Update pipelines/flux_pipeline/transformer.py

pipelines/flux_pipeline/transformer.py

@@ -41,19 +41,6 @@ from diffusers.utils.torch_utils import maybe_allow_in_graph
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
-def log_scale_masking(value, min_value=1, max_value=10):
-    # Convert the value into a positive domain for the logarithmic function
-    normalized_value = 1*value
-
-    # Apply logarithmic scaling
-    # log_scaled_value = 1-np.exp(-normalized_value)
-    log_scaled_value = 2.0* math.log(normalized_value+1, 2) / math.log(2, 2)  # np.log1p(x) = log(1 + x)
-    # print(log_scaled_value)
-
-    # Rescale to original range
-    scaled_value = log_scaled_value * (max_value - min_value) + min_value
-
-    return min(max_value, int(scaled_value))
 
 class FluxAttnProcessor2_0:
     """Attention processor used typically in processing the SD3-like self-attention projections."""
@@ -137,7 +124,7 @@ class FluxAttnProcessor2_0:
         if neg_mode:
             res = int(math.sqrt((end_of_hidden_states-(text_seq if encoder_hidden_states is None else 0)) // num))
             hw = res*res
-            mask_ = torch.ones(1, res, num*res, res, num*res).to(query.device)
+            mask_ = torch.zeros(1, res, num*res, res, num*res).to(query.device)
             for i in range(num):
                 mask_[:, :, i*res:(i+1)*res, :, i*res:(i+1)*res] = 1
             mask_ = rearrange(mask_, "b h w h1 w1 -> b (h w) (h1 w1)")