Merge pull request #454 from bghira/main

better error message for text embed failure | fix for pipeline unload deleting pipeline before unloading the weights | sd3: use 64px interval for bucketing instead of 8px

helpers/arguments.py

@@ -1519,11 +1519,16 @@ def parse_args(input_args=None):
         logger.info(f"Text Cache location: {args.cache_dir_text}")
     else:
         deepfloyd_pixel_alignment = 8
-        if args.aspect_bucket_alignment != deepfloyd_pixel_alignment:
+        if not args.sd3 and args.aspect_bucket_alignment != deepfloyd_pixel_alignment:
             logger.warning(
                 f"Overriding aspect bucket alignment pixel interval to {deepfloyd_pixel_alignment}px instead of {args.aspect_bucket_alignment}px."
             )
             args.aspect_bucket_alignment = deepfloyd_pixel_alignment
+        elif args.sd3:
+            logger.warning(
+                "Stable Diffusion 3 requires a pixel alignment interval of 64px. Updating value."
+            )
+            args.aspect_bucket_alignment = 64
 
     if "deepfloyd-stage2" in args.model_type and args.resolution < 256:
         logger.warning(

helpers/caching/sdxl_embeds.py

@@ -585,7 +585,9 @@ def compute_embeddings_for_sdxl_prompts(
                             f"\n-> id: {self.id}, data_backend id: {self.data_backend.id}"
                         )
                         should_encode = True
-                        raise Exception("This won't work. We cannot continue.")
+                        raise Exception(
+                            "Cache retrieval for text embed file failed. Ensure your dataloader config value for skip_file_discovery does not contain 'text', and that preserve_data_backend_cache is disabled or unset."
+                        )
                 if should_encode:
                     # If load_from_cache is True, should_encode would be False unless we failed to load.
                     # self.debug_log(f"Encoding prompt: {prompt}")
@@ -704,7 +706,9 @@ def compute_embeddings_for_legacy_prompts(
                             f"\n-> error: {e}"
                         )
                         should_encode = True
-                        raise Exception("This won't work. We cannot continue.")
+                        raise Exception(
+                            "Cache retrieval for text embed file failed. Ensure your dataloader config value for skip_file_discovery does not contain 'text', and that preserve_data_backend_cache is disabled or unset."
+                        )
 
                 if should_encode:
                     # self.debug_log(f"Encoding prompt: {prompt}")
@@ -822,7 +826,9 @@ def compute_embeddings_for_sd3_prompts(
                             f"\n-> id: {self.id}, data_backend id: {self.data_backend.id}"
                         )
                         should_encode = True
-                        raise Exception("This won't work. We cannot continue.")
+                        raise Exception(
+                            "Cache retrieval for text embed file failed. Ensure your dataloader config value for skip_file_discovery does not contain 'text', and that preserve_data_backend_cache is disabled or unset."
+                        )
                 if should_encode:
                     # If load_from_cache is True, should_encode would be False unless we failed to load.
                     self.debug_log(f"Encoding prompt: {prompt}")

inference_comparison.py

@@ -146,14 +146,14 @@ def combine_and_label_images(images_info, output_path):
         # Generate image with specified settings
         settings = pipeline_info.get("settings", {})
         image = pipeline(prompt, generator=manual_seed(420420420), **settings).images[0]
+        # Unload LoRA weights if they were loaded
+        if "lora" in pipeline_info:
+            pipeline.unload_lora_weights()
         del pipeline
         image.save(image_path, format="PNG")
 
         images_info.append((image, pipeline_info["label"]))
 
-        # Unload LoRA weights if they were loaded
-        if "lora" in pipeline_info:
-            pipeline.unload_lora_weights()
 
     # Combine and label images
     combine_and_label_images(images_info, f"{target_dir}/combined_image.png")