Implement Differential Diffusion

comfy/samplers.py

@@ -276,6 +276,8 @@ def __init__(self, model):
         self.inner_model = model
     def forward(self, x, sigma, uncond, cond, cond_scale, denoise_mask, model_options={}, seed=None):
         if denoise_mask is not None:
+            if "denoise_mask_function" in model_options:
+                denoise_mask = model_options["denoise_mask_function"](sigma, denoise_mask)
             latent_mask = 1. - denoise_mask
             x = x * denoise_mask + (self.latent_image + self.noise * sigma.reshape([sigma.shape[0]] + [1] * (len(self.noise.shape) - 1))) * latent_mask
         out = self.inner_model(x, sigma, cond=cond, uncond=uncond, cond_scale=cond_scale, model_options=model_options, seed=seed)

comfy_extras/nodes_differential_diffusion.py

@@ -0,0 +1,97 @@
+# code adapted from https://github.com/exx8/differential-diffusion
+
+import torch
+import inspect
+
+class DifferentialDiffusion():
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"model": ("MODEL", ),
+                            }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "apply"
+    CATEGORY = "_for_testing"
+    INIT = False
+
+    @classmethod
+    def IS_CHANGED(s, *args, **kwargs):
+        DifferentialDiffusion.INIT = s.INIT = True
+        return ""
+
+    def __init__(self) -> None:
+        DifferentialDiffusion.INIT = False
+        self.sigmas: torch.Tensor = None
+        self.thresholds: torch.Tensor = None
+        self.mask_i = None
+        self.valid_sigmas = False
+        self.varying_sigmas_samplers = ["dpmpp_2s", "dpmpp_sde", "dpm_2", "heun", "restart"]
+
+    def apply(self, model):
+        model = model.clone()
+        model.model_options["denoise_mask_function"] = self.forward
+        return (model,)
+
+    def init_sigmas(self, sigma: torch.Tensor, denoise_mask: torch.Tensor):
+        self.__init__()
+        self.sigmas, sampler = find_outer_instance("sigmas", callback=get_sigmas_and_sampler) or (None, "")
+        self.valid_sigmas = not ("sample_" not in sampler or any(s in sampler for s in self.varying_sigmas_samplers)) or "generic" in sampler
+        if self.sigmas is None:
+            self.sigmas = sigma[:1].repeat(2)
+            self.sigmas[-1].zero_()
+        self.sigmas_min = self.sigmas.min()
+        self.sigmas_max = self.sigmas.max()
+        self.thresholds = torch.linspace(1, 0, self.sigmas.shape[0], dtype=sigma.dtype, device=sigma.device)
+        self.thresholds_min_len = self.thresholds.shape[0] - 1
+        if self.valid_sigmas:
+            thresholds = self.thresholds[:-1].reshape(-1, 1, 1, 1, 1)
+            mask = denoise_mask.unsqueeze(0)
+            mask = (mask >= thresholds).to(denoise_mask.dtype)
+            self.mask_i = iter(mask)
+
+    def forward(self, sigma: torch.Tensor, denoise_mask: torch.Tensor):
+        if self.sigmas is None or DifferentialDiffusion.INIT:
+            self.init_sigmas(sigma, denoise_mask)
+        if self.valid_sigmas:
+            try:
+                return next(self.mask_i)
+            except StopIteration:
+                self.valid_sigmas = False
+        if self.thresholds_min_len > 1:
+            nearest_idx = (self.sigmas - sigma[0]).abs().argmin()
+            if not self.thresholds_min_len > nearest_idx:
+                nearest_idx = -2
+            threshold = self.thresholds[nearest_idx]
+        else:
+            threshold = (sigma[0] - self.sigmas_min) / (self.sigmas_max - self.sigmas_min)
+        return (denoise_mask >= threshold).to(denoise_mask.dtype)
+
+def get_sigmas_and_sampler(frame, target):
+    found = frame.f_locals[target]
+    if isinstance(found, torch.Tensor) and found[-1] < 0.1:
+        return found, frame.f_code.co_name
+    return False
+
+def find_outer_instance(target: str, target_type=None, callback=None):
+    frame = inspect.currentframe()
+    i = 0
+    while frame and i < 100:
+        if target in frame.f_locals:
+            if callback is not None:
+                res = callback(frame, target)
+                if res:
+                    return res
+            else:
+                found = frame.f_locals[target]
+                if isinstance(found, target_type):
+                    return found
+        frame = frame.f_back
+        i += 1
+    return None
+
+
+NODE_CLASS_MAPPINGS = {
+    "DifferentialDiffusion": DifferentialDiffusion,
+}
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "DifferentialDiffusion": "Differential Diffusion",
+}

nodes.py

@@ -1961,6 +1961,7 @@ def init_custom_nodes():
         "nodes_photomaker.py",
         "nodes_cond.py",
         "nodes_stable_cascade.py",
+        "nodes_differential_diffusion.py",
     ]
 
     for node_file in extras_files: