Update app.py

app.py

@@ -80,22 +80,22 @@ negative_prompt = "Aerial view, aerial view, overexposed, low quality, deformati
 
 @spaces.GPU(duration=90)
 def generate(segment, image, prompt, size, guidance_scale, num_inference_steps, frames, seed, progress=gr.Progress(track_tqdm=True) ):
-
-    random.seed(time.time())
-    seed = int(random.randrange(4294967294))
     if segment==1:
+        random.seed(time.time())
+        seed = int(random.randrange(4294967294))
         #Offload.offload(
         #    pipeline=pipe,
         #    config=offload_config,
         #)
         pipe.text_encoder.to("cuda")
         pipe.text_encoder_2.to("cuda")
-        with torch.no_grad():
-            prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_attention_mask, pooled_prompt_embeds, negative_pooled_prompt_embeds = pipe.encode_prompt(
-                prompt=prompt, do_classifier_free_guidance=True, negative_prompt=negative_prompt, device=device
-            )
+        #with torch.no_grad():
+        prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_attention_mask, pooled_prompt_embeds, negative_pooled_prompt_embeds = pipe.encode_prompt(
+            prompt=prompt, do_classifier_free_guidance=True, negative_prompt=negative_prompt, device=device
+        )
         pipe.text_encoder.to("cpu")
         pipe.text_encoder_2.to("cpu")
+        generator = torch.Generator(device='cuda').manual_seed(seed)
         transformer_dtype = pipe.transformer.dtype
         prompt_embeds = prompt_embeds.to(transformer_dtype)
         prompt_attention_mask = prompt_attention_mask.to(transformer_dtype)