Fixed an issue with CFG time embeds on SDXL

toolkit/stable_diffusion_model.py

@@ -444,8 +444,8 @@ class StableDiffusion:
 
                 if do_classifier_free_guidance:
                     # todo check this with larget batches
-                    train_util.concat_embeddings(
-                        add_time_ids, add_time_ids, 1
+                    add_time_ids = train_util.concat_embeddings(
+                        add_time_ids, add_time_ids, int(latents.shape[0])
                     )
                 else:
                     # concat to fit batch size