Merge branch 'master' into api-nodes

.github/workflows/stable-release.yml

@@ -36,7 +36,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           ref: ${{ inputs.git_tag }}
-          fetch-depth: 0
+          fetch-depth: 150
           persist-credentials: false
       - uses: actions/cache/restore@v4
         id: cache
@@ -70,7 +70,7 @@ jobs:
             cd ..
 
           git clone --depth 1 https://github.com/comfyanonymous/taesd
-          cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+          cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
           mkdir ComfyUI_windows_portable
           mv python_embeded ComfyUI_windows_portable
@@ -85,7 +85,7 @@ jobs:
 
           cd ..
 
-          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=8 -mfb=64 -md=32m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
           mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_nvidia.7z
 
           cd ComfyUI_windows_portable

.github/workflows/windows_release_nightly_pytorch.yml

@@ -56,7 +56,7 @@ jobs:
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd
-            cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+            cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
             mkdir ComfyUI_windows_portable_nightly_pytorch
             mv python_embeded ComfyUI_windows_portable_nightly_pytorch

.github/workflows/windows_release_package.yml

@@ -50,7 +50,7 @@ jobs:
 
         - uses: actions/checkout@v4
           with:
-            fetch-depth: 0
+            fetch-depth: 150
             persist-credentials: false
         - shell: bash
           run: |
@@ -67,7 +67,7 @@ jobs:
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd
-            cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+            cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
             mkdir ComfyUI_windows_portable
             mv python_embeded ComfyUI_windows_portable
@@ -82,7 +82,7 @@ jobs:
 
             cd ..
 
-            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=8 -mfb=64 -md=32m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
             mv ComfyUI_windows_portable.7z ComfyUI/new_ComfyUI_windows_portable_nvidia_cu${{ inputs.cu }}_or_cpu.7z
 
             cd ComfyUI_windows_portable

README.md

@@ -216,9 +216,9 @@ Additional discussion and help can be found [here](https://github.com/comfyanony
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu126```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128```
 
-This is the command to install pytorch nightly instead which supports the new blackwell 50xx series GPUs and might have performance improvements.
+This is the command to install pytorch nightly instead which might have performance improvements.
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128```
 

comfy/cli_args.py

@@ -66,6 +66,7 @@ fpunet_group.add_argument("--bf16-unet", action="store_true", help="Run the diff
 fpunet_group.add_argument("--fp16-unet", action="store_true", help="Run the diffusion model in fp16")
 fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.")
 fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.")
+fpunet_group.add_argument("--fp8_e8m0fnu-unet", action="store_true", help="Store unet weights in fp8_e8m0fnu.")
 
 fpvae_group = parser.add_mutually_exclusive_group()
 fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.")

comfy/image_encoders/dino2.py

@@ -116,7 +116,7 @@ class Dino2Embeddings(torch.nn.Module):
     def forward(self, pixel_values):
         x = self.patch_embeddings(pixel_values)
         # TODO: mask_token?
-        x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
+        x = torch.cat((self.cls_token.to(device=x.device, dtype=x.dtype).expand(x.shape[0], -1, -1), x), dim=1)
         x = x + comfy.model_management.cast_to_device(self.position_embeddings, x.device, x.dtype)
         return x
 

comfy/model_management.py

@@ -725,6 +725,8 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e4m3fn
     if args.fp8_e5m2_unet:
         return torch.float8_e5m2
+    if args.fp8_e8m0fnu_unet:
+        return torch.float8_e8m0fnu
 
     fp8_dtype = None
     if weight_dtype in FLOAT8_TYPES:

comfy/weight_adapter/__init__.py

@@ -3,6 +3,8 @@ from .lora import LoRAAdapter
 from .loha import LoHaAdapter
 from .lokr import LoKrAdapter
 from .glora import GLoRAAdapter
+from .oft import OFTAdapter
+from .boft import BOFTAdapter
 
 
 adapters: list[type[WeightAdapterBase]] = [
@@ -10,4 +12,6 @@ adapters: list[type[WeightAdapterBase]] = [
     LoHaAdapter,
     LoKrAdapter,
     GLoRAAdapter,
+    OFTAdapter,
+    BOFTAdapter,
 ]

comfy/weight_adapter/base.py

@@ -43,13 +43,23 @@ def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediat
     dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype)
     lora_diff *= alpha
     weight_calc = weight + function(lora_diff).type(weight.dtype)
-    weight_norm = (
-        weight_calc.transpose(0, 1)
-        .reshape(weight_calc.shape[1], -1)
-        .norm(dim=1, keepdim=True)
-        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
-        .transpose(0, 1)
-    )
+
+    wd_on_output_axis = dora_scale.shape[0] == weight_calc.shape[0]
+    if wd_on_output_axis:
+        weight_norm = (
+            weight.reshape(weight.shape[0], -1)
+            .norm(dim=1, keepdim=True)
+            .reshape(weight.shape[0], *[1] * (weight.dim() - 1))
+        )
+    else:
+        weight_norm = (
+            weight_calc.transpose(0, 1)
+            .reshape(weight_calc.shape[1], -1)
+            .norm(dim=1, keepdim=True)
+            .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
+            .transpose(0, 1)
+        )
+    weight_norm = weight_norm + torch.finfo(weight.dtype).eps
 
     weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
     if strength != 1.0:

comfy/weight_adapter/boft.py

@@ -0,0 +1,115 @@
+import logging
+from typing import Optional
+
+import torch
+import comfy.model_management
+from .base import WeightAdapterBase, weight_decompose
+
+
+class BOFTAdapter(WeightAdapterBase):
+    name = "boft"
+
+    def __init__(self, loaded_keys, weights):
+        self.loaded_keys = loaded_keys
+        self.weights = weights
+
+    @classmethod
+    def load(
+        cls,
+        x: str,
+        lora: dict[str, torch.Tensor],
+        alpha: float,
+        dora_scale: torch.Tensor,
+        loaded_keys: set[str] = None,
+    ) -> Optional["BOFTAdapter"]:
+        if loaded_keys is None:
+            loaded_keys = set()
+        blocks_name = "{}.boft_blocks".format(x)
+        rescale_name = "{}.rescale".format(x)
+
+        blocks = None
+        if blocks_name in lora.keys():
+            blocks = lora[blocks_name]
+            if blocks.ndim == 4:
+                loaded_keys.add(blocks_name)
+
+        rescale = None
+        if rescale_name in lora.keys():
+            rescale = lora[rescale_name]
+            loaded_keys.add(rescale_name)
+
+        if blocks is not None:
+            weights = (blocks, rescale, alpha, dora_scale)
+            return cls(loaded_keys, weights)
+        else:
+            return None
+
+    def calculate_weight(
+        self,
+        weight,
+        key,
+        strength,
+        strength_model,
+        offset,
+        function,
+        intermediate_dtype=torch.float32,
+        original_weight=None,
+    ):
+        v = self.weights
+        blocks = v[0]
+        rescale = v[1]
+        alpha = v[2]
+        dora_scale = v[3]
+
+        blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype)
+        if rescale is not None:
+            rescale = comfy.model_management.cast_to_device(rescale, weight.device, intermediate_dtype)
+
+        boft_m, block_num, boft_b, *_ = blocks.shape
+
+        try:
+            # Get r
+            I = torch.eye(boft_b, device=blocks.device, dtype=blocks.dtype)
+            # for Q = -Q^T
+            q = blocks - blocks.transpose(1, 2)
+            normed_q = q
+            if alpha > 0: # alpha in boft/bboft is for constraint
+                q_norm = torch.norm(q) + 1e-8
+                if q_norm > alpha:
+                    normed_q = q * alpha / q_norm
+            # use float() to prevent unsupported type in .inverse()
+            r = (I + normed_q) @ (I - normed_q).float().inverse()
+            r = r.to(original_weight)
+
+            inp = org = original_weight
+
+            r_b = boft_b//2
+            for i in range(boft_m):
+                bi = r[i]
+                g = 2
+                k = 2**i * r_b
+                if strength != 1:
+                    bi = bi * strength + (1-strength) * I
+                inp = (
+                    inp.unflatten(-1, (-1, g, k))
+                    .transpose(-2, -1)
+                    .flatten(-3)
+                    .unflatten(-1, (-1, boft_b))
+                )
+                inp = torch.einsum("b n m, b n ... -> b m ...", inp, bi)
+                inp = (
+                    inp.flatten(-2).unflatten(-1, (-1, k, g)).transpose(-2, -1).flatten(-3)
+                )
+
+            if rescale is not None:
+                inp = inp * rescale
+
+            lora_diff = inp - org
+            lora_diff = comfy.model_management.cast_to_device(lora_diff, weight.device, intermediate_dtype)
+            if dora_scale is not None:
+                weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
+            else:
+                weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+        except Exception as e:
+            logging.error("ERROR {} {} {}".format(self.name, key, e))
+        return weight

comfy/weight_adapter/oft.py

@@ -0,0 +1,94 @@
+import logging
+from typing import Optional
+
+import torch
+import comfy.model_management
+from .base import WeightAdapterBase, weight_decompose
+
+
+class OFTAdapter(WeightAdapterBase):
+    name = "oft"
+
+    def __init__(self, loaded_keys, weights):
+        self.loaded_keys = loaded_keys
+        self.weights = weights
+
+    @classmethod
+    def load(
+        cls,
+        x: str,
+        lora: dict[str, torch.Tensor],
+        alpha: float,
+        dora_scale: torch.Tensor,
+        loaded_keys: set[str] = None,
+    ) -> Optional["OFTAdapter"]:
+        if loaded_keys is None:
+            loaded_keys = set()
+        blocks_name = "{}.oft_blocks".format(x)
+        rescale_name = "{}.rescale".format(x)
+
+        blocks = None
+        if blocks_name in lora.keys():
+            blocks = lora[blocks_name]
+            if blocks.ndim == 3:
+                loaded_keys.add(blocks_name)
+
+        rescale = None
+        if rescale_name in lora.keys():
+            rescale = lora[rescale_name]
+            loaded_keys.add(rescale_name)
+
+        if blocks is not None:
+            weights = (blocks, rescale, alpha, dora_scale)
+            return cls(loaded_keys, weights)
+        else:
+            return None
+
+    def calculate_weight(
+        self,
+        weight,
+        key,
+        strength,
+        strength_model,
+        offset,
+        function,
+        intermediate_dtype=torch.float32,
+        original_weight=None,
+    ):
+        v = self.weights
+        blocks = v[0]
+        rescale = v[1]
+        alpha = v[2]
+        dora_scale = v[3]
+
+        blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype)
+        if rescale is not None:
+            rescale = comfy.model_management.cast_to_device(rescale, weight.device, intermediate_dtype)
+
+        block_num, block_size, *_ = blocks.shape
+
+        try:
+            # Get r
+            I = torch.eye(block_size, device=blocks.device, dtype=blocks.dtype)
+            # for Q = -Q^T
+            q = blocks - blocks.transpose(1, 2)
+            normed_q = q
+            if alpha > 0: # alpha in oft/boft is for constraint
+                q_norm = torch.norm(q) + 1e-8
+                if q_norm > alpha:
+                    normed_q = q * alpha / q_norm
+            # use float() to prevent unsupported type in .inverse()
+            r = (I + normed_q) @ (I - normed_q).float().inverse()
+            r = r.to(original_weight)
+            lora_diff = torch.einsum(
+                "k n m, k n ... -> k m ...",
+                (r * strength) - strength * I,
+                original_weight,
+            )
+            if dora_scale is not None:
+                weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
+            else:
+                weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+        except Exception as e:
+            logging.error("ERROR {} {} {}".format(self.name, key, e))
+        return weight

comfy_extras/nodes_mask.py

@@ -3,7 +3,10 @@ import scipy.ndimage
 import torch
 import comfy.utils
 import node_helpers
+import folder_paths
+import random
 
+import nodes
 from nodes import MAX_RESOLUTION
 
 def composite(destination, source, x, y, mask = None, multiplier = 8, resize_source = False):
@@ -362,6 +365,30 @@ class ThresholdMask:
         mask = (mask > value).float()
         return (mask,)
 
+# Mask Preview - original implement from
+# https://github.com/cubiq/ComfyUI_essentials/blob/9d9f4bedfc9f0321c19faf71855e228c93bd0dc9/mask.py#L81
+# upstream requested in https://github.com/Kosinkadink/rfcs/blob/main/rfcs/0000-corenodes.md#preview-nodes
+class MaskPreview(nodes.SaveImage):
+    def __init__(self):
+        self.output_dir = folder_paths.get_temp_directory()
+        self.type = "temp"
+        self.prefix_append = "_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for x in range(5))
+        self.compress_level = 4
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {"mask": ("MASK",), },
+            "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},
+        }
+
+    FUNCTION = "execute"
+    CATEGORY = "mask"
+
+    def execute(self, mask, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None):
+        preview = mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3)
+        return self.save_images(preview, filename_prefix, prompt, extra_pnginfo)
+
 
 NODE_CLASS_MAPPINGS = {
     "LatentCompositeMasked": LatentCompositeMasked,
@@ -376,6 +403,7 @@ NODE_CLASS_MAPPINGS = {
     "FeatherMask": FeatherMask,
     "GrowMask": GrowMask,
     "ThresholdMask": ThresholdMask,
+    "MaskPreview": MaskPreview
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {

comfy_extras/nodes_video.py

@@ -50,13 +50,15 @@ class SaveWEBM:
             for x in extra_pnginfo:
                 container.metadata[x] = json.dumps(extra_pnginfo[x])
 
-        codec_map = {"vp9": "libvpx-vp9", "av1": "libaom-av1"}
+        codec_map = {"vp9": "libvpx-vp9", "av1": "libsvtav1"}
         stream = container.add_stream(codec_map[codec], rate=Fraction(round(fps * 1000), 1000))
         stream.width = images.shape[-2]
         stream.height = images.shape[-3]
-        stream.pix_fmt = "yuv420p"
+        stream.pix_fmt = "yuv420p10le" if codec == "av1" else "yuv420p"
         stream.bit_rate = 0
         stream.options = {'crf': str(crf)}
+        if codec == "av1":
+            stream.options["preset"] = "6"
 
         for frame in images:
             frame = av.VideoFrame.from_ndarray(torch.clamp(frame[..., :3] * 255, min=0, max=255).to(device=torch.device("cpu"), dtype=torch.uint8).numpy(), format="rgb24")

requirements.txt

@@ -22,5 +22,5 @@ psutil
 kornia>=0.7.1
 spandrel
 soundfile
-av
+av>=14.1.0
 pydantic~=2.0