Add CFGZeroStar node.

Works on all models that use a negative prompt but is meant for rectified
flow models.

comfy_extras/nodes_cfg.py

@@ -0,0 +1,45 @@
+import torch
+
+# https://github.com/WeichenFan/CFG-Zero-star
+def optimized_scale(positive, negative):
+    positive_flat = positive.reshape(positive.shape[0], -1)
+    negative_flat = negative.reshape(negative.shape[0], -1)
+
+    # Calculate dot production
+    dot_product = torch.sum(positive_flat * negative_flat, dim=1, keepdim=True)
+
+    # Squared norm of uncondition
+    squared_norm = torch.sum(negative_flat ** 2, dim=1, keepdim=True) + 1e-8
+
+    # st_star = v_cond^T * v_uncond / ||v_uncond||^2
+    st_star = dot_product / squared_norm
+
+    return st_star.reshape([positive.shape[0]] + [1] * (positive.ndim - 1))
+
+class CFGZeroStar:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"model": ("MODEL",),
+                            }}
+    RETURN_TYPES = ("MODEL",)
+    RETURN_NAMES = ("patched_model",)
+    FUNCTION = "patch"
+    CATEGORY = "advanced/guidance"
+
+    def patch(self, model):
+        m = model.clone()
+        def cfg_zero_star(args):
+            guidance_scale = args['cond_scale']
+            x = args['input']
+            cond_p = args['cond_denoised']
+            uncond_p = args['uncond_denoised']
+            out = args["denoised"]
+            alpha = optimized_scale(x - cond_p, x - uncond_p)
+
+            return out + uncond_p * (alpha - 1.0)  + guidance_scale * uncond_p * (1.0 - alpha)
+        m.set_model_sampler_post_cfg_function(cfg_zero_star)
+        return (m, )
+
+NODE_CLASS_MAPPINGS = {
+    "CFGZeroStar": CFGZeroStar
+}

nodes.py

@@ -2267,6 +2267,7 @@ def init_builtin_extra_nodes():
         "nodes_lotus.py",
         "nodes_hunyuan3d.py",
         "nodes_primitive.py",
+        "nodes_cfg.py",
     ]
 
     import_failed = []