Merge branch 'comfyanonymous:master' into refactor/onprompt

.github/workflows/windows_release_nightly_pytorch.yml

@@ -31,7 +31,7 @@ jobs:
             echo 'import site' >> ./python311._pth
             curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
             ./python.exe get-pip.py
-            python -m pip wheel torch torchvision torchaudio aiohttp==3.8.4 --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu121 -r ../ComfyUI/requirements.txt pygit2 -w ../temp_wheel_dir
+            python -m pip wheel torch torchvision torchaudio aiohttp==3.8.5 --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu121 -r ../ComfyUI/requirements.txt pygit2 -w ../temp_wheel_dir
             ls ../temp_wheel_dir
             ./python.exe -s -m pip install --pre ../temp_wheel_dir/*
             sed -i '1i../ComfyUI' ./python311._pth

README.md

@@ -47,6 +47,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 | Ctrl + O                  | Load workflow                                                                                                      |
 | Ctrl + A                  | Select all nodes                                                                                                   |
 | Ctrl + M                  | Mute/unmute selected nodes                                                                                         |
+| Ctrl + B                  | Bypass selected nodes (acts like the node was removed from the graph and the wires reconnected through)            |
 | Delete/Backspace          | Delete selected nodes                                                                                              |
 | Ctrl + Delete/Backspace   | Delete the current graph                                                                                           |
 | Space                     | Move the canvas around when held and moving the cursor                                                             |

comfy/cli_args.py

@@ -82,6 +82,9 @@ vram_group.add_argument("--novram", action="store_true", help="When lowvram isn'
 vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).")
 
 
+parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
+
+
 parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
 parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")
 parser.add_argument("--windows-standalone-build", action="store_true", help="Windows standalone build: Enable convenient things that most people using the standalone windows build will probably enjoy (like auto opening the page on startup).")

comfy/clip_vision.py

@@ -25,6 +25,7 @@ class ClipVisionModel():
 
     def encode_image(self, image):
         img = torch.clip((255. * image), 0, 255).round().int()
+        img = list(map(lambda a: a, img))
         inputs = self.processor(images=img, return_tensors="pt")
         outputs = self.model(**inputs)
         return outputs

comfy/gligen.py

@@ -244,30 +244,15 @@ class Gligen(nn.Module):
         self.position_net = position_net
         self.key_dim = key_dim
         self.max_objs = 30
-        self.lowvram = False
+        self.current_device = torch.device("cpu")
 
     def _set_position(self, boxes, masks, positive_embeddings):
-        if self.lowvram == True:
-            self.position_net.to(boxes.device)
-
         objs = self.position_net(boxes, masks, positive_embeddings)
-
+        def func(x, extra_options):
-        if self.lowvram == True:
+            key = extra_options["transformer_index"]
-            self.position_net.cpu()
+            module = self.module_list[key]
-            def func_lowvram(x, extra_options):
+            return module(x, objs)
-                key = extra_options["transformer_index"]
+        return func
-                module = self.module_list[key]
-                module.to(x.device)
-                r = module(x, objs)
-                module.cpu()
-                return r
-            return func_lowvram
-        else:
-            def func(x, extra_options):
-                key = extra_options["transformer_index"]
-                module = self.module_list[key]
-                return module(x, objs)
-            return func
 
     def set_position(self, latent_image_shape, position_params, device):
         batch, c, h, w = latent_image_shape
@@ -312,14 +297,6 @@ class Gligen(nn.Module):
             masks.to(device),
             conds.to(device))
 
-    def set_lowvram(self, value=True):
-        self.lowvram = value
-
-    def cleanup(self):
-        self.lowvram = False
-
-    def get_models(self):
-        return [self]
 
 def load_gligen(sd):
     sd_k = sd.keys()

comfy/k_diffusion/sampling.py

@@ -649,7 +649,7 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     s_in = x.new_ones([x.shape[0]])
 
     denoised_1, denoised_2 = None, None
-    h_1, h_2 = None, None
+    h, h_1, h_2 = None, None, None
 
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)

comfy/model_detection.py

@@ -121,9 +121,20 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_fp16):
     return model_config_from_unet_config(unet_config)
 
 
-def model_config_from_diffusers_unet(state_dict, use_fp16):
+def unet_config_from_diffusers_unet(state_dict, use_fp16):
     match = {}
-    match["context_dim"] = state_dict["down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_k.weight"].shape[1]
+    attention_resolutions = []
+
+    attn_res = 1
+    for i in range(5):
+        k = "down_blocks.{}.attentions.1.transformer_blocks.0.attn2.to_k.weight".format(i)
+        if k in state_dict:
+            match["context_dim"] = state_dict[k].shape[1]
+            attention_resolutions.append(attn_res)
+        attn_res *= 2
+
+    match["attention_resolutions"] = attention_resolutions
+
     match["model_channels"] = state_dict["conv_in.weight"].shape[0]
     match["in_channels"] = state_dict["conv_in.weight"].shape[1]
     match["adm_in_channels"] = None
@@ -135,22 +146,22 @@ def model_config_from_diffusers_unet(state_dict, use_fp16):
     SDXL = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
             'num_classes': 'sequential', 'adm_in_channels': 2816, 'use_fp16': use_fp16, 'in_channels': 4, 'model_channels': 320,
             'num_res_blocks': 2, 'attention_resolutions': [2, 4], 'transformer_depth': [0, 2, 10], 'channel_mult': [1, 2, 4],
-            'transformer_depth_middle': 10, 'use_linear_in_transformer': True, 'context_dim': 2048}
+            'transformer_depth_middle': 10, 'use_linear_in_transformer': True, 'context_dim': 2048, "num_head_channels": 64}
 
     SDXL_refiner = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
                     'num_classes': 'sequential', 'adm_in_channels': 2560, 'use_fp16': use_fp16, 'in_channels': 4, 'model_channels': 384,
                     'num_res_blocks': 2, 'attention_resolutions': [2, 4], 'transformer_depth': [0, 4, 4, 0], 'channel_mult': [1, 2, 4, 4],
-                    'transformer_depth_middle': 4, 'use_linear_in_transformer': True, 'context_dim': 1280}
+                    'transformer_depth_middle': 4, 'use_linear_in_transformer': True, 'context_dim': 1280, "num_head_channels": 64}
 
     SD21 = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
             'adm_in_channels': None, 'use_fp16': use_fp16, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': 2,
             'attention_resolutions': [1, 2, 4], 'transformer_depth': [1, 1, 1, 0], 'channel_mult': [1, 2, 4, 4],
-            'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024}
+            'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, "num_head_channels": 64}
 
     SD21_uncliph = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
                     'num_classes': 'sequential', 'adm_in_channels': 2048, 'use_fp16': use_fp16, 'in_channels': 4, 'model_channels': 320,
                     'num_res_blocks': 2, 'attention_resolutions': [1, 2, 4], 'transformer_depth': [1, 1, 1, 0], 'channel_mult': [1, 2, 4, 4],
-                    'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024}
+                    'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, "num_head_channels": 64}
 
     SD21_unclipl = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
                     'num_classes': 'sequential', 'adm_in_channels': 1536, 'use_fp16': use_fp16, 'in_channels': 4, 'model_channels': 320,
@@ -160,9 +171,20 @@ def model_config_from_diffusers_unet(state_dict, use_fp16):
     SD15 = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
             'adm_in_channels': None, 'use_fp16': use_fp16, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': 2,
             'attention_resolutions': [1, 2, 4], 'transformer_depth': [1, 1, 1, 0], 'channel_mult': [1, 2, 4, 4],
-            'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768}
+            'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, "num_heads": 8}
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl]
+    SDXL_mid_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'num_classes': 'sequential', 'adm_in_channels': 2816, 'use_fp16': use_fp16, 'in_channels': 4, 'model_channels': 320,
+            'num_res_blocks': 2, 'attention_resolutions': [4], 'transformer_depth': [0, 0, 1], 'channel_mult': [1, 2, 4],
+            'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 2048, "num_head_channels": 64}
+
+    SDXL_small_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'num_classes': 'sequential', 'adm_in_channels': 2816, 'use_fp16': use_fp16, 'in_channels': 4, 'model_channels': 320,
+            'num_res_blocks': 2, 'attention_resolutions': [], 'transformer_depth': [0, 0, 0], 'channel_mult': [1, 2, 4],
+            'transformer_depth_middle': 0, 'use_linear_in_transformer': True, "num_head_channels": 64, 'context_dim': 1}
+
+
+    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet]
 
     for unet_config in supported_models:
         matches = True
@@ -171,5 +193,11 @@ def model_config_from_diffusers_unet(state_dict, use_fp16):
                 matches = False
                 break
         if matches:
-            return model_config_from_unet_config(unet_config)
+            return unet_config
+    return None
+
+def model_config_from_diffusers_unet(state_dict, use_fp16):
+    unet_config = unet_config_from_diffusers_unet(state_dict, use_fp16)
+    if unet_config is not None:
+        return model_config_from_unet_config(unet_config)
     return None

comfy/model_management.py

@@ -2,6 +2,7 @@ import psutil
 from enum import Enum
 from comfy.cli_args import args
 import torch
+import sys
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -201,6 +202,10 @@ if cpu_state == CPUState.MPS:
 
 print(f"Set vram state to: {vram_state.name}")
 
+DISABLE_SMART_MEMORY = args.disable_smart_memory
+
+if DISABLE_SMART_MEMORY:
+    print("Disabling smart memory management")
 
 def get_torch_device_name(device):
     if hasattr(device, 'type'):
@@ -221,132 +226,164 @@ except:
     print("Could not pick default device.")
 
 
-current_loaded_model = None
+current_loaded_models = []
-current_gpu_controlnets = []
 
-model_accelerated = False
+class LoadedModel:
+    def __init__(self, model):
+        self.model = model
+        self.model_accelerated = False
+        self.device = model.load_device
 
+    def model_memory(self):
+        return self.model.model_size()
 
-def unload_model():
+    def model_memory_required(self, device):
-    global current_loaded_model
+        if device == self.model.current_device:
-    global model_accelerated
+            return 0
-    global current_gpu_controlnets
+        else:
-    global vram_state
+            return self.model_memory()
 
-    if current_loaded_model is not None:
+    def model_load(self, lowvram_model_memory=0):
-        if model_accelerated:
+        patch_model_to = None
-            accelerate.hooks.remove_hook_from_submodules(current_loaded_model.model)
+        if lowvram_model_memory == 0:
-            model_accelerated = False
+            patch_model_to = self.device
 
-        current_loaded_model.unpatch_model()
+        self.model.model_patches_to(self.device)
-        current_loaded_model.model.to(current_loaded_model.offload_device)
+        self.model.model_patches_to(self.model.model_dtype())
-        current_loaded_model.model_patches_to(current_loaded_model.offload_device)
-        current_loaded_model = None
-        if vram_state != VRAMState.HIGH_VRAM:
-            soft_empty_cache()
 
-    if vram_state != VRAMState.HIGH_VRAM:
+        try:
-        if len(current_gpu_controlnets) > 0:
+            self.real_model = self.model.patch_model(device_to=patch_model_to) #TODO: do something with loras and offloading to CPU
-            for n in current_gpu_controlnets:
+        except Exception as e:
-                n.cpu()
+            self.model.unpatch_model(self.model.offload_device)
-            current_gpu_controlnets = []
+            self.model_unload()
+            raise e
+
+        if lowvram_model_memory > 0:
+            print("loading in lowvram mode", lowvram_model_memory/(1024 * 1024))
+            device_map = accelerate.infer_auto_device_map(self.real_model, max_memory={0: "{}MiB".format(lowvram_model_memory // (1024 * 1024)), "cpu": "16GiB"})
+            accelerate.dispatch_model(self.real_model, device_map=device_map, main_device=self.device)
+            self.model_accelerated = True
+
+        return self.real_model
+
+    def model_unload(self):
+        if self.model_accelerated:
+            accelerate.hooks.remove_hook_from_submodules(self.real_model)
+            self.model_accelerated = False
+
+        self.model.unpatch_model(self.model.offload_device)
+        self.model.model_patches_to(self.model.offload_device)
+
+    def __eq__(self, other):
+        return self.model is other.model
 
 def minimum_inference_memory():
-    return (768 * 1024 * 1024)
+    return (1024 * 1024 * 1024)
+
+def unload_model_clones(model):
+    to_unload = []
+    for i in range(len(current_loaded_models)):
+        if model.is_clone(current_loaded_models[i].model):
+            to_unload = [i] + to_unload
+
+    for i in to_unload:
+        print("unload clone", i)
+        current_loaded_models.pop(i).model_unload()
+
+def free_memory(memory_required, device, keep_loaded=[]):
+    unloaded_model = False
+    for i in range(len(current_loaded_models) -1, -1, -1):
+        if DISABLE_SMART_MEMORY:
+            current_free_mem = 0
+        else:
+            current_free_mem = get_free_memory(device)
+        if current_free_mem > memory_required:
+            break
+        shift_model = current_loaded_models[i]
+        if shift_model.device == device:
+            if shift_model not in keep_loaded:
+                current_loaded_models.pop(i).model_unload()
+                unloaded_model = True
+
+    if unloaded_model:
+        soft_empty_cache()
+
+
+def load_models_gpu(models, memory_required=0):
+    global vram_state
+
+    inference_memory = minimum_inference_memory()
+    extra_mem = max(inference_memory, memory_required)
+
+    models_to_load = []
+    models_already_loaded = []
+    for x in models:
+        loaded_model = LoadedModel(x)
+
+        if loaded_model in current_loaded_models:
+            index = current_loaded_models.index(loaded_model)
+            current_loaded_models.insert(0, current_loaded_models.pop(index))
+            models_already_loaded.append(loaded_model)
+        else:
+            models_to_load.append(loaded_model)
+
+    if len(models_to_load) == 0:
+        devs = set(map(lambda a: a.device, models_already_loaded))
+        for d in devs:
+            if d != torch.device("cpu"):
+                free_memory(extra_mem, d, models_already_loaded)
+        return
+
+    print("loading new")
+
+    total_memory_required = {}
+    for loaded_model in models_to_load:
+        unload_model_clones(loaded_model.model)
+        total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
+
+    for device in total_memory_required:
+        if device != torch.device("cpu"):
+            free_memory(total_memory_required[device] * 1.3 + extra_mem, device, models_already_loaded)
+
+    for loaded_model in models_to_load:
+        model = loaded_model.model
+        torch_dev = model.load_device
+        if is_device_cpu(torch_dev):
+            vram_set_state = VRAMState.DISABLED
+        else:
+            vram_set_state = vram_state
+        lowvram_model_memory = 0
+        if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
+            model_size = loaded_model.model_memory_required(torch_dev)
+            current_free_mem = get_free_memory(torch_dev)
+            lowvram_model_memory = int(max(256 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 ))
+            if model_size > (current_free_mem - inference_memory): #only switch to lowvram if really necessary
+                vram_set_state = VRAMState.LOW_VRAM
+            else:
+                lowvram_model_memory = 0
+
+        if vram_set_state == VRAMState.NO_VRAM:
+            lowvram_model_memory = 256 * 1024 * 1024
+
+        cur_loaded_model = loaded_model.model_load(lowvram_model_memory)
+        current_loaded_models.insert(0, loaded_model)
+    return
+
 
 def load_model_gpu(model):
-    global current_loaded_model
+    return load_models_gpu([model])
-    global vram_state
-    global model_accelerated
 
-    if model is current_loaded_model:
+def cleanup_models():
-        return
+    to_delete = []
-    unload_model()
+    for i in range(len(current_loaded_models)):
+        print(sys.getrefcount(current_loaded_models[i].model))
+        if sys.getrefcount(current_loaded_models[i].model) <= 2:
+            to_delete = [i] + to_delete
 
-    torch_dev = model.load_device
+    for i in to_delete:
-    model.model_patches_to(torch_dev)
+        x = current_loaded_models.pop(i)
-    model.model_patches_to(model.model_dtype())
+        x.model_unload()
-    current_loaded_model = model
+        del x
-
-    if is_device_cpu(torch_dev):
-        vram_set_state = VRAMState.DISABLED
-    else:
-        vram_set_state = vram_state
-
-    if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
-        model_size = model.model_size()
-        current_free_mem = get_free_memory(torch_dev)
-        lowvram_model_memory = int(max(256 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 ))
-        if model_size > (current_free_mem - minimum_inference_memory()): #only switch to lowvram if really necessary
-            vram_set_state = VRAMState.LOW_VRAM
-
-    real_model = model.model
-    patch_model_to = None
-    if vram_set_state == VRAMState.DISABLED:
-        pass
-    elif vram_set_state == VRAMState.NORMAL_VRAM or vram_set_state == VRAMState.HIGH_VRAM or vram_set_state == VRAMState.SHARED:
-        model_accelerated = False
-        patch_model_to = torch_dev
-
-    try:
-        real_model = model.patch_model(device_to=patch_model_to)
-    except Exception as e:
-        model.unpatch_model()
-        unload_model()
-        raise e
-
-    if patch_model_to is not None:
-        real_model.to(torch_dev)
-
-    if vram_set_state == VRAMState.NO_VRAM:
-        device_map = accelerate.infer_auto_device_map(real_model, max_memory={0: "256MiB", "cpu": "16GiB"})
-        accelerate.dispatch_model(real_model, device_map=device_map, main_device=torch_dev)
-        model_accelerated = True
-    elif vram_set_state == VRAMState.LOW_VRAM:
-        device_map = accelerate.infer_auto_device_map(real_model, max_memory={0: "{}MiB".format(lowvram_model_memory // (1024 * 1024)), "cpu": "16GiB"})
-        accelerate.dispatch_model(real_model, device_map=device_map, main_device=torch_dev)
-        model_accelerated = True
-
-    return current_loaded_model
-
-def load_controlnet_gpu(control_models):
-    global current_gpu_controlnets
-    global vram_state
-    if vram_state == VRAMState.DISABLED:
-        return
-
-    if vram_state == VRAMState.LOW_VRAM or vram_state == VRAMState.NO_VRAM:
-        for m in control_models:
-            if hasattr(m, 'set_lowvram'):
-                m.set_lowvram(True)
-        #don't load controlnets like this if low vram because they will be loaded right before running and unloaded right after
-        return
-
-    models = []
-    for m in control_models:
-        models += m.get_models()
-
-    for m in current_gpu_controlnets:
-        if m not in models:
-            m.cpu()
-
-    device = get_torch_device()
-    current_gpu_controlnets = []
-    for m in models:
-        current_gpu_controlnets.append(m.to(device))
-
-
-def load_if_low_vram(model):
-    global vram_state
-    if vram_state == VRAMState.LOW_VRAM or vram_state == VRAMState.NO_VRAM:
-        return model.to(get_torch_device())
-    return model
-
-def unload_if_low_vram(model):
-    global vram_state
-    if vram_state == VRAMState.LOW_VRAM or vram_state == VRAMState.NO_VRAM:
-        return model.cpu()
-    return model
 
 def unet_offload_device():
     if vram_state == VRAMState.HIGH_VRAM:
@@ -354,6 +391,25 @@ def unet_offload_device():
     else:
         return torch.device("cpu")
 
+def unet_inital_load_device(parameters, dtype):
+    torch_dev = get_torch_device()
+    if vram_state == VRAMState.HIGH_VRAM:
+        return torch_dev
+
+    cpu_dev = torch.device("cpu")
+    dtype_size = 4
+    if dtype == torch.float16 or dtype == torch.bfloat16:
+        dtype_size = 2
+
+    model_size = dtype_size * parameters
+
+    mem_dev = get_free_memory(torch_dev)
+    mem_cpu = get_free_memory(cpu_dev)
+    if mem_dev > mem_cpu and model_size < mem_dev:
+        return torch_dev
+    else:
+        return cpu_dev
+
 def text_encoder_offload_device():
     if args.gpu_only:
         return get_torch_device()
@@ -456,6 +512,13 @@ def get_free_memory(dev=None, torch_free_too=False):
     else:
         return mem_free_total
 
+def batch_area_memory(area):
+    if xformers_enabled() or pytorch_attention_flash_attention():
+        #TODO: these formulas are copied from maximum_batch_area below
+        return (area / 20) * (1024 * 1024)
+    else:
+        return (((area * 0.6) / 0.9) + 1024) * (1024 * 1024)
+
 def maximum_batch_area():
     global vram_state
     if vram_state == VRAMState.NO_VRAM:

comfy/sample.py

@@ -51,19 +51,24 @@ def get_models_from_cond(cond, model_type):
             models += [c[1][model_type]]
     return models
 
-def load_additional_models(positive, negative, dtype):
+def get_additional_models(positive, negative):
     """loads additional models in positive and negative conditioning"""
-    control_nets = get_models_from_cond(positive, "control") + get_models_from_cond(negative, "control")
+    control_nets = set(get_models_from_cond(positive, "control") + get_models_from_cond(negative, "control"))
+
+    control_models = []
+    for m in control_nets:
+        control_models += m.get_models()
+
     gligen = get_models_from_cond(positive, "gligen") + get_models_from_cond(negative, "gligen")
-    gligen = [x[1].to(dtype) for x in gligen]
+    gligen = [x[1] for x in gligen]
-    models = control_nets + gligen
+    models = control_models + gligen
-    comfy.model_management.load_controlnet_gpu(models)
     return models
 
 def cleanup_additional_models(models):
     """cleanup additional models that were loaded"""
     for m in models:
-        m.cleanup()
+        if hasattr(m, 'cleanup'):
+            m.cleanup()
 
 def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=1.0, disable_noise=False, start_step=None, last_step=None, force_full_denoise=False, noise_mask=None, sigmas=None, callback=None, disable_pbar=False, seed=None):
     device = comfy.model_management.get_torch_device()
@@ -72,7 +77,8 @@ def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative
         noise_mask = prepare_mask(noise_mask, noise.shape, device)
 
     real_model = None
-    comfy.model_management.load_model_gpu(model)
+    models = get_additional_models(positive, negative)
+    comfy.model_management.load_models_gpu([model] + models, comfy.model_management.batch_area_memory(noise.shape[0] * noise.shape[2] * noise.shape[3]))
     real_model = model.model
 
     noise = noise.to(device)
@@ -81,7 +87,6 @@ def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative
     positive_copy = broadcast_cond(positive, noise.shape[0], device)
     negative_copy = broadcast_cond(negative, noise.shape[0], device)
 
-    models = load_additional_models(positive, negative, model.model_dtype())
 
     sampler = comfy.samplers.KSampler(real_model, steps=steps, device=device, sampler=sampler_name, scheduler=scheduler, denoise=denoise, model_options=model.model_options)
 

comfy/samplers.py

@@ -88,9 +88,9 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, con
                 gligen_type = gligen[0]
                 gligen_model = gligen[1]
                 if gligen_type == "position":
-                    gligen_patch = gligen_model.set_position(input_x.shape, gligen[2], input_x.device)
+                    gligen_patch = gligen_model.model.set_position(input_x.shape, gligen[2], input_x.device)
                 else:
-                    gligen_patch = gligen_model.set_empty(input_x.shape, input_x.device)
+                    gligen_patch = gligen_model.model.set_empty(input_x.shape, input_x.device)
 
                 patches['middle_patch'] = [gligen_patch]
 

comfy/sd.py

@@ -244,7 +244,7 @@ def set_attr(obj, attr, value):
     del prev
 
 class ModelPatcher:
-    def __init__(self, model, load_device, offload_device, size=0):
+    def __init__(self, model, load_device, offload_device, size=0, current_device=None):
         self.size = size
         self.model = model
         self.patches = {}
@@ -253,6 +253,10 @@ class ModelPatcher:
         self.model_size()
         self.load_device = load_device
         self.offload_device = offload_device
+        if current_device is None:
+            self.current_device = self.offload_device
+        else:
+            self.current_device = current_device
 
     def model_size(self):
         if self.size > 0:
@@ -267,7 +271,7 @@ class ModelPatcher:
         return size
 
     def clone(self):
-        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size)
+        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, self.current_device)
         n.patches = {}
         for k in self.patches:
             n.patches[k] = self.patches[k][:]
@@ -276,6 +280,11 @@ class ModelPatcher:
         n.model_keys = self.model_keys
         return n
 
+    def is_clone(self, other):
+        if hasattr(other, 'model') and self.model is other.model:
+            return True
+        return False
+
     def set_model_sampler_cfg_function(self, sampler_cfg_function):
         if len(inspect.signature(sampler_cfg_function).parameters) == 3:
             self.model_options["sampler_cfg_function"] = lambda args: sampler_cfg_function(args["cond"], args["uncond"], args["cond_scale"]) #Old way
@@ -390,6 +399,11 @@ class ModelPatcher:
             out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype)
             set_attr(self.model, key, out_weight)
             del temp_weight
+
+        if device_to is not None:
+            self.model.to(device_to)
+            self.current_device = device_to
+
         return self.model
 
     def calculate_weight(self, patches, weight, key):
@@ -482,7 +496,7 @@ class ModelPatcher:
 
         return weight
 
-    def unpatch_model(self):
+    def unpatch_model(self, device_to=None):
         keys = list(self.backup.keys())
 
         for k in keys:
@@ -490,6 +504,11 @@ class ModelPatcher:
 
         self.backup = {}
 
+        if device_to is not None:
+            self.model.to(device_to)
+            self.current_device = device_to
+
+
 def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
     key_map = model_lora_keys_unet(model.model)
     key_map = model_lora_keys_clip(clip.cond_stage_model, key_map)
@@ -555,7 +574,7 @@ class CLIP:
         else:
             self.cond_stage_model.reset_clip_layer()
 
-        model_management.load_model_gpu(self.patcher)
+        self.load_model()
         cond, pooled = self.cond_stage_model.encode_token_weights(tokens)
         if return_pooled:
             return cond, pooled
@@ -571,11 +590,9 @@ class CLIP:
     def get_sd(self):
         return self.cond_stage_model.state_dict()
 
-    def patch_model(self):
+    def load_model(self):
-        self.patcher.patch_model()
+        model_management.load_model_gpu(self.patcher)
-
+        return self.patcher
-    def unpatch_model(self):
-        self.patcher.unpatch_model()
 
     def get_key_patches(self):
         return self.patcher.get_key_patches()
@@ -630,11 +647,12 @@ class VAE:
         return samples
 
     def decode(self, samples_in):
-        model_management.unload_model()
         self.first_stage_model = self.first_stage_model.to(self.device)
         try:
+            memory_used = (2562 * samples_in.shape[2] * samples_in.shape[3] * 64) * 1.4
+            model_management.free_memory(memory_used, self.device)
             free_memory = model_management.get_free_memory(self.device)
-            batch_number = int((free_memory * 0.7) / (2562 * samples_in.shape[2] * samples_in.shape[3] * 64))
+            batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
 
             pixel_samples = torch.empty((samples_in.shape[0], 3, round(samples_in.shape[2] * 8), round(samples_in.shape[3] * 8)), device="cpu")
@@ -650,19 +668,19 @@ class VAE:
         return pixel_samples
 
     def decode_tiled(self, samples, tile_x=64, tile_y=64, overlap = 16):
-        model_management.unload_model()
         self.first_stage_model = self.first_stage_model.to(self.device)
         output = self.decode_tiled_(samples, tile_x, tile_y, overlap)
         self.first_stage_model = self.first_stage_model.to(self.offload_device)
         return output.movedim(1,-1)
 
     def encode(self, pixel_samples):
-        model_management.unload_model()
         self.first_stage_model = self.first_stage_model.to(self.device)
         pixel_samples = pixel_samples.movedim(-1,1)
         try:
+            memory_used = (2078 * pixel_samples.shape[2] * pixel_samples.shape[3]) * 1.4 #NOTE: this constant along with the one in the decode above are estimated from the mem usage for the VAE and could change.
+            model_management.free_memory(memory_used, self.device)
             free_memory = model_management.get_free_memory(self.device)
-            batch_number = int((free_memory * 0.7) / (2078 * pixel_samples.shape[2] * pixel_samples.shape[3])) #NOTE: this constant along with the one in the decode above are estimated from the mem usage for the VAE and could change.
+            batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
             samples = torch.empty((pixel_samples.shape[0], 4, round(pixel_samples.shape[2] // 8), round(pixel_samples.shape[3] // 8)), device="cpu")
             for x in range(0, pixel_samples.shape[0], batch_number):
@@ -677,7 +695,6 @@ class VAE:
         return samples
 
     def encode_tiled(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
-        model_management.unload_model()
         self.first_stage_model = self.first_stage_model.to(self.device)
         pixel_samples = pixel_samples.movedim(-1,1)
         samples = self.encode_tiled_(pixel_samples, tile_x=tile_x, tile_y=tile_y, overlap=overlap)
@@ -757,6 +774,7 @@ class ControlNet(ControlBase):
     def __init__(self, control_model, global_average_pooling=False, device=None):
         super().__init__(device)
         self.control_model = control_model
+        self.control_model_wrapped = ModelPatcher(self.control_model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
         self.global_average_pooling = global_average_pooling
 
     def get_control(self, x_noisy, t, cond, batched_number):
@@ -786,11 +804,9 @@ class ControlNet(ControlBase):
             precision_scope = contextlib.nullcontext
 
         with precision_scope(model_management.get_autocast_device(self.device)):
-            self.control_model = model_management.load_if_low_vram(self.control_model)
             context = torch.cat(cond['c_crossattn'], 1)
             y = cond.get('c_adm', None)
             control = self.control_model(x=x_noisy, hint=self.cond_hint, timesteps=t, context=context, y=y)
-            self.control_model = model_management.unload_if_low_vram(self.control_model)
         out = {'middle':[], 'output': []}
         autocast_enabled = torch.is_autocast_enabled()
 
@@ -825,7 +841,7 @@ class ControlNet(ControlBase):
 
     def get_models(self):
         out = super().get_models()
-        out.append(self.control_model)
+        out.append(self.control_model_wrapped)
         return out
 
 
@@ -835,7 +851,7 @@ def load_controlnet(ckpt_path, model=None):
     controlnet_config = None
     if "controlnet_cond_embedding.conv_in.weight" in controlnet_data: #diffusers format
         use_fp16 = model_management.should_use_fp16()
-        controlnet_config = model_detection.model_config_from_diffusers_unet(controlnet_data, use_fp16).unet_config
+        controlnet_config = model_detection.unet_config_from_diffusers_unet(controlnet_data, use_fp16)
         diffusers_keys = utils.unet_to_diffusers(controlnet_config)
         diffusers_keys["controlnet_mid_block.weight"] = "middle_block_out.0.weight"
         diffusers_keys["controlnet_mid_block.bias"] = "middle_block_out.0.bias"
@@ -874,6 +890,9 @@ def load_controlnet(ckpt_path, model=None):
             if k in controlnet_data:
                 new_sd[diffusers_keys[k]] = controlnet_data.pop(k)
 
+        leftover_keys = controlnet_data.keys()
+        if len(leftover_keys) > 0:
+            print("leftover keys:", leftover_keys)
         controlnet_data = new_sd
 
     pth_key = 'control_model.zero_convs.0.0.weight'
@@ -901,8 +920,8 @@ def load_controlnet(ckpt_path, model=None):
     if pth:
         if 'difference' in controlnet_data:
             if model is not None:
-                m = model.patch_model()
+                model_management.load_models_gpu([model])
-                model_sd = m.state_dict()
+                model_sd = model.model_state_dict()
                 for x in controlnet_data:
                     c_m = "control_model."
                     if x.startswith(c_m):
@@ -910,7 +929,6 @@ def load_controlnet(ckpt_path, model=None):
                         if sd_key in model_sd:
                             cd = controlnet_data[x]
                             cd += model_sd[sd_key].type(cd.dtype).to(cd.device)
-                model.unpatch_model()
             else:
                 print("WARNING: Loaded a diff controlnet without a model. It will very likely not work.")
 
@@ -1001,7 +1019,6 @@ class T2IAdapter(ControlBase):
         self.copy_to(c)
         return c
 
-
 def load_t2i_adapter(t2i_data):
     keys = t2i_data.keys()
     if 'adapter' in keys:
@@ -1087,7 +1104,7 @@ def load_gligen(ckpt_path):
     model = gligen.load_gligen(data)
     if model_management.should_use_fp16():
         model = model.half()
-    return model
+    return ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
 
 def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_clip=True, embedding_directory=None, state_dict=None, config=None):
     #TODO: this function is a mess and should be removed eventually
@@ -1199,8 +1216,13 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
         if output_clipvision:
             clipvision = clip_vision.load_clipvision_from_sd(sd, model_config.clip_vision_prefix, True)
 
+    dtype = torch.float32
+    if fp16:
+        dtype = torch.float16
+
+    inital_load_device = model_management.unet_inital_load_device(parameters, dtype)
     offload_device = model_management.unet_offload_device()
-    model = model_config.get_model(sd, "model.diffusion_model.", device=offload_device)
+    model = model_config.get_model(sd, "model.diffusion_model.", device=inital_load_device)
     model.load_model_weights(sd, "model.diffusion_model.")
 
     if output_vae:
@@ -1221,7 +1243,12 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
     if len(left_over) > 0:
         print("left over keys:", left_over)
 
-    return (ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device), clip, vae, clipvision)
+    model_patcher = ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device(), current_device=inital_load_device)
+    if inital_load_device != torch.device("cpu"):
+        print("loaded straight to GPU")
+        model_management.load_model_gpu(model_patcher)
+
+    return (model_patcher, clip, vae, clipvision)
 
 
 def load_unet(unet_path): #load unet in diffusers format
@@ -1249,14 +1276,6 @@ def load_unet(unet_path): #load unet in diffusers format
     return ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=offload_device)
 
 def save_checkpoint(output_path, model, clip, vae, metadata=None):
-    try:
+    model_management.load_models_gpu([model, clip.load_model()])
-        model.patch_model()
+    sd = model.model.state_dict_for_saving(clip.get_sd(), vae.get_sd())
-        clip.patch_model()
+    utils.save_torch_file(sd, output_path, metadata=metadata)
-        sd = model.model.state_dict_for_saving(clip.get_sd(), vae.get_sd())
-        utils.save_torch_file(sd, output_path, metadata=metadata)
-        model.unpatch_model()
-        clip.unpatch_model()
-    except Exception as e:
-        model.unpatch_model()
-        clip.unpatch_model()
-        raise e

comfy_extras/nodes_mask.py

@@ -1,15 +1,19 @@
+import numpy as np
+from scipy.ndimage import grey_dilation
 import torch
 
 from nodes import MAX_RESOLUTION
 
-def composite(destination, source, x, y, mask = None, multiplier = 8):
+def composite(destination, source, x, y, mask = None, multiplier = 8, resize_source = False):
+    if resize_source:
+        source = torch.nn.functional.interpolate(source, size=(destination.shape[2], destination.shape[3]), mode="bilinear")
+
     x = max(-source.shape[3] * multiplier, min(x, destination.shape[3] * multiplier))
     y = max(-source.shape[2] * multiplier, min(y, destination.shape[2] * multiplier))
 
     left, top = (x // multiplier, y // multiplier)
     right, bottom = (left + source.shape[3], top + source.shape[2],)
 
-
     if mask is None:
         mask = torch.ones_like(source)
     else:
@@ -40,6 +44,7 @@ class LatentCompositeMasked:
                 "source": ("LATENT",),
                 "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}),
                 "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}),
+                "resize_source": ("BOOLEAN", {"default": False}),
             },
             "optional": {
                 "mask": ("MASK",),
@@ -50,11 +55,11 @@ class LatentCompositeMasked:
 
     CATEGORY = "latent"
 
-    def composite(self, destination, source, x, y, mask = None):
+    def composite(self, destination, source, x, y, resize_source, mask = None):
         output = destination.copy()
         destination = destination["samples"].clone()
         source = source["samples"]
-        output["samples"] = composite(destination, source, x, y, mask, 8)
+        output["samples"] = composite(destination, source, x, y, mask, 8, resize_source)
         return (output,)
 
 class ImageCompositeMasked:
@@ -66,6 +71,7 @@ class ImageCompositeMasked:
                 "source": ("IMAGE",),
                 "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
                 "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "resize_source": ("BOOLEAN", {"default": False}),
             },
             "optional": {
                 "mask": ("MASK",),
@@ -76,9 +82,9 @@ class ImageCompositeMasked:
 
     CATEGORY = "image"
 
-    def composite(self, destination, source, x, y, mask = None):
+    def composite(self, destination, source, x, y, resize_source, mask = None):
         destination = destination.clone().movedim(-1, 1)
-        output = composite(destination, source.movedim(-1, 1), x, y, mask, 1).movedim(1, -1)
+        output = composite(destination, source.movedim(-1, 1), x, y, mask, 1, resize_source).movedim(1, -1)
         return (output,)
 
 class MaskToImage:
@@ -272,6 +278,35 @@ class FeatherMask:
             output[-y, :] *= feather_rate
 
         return (output,)
+    
+class GrowMask:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "mask": ("MASK",),
+                "expand": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}),
+                "tapered_corners": ("BOOLEAN", {"default": True}),
+            },
+        }
+    
+    CATEGORY = "mask"
+
+    RETURN_TYPES = ("MASK",)
+
+    FUNCTION = "expand_mask"
+
+    def expand_mask(self, mask, expand, tapered_corners):
+        c = 0 if tapered_corners else 1
+        kernel = np.array([[c, 1, c],
+                           [1, 1, 1],
+                           [c, 1, c]])
+        output = mask.numpy().copy()
+        while expand > 0:
+            output = grey_dilation(output, footprint=kernel)
+            expand -= 1
+        output = torch.from_numpy(output)
+        return (output,)
 
 
 
@@ -285,6 +320,7 @@ NODE_CLASS_MAPPINGS = {
     "CropMask": CropMask,
     "MaskComposite": MaskComposite,
     "FeatherMask": FeatherMask,
+    "GrowMask": GrowMask,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {

comfy_extras/nodes_post_processing.py

@@ -2,6 +2,7 @@ import numpy as np
 import torch
 import torch.nn.functional as F
 from PIL import Image
+import math
 
 import comfy.utils
 
@@ -209,9 +210,36 @@ class Sharpen:
 
         return (result,)
 
+class ImageScaleToTotalPixels:
+    upscale_methods = ["nearest-exact", "bilinear", "area", "bicubic"]
+    crop_methods = ["disabled", "center"]
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": ("IMAGE",), "upscale_method": (s.upscale_methods,),
+                              "megapixels": ("FLOAT", {"default": 1.0, "min": 0.01, "max": 16.0, "step": 0.01}),
+                            }}
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "upscale"
+
+    CATEGORY = "image/upscaling"
+
+    def upscale(self, image, upscale_method, megapixels):
+        samples = image.movedim(-1,1)
+        total = int(megapixels * 1024 * 1024)
+
+        scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+        width = round(samples.shape[3] * scale_by)
+        height = round(samples.shape[2] * scale_by)
+
+        s = comfy.utils.common_upscale(samples, width, height, upscale_method, "disabled")
+        s = s.movedim(1,-1)
+        return (s,)
+
 NODE_CLASS_MAPPINGS = {
     "ImageBlend": Blend,
     "ImageBlur": Blur,
     "ImageQuantize": Quantize,
     "ImageSharpen": Sharpen,
+    "ImageScaleToTotalPixels": ImageScaleToTotalPixels,
 }

execution.py

@@ -354,6 +354,7 @@ class PromptExecutor:
                     d = self.outputs_ui.pop(x)
                     del d
 
+            comfy.model_management.cleanup_models()
             if self.server.client_id is not None:
                 self.server.send_sync("execution_cached", { "nodes": list(current_outputs) , "prompt_id": prompt_id}, self.server.client_id)
             executed = set()

nodes.py

@@ -1465,6 +1465,28 @@ class ImageBatch:
         s = torch.cat((image1, image2), dim=0)
         return (s,)
 
+class EmptyImage:
+    def __init__(self, device="cpu"):
+        self.device = device
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}),
+                              "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}),
+                              "batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
+                              "color": ("INT", {"default": 0, "min": 0, "max": 0xFFFFFF, "step": 1, "display": "color"}),
+                              }}
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "generate"
+
+    CATEGORY = "image"
+
+    def generate(self, width, height, batch_size=1, color=0):
+        r = torch.full([batch_size, height, width, 1], ((color >> 16) & 0xFF) / 0xFF)
+        g = torch.full([batch_size, height, width, 1], ((color >> 8) & 0xFF) / 0xFF)
+        b = torch.full([batch_size, height, width, 1], ((color) & 0xFF) / 0xFF)
+        return (torch.cat((r, g, b), dim=-1), )
+
 class ImagePadForOutpaint:
 
     @classmethod
@@ -1551,6 +1573,7 @@ NODE_CLASS_MAPPINGS = {
     "ImageInvert": ImageInvert,
     "ImageBatch": ImageBatch,
     "ImagePadForOutpaint": ImagePadForOutpaint,
+    "EmptyImage": EmptyImage,
     "ConditioningAverage ": ConditioningAverage ,
     "ConditioningCombine": ConditioningCombine,
     "ConditioningConcat": ConditioningConcat,

web/index.html

@@ -6,6 +6,7 @@
 		<meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=no">
 		<link rel="stylesheet" type="text/css" href="./lib/litegraph.css" />
 		<link rel="stylesheet" type="text/css" href="./style.css" />
+		<link rel="stylesheet" type="text/css" href="./user.css" />
 		<script type="text/javascript" src="./lib/litegraph.core.js"></script>
 		<script type="text/javascript" src="./lib/litegraph.extensions.js" defer></script>
 		<script type="module">

web/scripts/app.js

@@ -284,6 +284,11 @@ export class ComfyApp {
 				}
 			}
 
+			options.push({
+					content: "Bypass",
+					callback: (obj) => { if (this.mode === 4) this.mode = 0; else this.mode = 4; this.graph.change(); }
+				});
+
 			// prevent conflict of clipspace content
 			if(!ComfyApp.clipspace_return_node) {
 				options.push({

web/user.css

@@ -0,0 +1 @@
+/* Put custom styles here */