fix

lib_neutral_prompt/cfg_denoiser_hijack.py

@@ -28,7 +28,7 @@ def combine_denoised_hijack(
         aux_cond_delta = prompt.accept(AuxCondDeltaVisitor(), args, cond_delta, 0)
 
         if prompt.local_transform is not None:
-            cond_delta, weight = apply_masked_transform(cond_delta, prompt.local_transform)
+            cond_delta = apply_affine_transform(cond_delta, prompt.local_transform)
             aux_cond_delta = apply_affine_transform(aux_cond_delta, prompt.local_transform)
 
         cfg_cond = denoised[batch_i] + aux_cond_delta * cond_scale
@@ -184,14 +184,14 @@ def get_cond_delta(prompt: neutral_prompt_parser.PromptExpr, args: CombineDenois
     weight = prompt.weight
 
     if prompt.local_transform is not None:
-        cond_delta, weight = apply_masked_transform(cond_delta, prompt.local_transform)
-        weight = weight
+        transformed_cond, weight = apply_masked_transform(cond_delta + args.uncond, prompt.local_transform, prompt.weight)
+        cond_delta = transformed_cond - args.uncond
 
     return cond_delta, weight
 
 
-def apply_masked_transform(tensor: torch.Tensor, affine: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-    tensor_with_mask = torch.cat([tensor, create_cosine_feathered_mask(tensor.shape[-2:]).unsqueeze(0).to(tensor.device)])
+def apply_masked_transform(tensor: torch.Tensor, affine: torch.Tensor, weight: float) -> Tuple[torch.Tensor, torch.Tensor]:
+    tensor_with_mask = torch.cat([tensor, create_cosine_feathered_mask(tensor.shape[-2:], weight).unsqueeze(0).to(tensor.device)])
     transformed = apply_affine_transform(tensor_with_mask, affine)
     return transformed[:-1], transformed[-1]
 
@@ -203,11 +203,11 @@ def apply_affine_transform(tensor, affine):
     affine[1, 0] /= aspect_ratio
 
     grid = F.affine_grid(affine.unsqueeze(0), tensor.unsqueeze(0).size(), align_corners=False)
-    transformed_tensors = F.grid_sample(tensor.unsqueeze(0), grid, align_corners=False)
+    transformed_tensors = F.grid_sample(tensor.unsqueeze(0), grid, 'nearest', align_corners=False)
     return transformed_tensors.squeeze(0)
 
 
-def create_cosine_feathered_mask(size):
+def create_cosine_feathered_mask(size, weight: float):
     """
     Create a cosine-based feathered mask.
     """
@@ -215,7 +215,7 @@ def create_cosine_feathered_mask(size):
     dist = torch.sqrt(x**2 + y**2)
     mask = 0.5 * (1 + torch.cos(torch.pi * dist))
     mask[dist > 1] = 0
-    return mask.float()
+    return mask.float() * weight
 
 
 def get_perpendicular_component(normal: torch.Tensor, vector: torch.Tensor) -> torch.Tensor:

scripts/neutral_prompt.py

@@ -7,9 +7,6 @@ import torch
 import functools
 
 
-sampling_step = 0
-
-
 class NeutralPromptScript(scripts.Script):
     def __init__(self):
         self.accordion_interface = None
@@ -109,25 +106,14 @@ class CombinePreNoiseArgs:
     cond_indices: List[Tuple[int, float]]
 
 
-noises = []
-
-
 def on_cfg_denoiser(params: script_callbacks.CFGDenoiserParams):
     if not global_state.is_enabled:
         return
 
-    global noises, sampling_step
-    sampling_step += 1
-    if sampling_step == 1:
-        noises = params.x.clone()
-        return
-
-    for batch_i, (prompt, cond_indices) in enumerate(zip(global_state.prompt_exprs, global_state.batch_cond_indices)):
+    for prompt, cond_indices in zip(global_state.prompt_exprs, global_state.batch_cond_indices):
         args = CombinePreNoiseArgs(params.x, cond_indices)
         inv_transforms = prompt.accept(GlobalToLocalAffineVisitor(), args, 0)
-        for cond_index, weight in cond_indices:
-            # noisy_component = noises[cond_index] * torch.sum(noises[cond_index] * params.x[cond_index]) / torch.norm(noises[cond_index]) ** 2
-            # params.x[cond_index] = apply_affine_transform(params.x[cond_index] - noisy_component, inv_transforms[cond_index]) + noisy_component
+        for cond_index, _ in cond_indices:
             params.x[cond_index] = apply_affine_transform(params.x[cond_index], inv_transforms[cond_index])