Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Working RescaleCFG node.

comfyanonymous58d5d71a9390
diff --git a/comfy/samplers.py b/comfy/samplers.py
index d7ff89850..a839ee9e2 100644
--- a/comfy/samplers.py
+++ b/comfy/samplers.py
@@ -248,7 +248,7 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, mod
cond, uncond = calc_cond_uncond_batch(model_function, cond, uncond, x, timestep, max_total_area, model_options)
if "sampler_cfg_function" in model_options:
- args = {"cond": x - cond, "uncond": x - uncond, "cond_scale": cond_scale, "timestep": timestep, "input": x}
+ args = {"cond": x - cond, "uncond": x - uncond, "cond_scale": cond_scale, "timestep": timestep, "input": x, "sigma": timestep}
return x - model_options["sampler_cfg_function"](args)
else:
return uncond + (cond - uncond) * cond_scale
diff --git a/comfy_extras/nodes_model_advanced.py b/comfy_extras/nodes_model_advanced.py
index 42596fbd5..09d2d9072 100644
--- a/comfy_extras/nodes_model_advanced.py
+++ b/comfy_extras/nodes_model_advanced.py
@@ -123,6 +123,45 @@ class ModelSamplingDiscrete:
m.add_object_patch("model_sampling", model_sampling)
return (m, )
+class RescaleCFG:
+ @classmethod
+ def INPUT_TYPES(s):
+ return {"required": { "model": ("MODEL",),
+ "multiplier": ("FLOAT", {"default": 0.7, "min": 0.0, "max": 1.0, "step": 0.01}),
+ }}
+ RETURN_TYPES = ("MODEL",)
+ FUNCTION = "patch"
+
+ CATEGORY = "advanced/model"
+
+ def patch(self, model, multiplier):
+ def rescale_cfg(args):
+ cond = args["cond"]
+ uncond = args["uncond"]
+ cond_scale = args["cond_scale"]
+ sigma = args["sigma"]
+ x_orig = args["input"]
+
+ #rescale cfg has to be done on v-pred model output
+ x = x_orig / (sigma * sigma + 1.0)
+ cond = ((x - (x_orig - cond)) * (sigma ** 2 + 1.0) ** 0.5) / (sigma)
+ uncond = ((x - (x_orig - uncond)) * (sigma ** 2 + 1.0) ** 0.5) / (sigma)
+
+ #rescalecfg
+ x_cfg = uncond + cond_scale * (cond - uncond)
+ ro_pos = torch.std(cond, dim=(1,2,3), keepdim=True)
+ ro_cfg = torch.std(x_cfg, dim=(1,2,3), keepdim=True)
+
+ x_rescaled = x_cfg * (ro_pos / ro_cfg)
+ x_final = multiplier * x_rescaled + (1.0 - multiplier) * x_cfg
+
+ return x_orig - (x - x_final * sigma / (sigma * sigma + 1.0) ** 0.5)
+
+ m = model.clone()
+ m.set_model_sampler_cfg_function(rescale_cfg)
+ return (m, )
+
NODE_CLASS_MAPPINGS = {
"ModelSamplingDiscrete": ModelSamplingDiscrete,
+ "RescaleCFG": RescaleCFG,
}