Merge branch 'master' into dr-support-pip-cm

README.md

@@ -255,7 +255,7 @@ This is the command to install the Pytorch xpu nightly which might have some per
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu129```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130```
 
 This is the command to install pytorch nightly instead which might have performance improvements.
 

comfy/patcher_extension.py

@@ -150,7 +150,7 @@ def merge_nested_dicts(dict1: dict, dict2: dict, copy_dict1=True):
     for key, value in dict2.items():
         if isinstance(value, dict):
             curr_value = merged_dict.setdefault(key, {})
-            merged_dict[key] = merge_nested_dicts(value, curr_value)
+            merged_dict[key] = merge_nested_dicts(curr_value, value)
         elif isinstance(value, list):
             merged_dict.setdefault(key, []).extend(value)
         else:

comfy/samplers.py

@@ -306,17 +306,10 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens
                                                                                  copy_dict1=False)
 
             if patches is not None:
-                # TODO: replace with merge_nested_dicts function
-                if "patches" in transformer_options:
-                    cur_patches = transformer_options["patches"].copy()
-                    for p in patches:
-                        if p in cur_patches:
-                            cur_patches[p] = cur_patches[p] + patches[p]
-                        else:
-                            cur_patches[p] = patches[p]
-                    transformer_options["patches"] = cur_patches
-                else:
-                    transformer_options["patches"] = patches
+                transformer_options["patches"] = comfy.patcher_extension.merge_nested_dicts(
+                    transformer_options.get("patches", {}),
+                    patches
+                )
 
             transformer_options["cond_or_uncond"] = cond_or_uncond[:]
             transformer_options["uuids"] = uuids[:]

comfy_api_nodes/nodes_veo2.py

@@ -27,6 +27,13 @@ from comfy_api_nodes.apinode_utils import (
 )
 
 AVERAGE_DURATION_VIDEO_GEN = 32
+MODELS_MAP = {
+    "veo-2.0-generate-001": "veo-2.0-generate-001",
+    "veo-3.1-generate": "veo-3.1-generate-preview",
+    "veo-3.1-fast-generate": "veo-3.1-fast-generate-preview",
+    "veo-3.0-generate-001": "veo-3.0-generate-001",
+    "veo-3.0-fast-generate-001": "veo-3.0-fast-generate-001",
+}
 
 def convert_image_to_base64(image: torch.Tensor):
     if image is None:
@@ -158,6 +165,7 @@ class VeoVideoGenerationNode(IO.ComfyNode):
         model="veo-2.0-generate-001",
         generate_audio=False,
     ):
+        model = MODELS_MAP[model]
         # Prepare the instances for the request
         instances = []
 
@@ -385,7 +393,7 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode):
                 ),
                 IO.Combo.Input(
                     "model",
-                    options=["veo-3.0-generate-001", "veo-3.0-fast-generate-001"],
+                    options=list(MODELS_MAP.keys()),
                     default="veo-3.0-generate-001",
                     tooltip="Veo 3 model to use for video generation",
                     optional=True,

comfy_extras/nodes_eps.py

@@ -1,5 +1,7 @@
+import torch
 from typing_extensions import override
 
+from comfy.k_diffusion.sampling import sigma_to_half_log_snr
 from comfy_api.latest import ComfyExtension, io
 
 
@@ -63,12 +65,105 @@ class EpsilonScaling(io.ComfyNode):
         return io.NodeOutput(model_clone)
 
 
+def compute_tsr_rescaling_factor(
+    snr: torch.Tensor, tsr_k: float, tsr_variance: float
+) -> torch.Tensor:
+    """Compute the rescaling score ratio in Temporal Score Rescaling.
+
+    See equation (6) in https://arxiv.org/pdf/2510.01184v1.
+    """
+    posinf_mask = torch.isposinf(snr)
+    rescaling_factor = (snr * tsr_variance + 1) / (snr * tsr_variance / tsr_k + 1)
+    return torch.where(posinf_mask, tsr_k, rescaling_factor) # when snr → inf, r = tsr_k
+
+
+class TemporalScoreRescaling(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TemporalScoreRescaling",
+            display_name="TSR - Temporal Score Rescaling",
+            category="model_patches/unet",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input(
+                    "tsr_k",
+                    tooltip=(
+                        "Controls the rescaling strength.\n"
+                        "Lower k produces more detailed results; higher k produces smoother results in image generation. Setting k = 1 disables rescaling."
+                    ),
+                    default=0.95,
+                    min=0.01,
+                    max=100.0,
+                    step=0.001,
+                    display_mode=io.NumberDisplay.number,
+                ),
+                io.Float.Input(
+                    "tsr_sigma",
+                    tooltip=(
+                        "Controls how early rescaling takes effect.\n"
+                        "Larger values take effect earlier."
+                    ),
+                    default=1.0,
+                    min=0.01,
+                    max=100.0,
+                    step=0.001,
+                    display_mode=io.NumberDisplay.number,
+                ),
+            ],
+            outputs=[
+                io.Model.Output(
+                    display_name="patched_model",
+                ),
+            ],
+            description=(
+                "[Post-CFG Function]\n"
+                "TSR - Temporal Score Rescaling (2510.01184)\n\n"
+                "Rescaling the model's score or noise to steer the sampling diversity.\n"
+            ),
+        )
+
+    @classmethod
+    def execute(cls, model, tsr_k, tsr_sigma) -> io.NodeOutput:
+        tsr_variance = tsr_sigma**2
+
+        def temporal_score_rescaling(args):
+            denoised = args["denoised"]
+            x = args["input"]
+            sigma = args["sigma"]
+            curr_model = args["model"]
+
+            # No rescaling (r = 1) or no noise
+            if tsr_k == 1 or sigma == 0:
+                return denoised
+
+            model_sampling = curr_model.current_patcher.get_model_object("model_sampling")
+            half_log_snr = sigma_to_half_log_snr(sigma, model_sampling)
+            snr = (2 * half_log_snr).exp()
+
+            # No rescaling needed (r = 1)
+            if snr == 0:
+                return denoised
+
+            rescaling_r = compute_tsr_rescaling_factor(snr, tsr_k, tsr_variance)
+
+            # Derived from scaled_denoised = (x - r * sigma * noise) / alpha
+            alpha = sigma * half_log_snr.exp()
+            return torch.lerp(x / alpha, denoised, rescaling_r)
+
+        m = model.clone()
+        m.set_model_sampler_post_cfg_function(temporal_score_rescaling)
+        return io.NodeOutput(m)
+
+
 class EpsilonScalingExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
         return [
             EpsilonScaling,
+            TemporalScoreRescaling,
         ]
 
+
 async def comfy_entrypoint() -> EpsilonScalingExtension:
     return EpsilonScalingExtension()