fix

stable_diffusion_tf/stable_diffusion.py

@@ -1,3 +1,5 @@
+import logging
+
 import numpy as np
 from tqdm import tqdm
 import math
@@ -58,16 +60,19 @@ def generate(
       progress_call_back=None
   ):
     # 1) 将条件输入文本(prompt) Tokenize, prompt不能太长, 要小于77个token
-    inputs: List[int] = self.tokenizer.encode(prompt)  # a list of int, start with 49406, end with 49407
-    assert len(inputs) < MAX_TEXT_LEN, "Prompt is too long (should be < 77 tokens)"
-    phrase = inputs + [49407] * (MAX_TEXT_LEN - len(inputs))  # padding to MAX_TEXT_LEN
-    phrase = np.array(phrase)[None].astype("int32")  # shape=(1, MAX_TEXT_LEN)
-    phrase = np.repeat(phrase, batch_size, axis=0)
-
-    # Encode prompt tokens (and their positions) into a "context vector"
     pos_ids = np.array(list(range(MAX_TEXT_LEN)))[None].astype("int32")  # shape=(1, MAX_TEXT_LEN)
     pos_ids = np.repeat(pos_ids, batch_size, axis=0)
-    context = self.text_encoder.predict_on_batch([phrase, pos_ids])
+    if prompt:
+      inputs: List[int] = self.tokenizer.encode(prompt)  # a list of int, start with 49406, end with 49407
+      assert len(inputs) < MAX_TEXT_LEN, "Prompt is too long (should be < 77 tokens)"
+      phrase = inputs + [49407] * (MAX_TEXT_LEN - len(inputs))  # padding to MAX_TEXT_LEN
+      phrase = np.array(phrase)[None].astype("int32")  # shape=(1, MAX_TEXT_LEN)
+      phrase = np.repeat(phrase, batch_size, axis=0)
+      # Encode prompt tokens (and their positions) into a "context vector"
+      context = self.text_encoder.predict_on_batch([phrase, pos_ids])
+    else:
+      logging.info('prompt is empty!')
+      context = None
 
     input_image_tensor, input_image_array = None, None
     if input_image is not None:
@@ -212,11 +217,14 @@ def get_model_output(
     unconditional_latent = self.diffusion_model.predict_on_batch(
       [latent, t_emb, unconditional_context]
     )
-    latent = self.diffusion_model.predict_on_batch([latent, t_emb, context])
-    return unconditional_latent + unconditional_guidance_scale * (
-        latent - unconditional_latent
-    )
-
+    if context:
+      latent = self.diffusion_model.predict_on_batch([latent, t_emb, context])
+      return unconditional_latent + unconditional_guidance_scale * (
+          latent - unconditional_latent
+      )
+    else:
+      return unconditional_latent
+
   def get_x_prev_and_pred_x0(self, x, e_t, index, a_t, a_prev, temperature, seed):
     sigma_t = 0
     sqrt_one_minus_at = math.sqrt(1 - a_t)