diffirential guidance is WORKING (from what I can tell)

toolkit/config_modules.py

@@ -242,6 +242,7 @@ class ModelConfig:
         self.dtype: str = kwargs.get('dtype', 'float16')
         self.vae_path = kwargs.get('vae_path', None)
         self.refiner_name_or_path = kwargs.get('refiner_name_or_path', None)
+        self._original_refiner_name_or_path = self.refiner_name_or_path
         self.refiner_start_at = kwargs.get('refiner_start_at', 0.5)
 
         # only for SDXL models for now
@@ -286,6 +287,14 @@ class SliderTargetConfig:
         self.shuffle: bool = kwargs.get('shuffle', False)
 
 
+class GuidanceConfig:
+    def __init__(self, **kwargs):
+        self.target_class: str = kwargs.get('target_class', '')
+        self.guidance_scale: float = kwargs.get('guidance_scale', 1.0)
+        self.positive_prompt: str = kwargs.get('positive_prompt', '')
+        self.negative_prompt: str = kwargs.get('negative_prompt', '')
+
+
 class SliderConfigAnchors:
     def __init__(self, **kwargs):
         self.prompt = kwargs.get('prompt', '')

toolkit/data_loader.py

@@ -2,6 +2,7 @@ import copy
 import json
 import os
 import random
+import traceback
 from functools import lru_cache
 from typing import List, TYPE_CHECKING
 
@@ -378,6 +379,7 @@ class AiToolkitDataset(LatentCachingMixin, BucketsMixin, CaptionMixin, Dataset):
                 )
                 self.file_list.append(file_item)
             except Exception as e:
+                print(traceback.format_exc())
                 print(f"Error processing image: {file}")
                 print(e)
                 bad_count += 1

toolkit/data_transfer_object/data_loader.py

@@ -35,7 +35,6 @@ class FileItemDTO(
     ArgBreakMixin,
 ):
     def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
         self.path = kwargs.get('path', None)
         self.dataset_config: 'DatasetConfig' = kwargs.get('dataset_config', None)
         # process width and height
@@ -48,6 +47,7 @@ class FileItemDTO(
             h, w = img.size
         self.width: int = w
         self.height: int = h
+        super().__init__(*args, **kwargs)
 
         # self.caption_path: str = kwargs.get('caption_path', None)
         self.raw_caption: str = kwargs.get('raw_caption', None)

toolkit/dataloader_mixins.py

@@ -779,11 +779,21 @@ class PoiFileItemDTOMixin:
             if self.poi not in json_data['poi']:
                 raise Exception(f"Error: poi not found in caption file: {caption_path}")
             # poi has, x, y, width, height
-            poi = json_data['poi'][self.poi]
-            self.poi_x = int(poi['x'])
-            self.poi_y = int(poi['y'])
-            self.poi_width = int(poi['width'])
-            self.poi_height = int(poi['height'])
+            # do full image if no poi
+            self.poi_x = 0
+            self.poi_y = 0
+            self.poi_width = self.width
+            self.poi_height = self.height
+            try:
+                if self.poi in json_data['poi']:
+                    poi = json_data['poi'][self.poi]
+                    self.poi_x = int(poi['x'])
+                    self.poi_y = int(poi['y'])
+                    self.poi_width = int(poi['width'])
+                    self.poi_height = int(poi['height'])
+            except Exception as e:
+                pass
+
 
             # handle flipping
             if kwargs.get('flip_x', False):

toolkit/stable_diffusion_model.py

@@ -1059,7 +1059,7 @@ class StableDiffusion:
         refiner_config_path = os.path.join(ORIG_CONFIGS_ROOT, 'sd_xl_refiner.yaml')
         # load the refiner model
         dtype = get_torch_dtype(self.dtype)
-        model_path = self.model_config.refiner_name_or_path
+        model_path = self.model_config._original_refiner_name_or_path
         if not os.path.exists(model_path) or os.path.isdir(model_path):
             # TODO only load unet??
             refiner = StableDiffusionXLImg2ImgPipeline.from_pretrained(