Merge branch 'master' of https://github.com/comfyanonymous/ComfyUI

.ci/update_windows/update.py

@@ -63,7 +63,12 @@ except:
 print("checking out master branch")  # noqa: T201
 branch = repo.lookup_branch('master')
 if branch is None:
-    ref = repo.lookup_reference('refs/remotes/origin/master')
+    try:
+        ref = repo.lookup_reference('refs/remotes/origin/master')
+    except:
+        print("pulling.")  # noqa: T201
+        pull(repo)
+        ref = repo.lookup_reference('refs/remotes/origin/master')
     repo.checkout(ref)
     branch = repo.lookup_branch('master')
     if branch is None:

comfy/weight_adapter/boft.py

@@ -24,7 +24,7 @@ class BOFTAdapter(WeightAdapterBase):
     ) -> Optional["BOFTAdapter"]:
         if loaded_keys is None:
             loaded_keys = set()
-        blocks_name = "{}.boft_blocks".format(x)
+        blocks_name = "{}.oft_blocks".format(x)
         rescale_name = "{}.rescale".format(x)
 
         blocks = None
@@ -32,17 +32,18 @@ class BOFTAdapter(WeightAdapterBase):
             blocks = lora[blocks_name]
             if blocks.ndim == 4:
                 loaded_keys.add(blocks_name)
+            else:
+                blocks = None
+        if blocks is None:
+            return None
 
         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
+        weights = (blocks, rescale, alpha, dora_scale)
+        return cls(loaded_keys, weights)
 
     def calculate_weight(
         self,
@@ -71,7 +72,7 @@ class BOFTAdapter(WeightAdapterBase):
             # Get r
             I = torch.eye(boft_b, device=blocks.device, dtype=blocks.dtype)
             # for Q = -Q^T
-            q = blocks - blocks.transpose(1, 2)
+            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
@@ -79,9 +80,8 @@ class BOFTAdapter(WeightAdapterBase):
                     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 = r.to(weight)
+            inp = org = weight
 
             r_b = boft_b//2
             for i in range(boft_m):
@@ -91,14 +91,14 @@ class BOFTAdapter(WeightAdapterBase):
                 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.unflatten(0, (-1, g, k))
+                    .transpose(1, 2)
+                    .flatten(0, 2)
+                    .unflatten(0, (-1, boft_b))
                 )
-                inp = torch.einsum("b n m, b n ... -> b m ...", inp, bi)
+                inp = torch.einsum("b i j, b j ...-> b i ...", bi, inp)
                 inp = (
-                    inp.flatten(-2).unflatten(-1, (-1, k, g)).transpose(-2, -1).flatten(-3)
+                    inp.flatten(0, 1).unflatten(0, (-1, k, g)).transpose(1, 2).flatten(0, 2)
                 )
 
             if rescale is not None:
@@ -109,7 +109,7 @@ class BOFTAdapter(WeightAdapterBase):
             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))
+                weight += function((strength * lora_diff).type(weight.dtype))
         except Exception as e:
             logging.error("ERROR {} {} {}".format(self.name, key, e))
         return weight

comfy/weight_adapter/oft.py

@@ -32,17 +32,18 @@ class OFTAdapter(WeightAdapterBase):
             blocks = lora[blocks_name]
             if blocks.ndim == 3:
                 loaded_keys.add(blocks_name)
+            else:
+                blocks = None
+        if blocks is None:
+            return None
 
         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
+        weights = (blocks, rescale, alpha, dora_scale)
+        return cls(loaded_keys, weights)
 
     def calculate_weight(
         self,
@@ -79,16 +80,17 @@ class OFTAdapter(WeightAdapterBase):
                     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)
+            r = r.to(weight)
+            _, *shape = weight.shape
             lora_diff = torch.einsum(
                 "k n m, k n ... -> k m ...",
                 (r * strength) - strength * I,
-                original_weight,
-            )
+                weight.view(block_num, block_size, *shape),
+            ).view(-1, *shape)
             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))
+                weight += function((strength * lora_diff).type(weight.dtype))
         except Exception as e:
             logging.error("ERROR {} {} {}".format(self.name, key, e))
         return weight

comfy_extras/nodes_custom_sampler.py

@@ -1,3 +1,4 @@
+import math
 import comfy.samplers
 import comfy.sample
 from comfy.k_diffusion import sampling as k_diffusion_sampling
@@ -249,6 +250,55 @@ class SetFirstSigma:
         sigmas[0] = sigma
         return (sigmas, )
 
+class ExtendIntermediateSigmas:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"sigmas": ("SIGMAS", ),
+                     "steps": ("INT", {"default": 2, "min": 1, "max": 100}),
+                     "start_at_sigma": ("FLOAT", {"default": -1.0, "min": -1.0, "max": 20000.0, "step": 0.01, "round": False}),
+                     "end_at_sigma": ("FLOAT", {"default": 12.0, "min":  0.0, "max": 20000.0, "step": 0.01, "round": False}),
+                     "spacing": (['linear', 'cosine', 'sine'],),
+                    }
+               }
+    RETURN_TYPES = ("SIGMAS",)
+    CATEGORY = "sampling/custom_sampling/sigmas"
+
+    FUNCTION = "extend"
+
+    def extend(self, sigmas: torch.Tensor, steps: int, start_at_sigma: float, end_at_sigma: float, spacing: str):
+        if start_at_sigma < 0:
+            start_at_sigma = float("inf")
+
+        interpolator = {
+            'linear': lambda x: x,
+            'cosine': lambda x: torch.sin(x*math.pi/2),
+            'sine':   lambda x: 1 - torch.cos(x*math.pi/2)
+        }[spacing]
+
+        # linear space for our interpolation function
+        x = torch.linspace(0, 1, steps + 1, device=sigmas.device)[1:-1]
+        computed_spacing = interpolator(x)
+
+        extended_sigmas = []
+        for i in range(len(sigmas) - 1):
+            sigma_current = sigmas[i]
+            sigma_next = sigmas[i+1]
+
+            extended_sigmas.append(sigma_current)
+
+            if end_at_sigma <= sigma_current <= start_at_sigma:
+                interpolated_steps = computed_spacing * (sigma_next - sigma_current) + sigma_current
+                extended_sigmas.extend(interpolated_steps.tolist())
+
+        # Add the last sigma value
+        if len(sigmas) > 0:
+            extended_sigmas.append(sigmas[-1])
+
+        extended_sigmas = torch.FloatTensor(extended_sigmas)
+
+        return (extended_sigmas,)
+
 class KSamplerSelect:
     @classmethod
     def INPUT_TYPES(s):
@@ -735,6 +785,7 @@ NODE_CLASS_MAPPINGS = {
     "SplitSigmasDenoise": SplitSigmasDenoise,
     "FlipSigmas": FlipSigmas,
     "SetFirstSigma": SetFirstSigma,
+    "ExtendIntermediateSigmas": ExtendIntermediateSigmas,
 
     "CFGGuider": CFGGuider,
     "DualCFGGuider": DualCFGGuider,

comfy_extras/nodes_model_merging.py

@@ -276,7 +276,7 @@ class CLIPSave:
         comfy.model_management.load_models_gpu([clip.load_model()], force_patch_weights=True)
         clip_sd = clip.get_sd()
 
-        for prefix in ["clip_l.", "clip_g.", ""]:
+        for prefix in ["clip_l.", "clip_g.", "clip_h.", "t5xxl.", "pile_t5xl.", "mt5xl.", "umt5xxl.", "t5base.", "gemma2_2b.", "llama.", "hydit_clip.", ""]:
             k = list(filter(lambda a: a.startswith(prefix), clip_sd.keys()))
             current_clip_sd = {}
             for x in k:

comfy_extras/nodes_preview_any.py

@@ -0,0 +1,43 @@
+import json
+from comfy.comfy_types.node_typing import IO
+
+# Preview Any - original implement from
+# https://github.com/rgthree/rgthree-comfy/blob/main/py/display_any.py
+# upstream requested in https://github.com/Kosinkadink/rfcs/blob/main/rfcs/0000-corenodes.md#preview-nodes
+class PreviewAny():
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {"source": (IO.ANY, {})},
+        }
+
+    RETURN_TYPES = ()
+    FUNCTION = "main"
+    OUTPUT_NODE = True
+
+    CATEGORY = "utils"
+
+    def main(self, source=None):
+        value = 'None'
+        if isinstance(source, str):
+            value = source
+        elif isinstance(source, (int, float, bool)):
+            value = str(source)
+        elif source is not None:
+            try:
+                value = json.dumps(source)
+            except Exception:
+                try:
+                    value = str(source)
+                except Exception:
+                    value = 'source exists, but could not be serialized.'
+
+        return {"ui": {"text": (value,)}}
+
+NODE_CLASS_MAPPINGS = {
+    "PreviewAny": PreviewAny,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PreviewAny": "Preview Any",
+}

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.30"
+__version__ = "0.3.31"

main.py

@@ -13,7 +13,7 @@ import logging
 import sys
 
 if __name__ == "__main__":
-    #NOTE: These do not do anything on core ComfyUI which should already have no communication with the internet, they are for custom nodes.
+    #NOTE: These do not do anything on core ComfyUI, they are for custom nodes.
     os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1'
     os.environ['DO_NOT_TRACK'] = '1'
 

nodes.py

@@ -2336,6 +2336,7 @@ def init_builtin_extra_nodes():
         "nodes_optimalsteps.py",
         "nodes_hidream.py",
         "nodes_fresca.py",
+        "nodes_preview_any.py",
     ]
 
     api_nodes_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy_api_nodes")

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.30"
+version = "0.3.31"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.18.5
+comfyui-frontend-package==1.18.6
 comfyui-workflow-templates==0.1.3
 torch
 torchsde