horde import log_free_ram
 
 
 # Don't let the name fool you, this class is trying to load all the files that will be necessary
@@ -44,6 +45,7 @@ def load_checkpoint(
         output_clip=True,
         preloading=False,
     ):
+        log_free_ram()
         if file_type is not None:
             logger.debug(f"Loading model {horde_model_name}:{file_type}")
         else:
@@ -77,7 +79,7 @@ def load_checkpoint(
                 make_regular_vae(same_loaded_model[0][2])
 
             logger.debug("Model was previously loaded, returning it.")
-
+            log_free_ram()
             return same_loaded_model[0]
 
         if not ckpt_name:
@@ -133,6 +135,7 @@ def load_checkpoint(
             result[0].model.apply(make_regular)
             make_regular_vae(result[2])
 
+        log_free_ram()
         return result
 
 

hordelib/nodes/node_upscale_model_loader.py

@@ -18,11 +18,15 @@ def INPUT_TYPES(s):
     CATEGORY = "loaders"
 
     def load_model(self, model_name):
+        from hordelib.comfy_horde import log_free_ram
+
+        log_free_ram()
         model_path = folder_paths.get_full_path("upscale_models", model_name)
         sd = comfy.utils.load_torch_file(model_path, safe_load=True)
         if "module.layers.0.residual_group.blocks.0.norm1.weight" in sd:
             sd = comfy.utils.state_dict_prefix_replace(sd, {"module.": ""})
         out = model_loading.load_state_dict(sd).eval()
+        log_free_ram()
         return (out,)