Not going great, most of these are not great changes

img2img.py

@@ -9,63 +9,62 @@ from PIL import Image
 pipe = None
 start_time = time.time()
 torch.backends.cuda.matmul.allow_tf32 = True
-model_path = ("./models/stable-diffusion/card-generator-v1.safetensors")
-lora_path = "./models/stable-diffusion/Loras/blank-card-template-5.safetensors"
-detail_lora_path = "./models/stable-diffusion/Loras/add-detail-xl.safetensors"
-mimic_lora_path = "./models/stable-diffusion/Loras/EnvyMimicXL01.safetensors"
+model_path = "/media/drakosfire/Shared/models/stable-diffusion/card-generator-v1/card-generator-v1.safetensors"
+lora_path = "/media/drakosfire/Shared/models/stable-diffusion/card-generator-v1/blank-card-template-5.safetensors"
+detail_lora_path = "/media/drakosfire/Shared/models/stable-diffusion/card-generator-v1/add-detail-xl.safetensors"
 temp_image_path = "./image_temp/"
 card_pre_prompt = " blank magic card,high resolution, detailed intricate high quality border, textbox, high quality detailed magnum opus drawing of a "
 negative_prompts = "text, words, numbers, letters"
 image_list = []
 
+class img_generator(): 
+    def load_img_gen(prompt, item, mimic = None):
+        prompt = card_pre_prompt + item + ' ' + prompt
+        print(prompt)
+        
+        
+        pipe = StableDiffusionXLImg2ImgPipeline.from_single_file(model_path,
+                                                        custom_pipeline="low_stable_diffusion",                                                       
+                                                            torch_dtype=torch.float16, 
+                                                            variant="fp16").to("cuda")  
+        # Load LoRAs for controlling image
+        #pipe.load_lora_weights(lora_path, weight_name = "blank-card-template-5.safetensors",adapter_name = 'blank-card-template')    
+        pipe.load_lora_weights(detail_lora_path, weight_name = "add-detail-xl.safetensors", adapter_name = "add-detail-xl")
+        
+        # If mimic keyword has been detected, load the mimic LoRA and set adapter values
+        if mimic:
+            print("MIMIC!")
+            pipe.load_lora_weights(mimic_lora_path, weight_name = "EnvyMimicXL01.safetensors", adapter_name = "EnvyMimicXL")
+            pipe.set_adapters(['blank-card-template', "add-detail-xl", "EnvyMimicXL"], adapter_weights = [0.9,0.9,1.0])
+        else : 
+            pipe.set_adapters([ "add-detail-xl"], adapter_weights = [0.9])       
+        pipe.enable_vae_slicing()
+        return pipe, prompt
 
-def load_img_gen(prompt, item, mimic = None):
-    prompt = card_pre_prompt + item + ' ' + prompt
-    print(prompt)
-    # image_path = f"{user_input_template}"
-    # init_image = load_image(image_path).convert("RGB")
-    
-    pipe = StableDiffusionXLImg2ImgPipeline.from_single_file(model_path,
-                                                       custom_pipeline="low_stable_diffusion",                                                       
-                                                         torch_dtype=torch.float16, 
-                                                         variant="fp16").to("cuda")  
-    # Load LoRAs for controlling image
-    #pipe.load_lora_weights(lora_path, weight_name = "blank-card-template-5.safetensors",adapter_name = 'blank-card-template')    
-    pipe.load_lora_weights(detail_lora_path, weight_name = "add-detail-xl.safetensors", adapter_name = "add-detail-xl")
-    
-    # If mimic keyword has been detected, load the mimic LoRA and set adapter values
-    if mimic:
-        pipe.load_lora_weights(mimic_lora_path, weight_name = "EnvyMimicXL01.safetensors", adapter_name = "EnvyMimicXL")
-        pipe.set_adapters(['blank-card-template', "add-detail-xl", "EnvyMimicXL"], adapter_weights = [0.9,0.9,1.0])
-    else : 
-        pipe.set_adapters([ "add-detail-xl"], adapter_weights = [0.9])       
-    pipe.enable_vae_slicing()
-    return pipe, prompt
-
-def preview_and_generate_image(x,pipe, prompt, user_input_template, item):    
-    img_start = time.time()   
-    image = pipe(prompt=prompt,
-                strength = .9,
-                guidance_scale = 5,
-                image= user_input_template,
-                negative_promt = negative_prompts,
-                num_inference_steps=40,
-                height = 1024, width = 768).images[0]
-    
-    image = image.save(temp_image_path+str(x) + f"{item}.png")
-    output_image_path = temp_image_path+str(x) + f"{item}.png"
-    img_time = time.time() - img_start
-    img_its = 50/img_time
-    print(f"image gen time = {img_time} and {img_its} it/s")
+    def preview_and_generate_image(x,pipe, prompt, user_input_template, item):    
+        img_start = time.time()   
+        image = pipe(prompt=prompt,
+                    strength = .9,
+                    guidance_scale = 5,
+                    image= user_input_template,
+                    negative_prompt = negative_prompts,
+                    num_inference_steps=40,
+                    height = 1024, width = 768).images[0]
         
-    # Delete the image variable to keep VRAM open to load the LLM
-    del image
-    print(f"Memory after del {torch.cuda.memory_allocated()}")
-    print(image_list)
-    total_time = time.time() - start_time
-    print(total_time)
+        image = image.save(temp_image_path+str(x) + f"{item}.png")
+        output_image_path = temp_image_path+str(x) + f"{item}.png"
+        img_time = time.time() - img_start
+        img_its = 50/img_time
+        print(f"image gen time = {img_time} and {img_its} it/s")
+            
+        # Delete the image variable to keep VRAM open to load the LLM
+        del image
+        print(f"Memory after del {torch.cuda.memory_allocated()}")
+        print(image_list)
+        total_time = time.time() - start_time
+        print(total_time)
 
-    return output_image_path
+        return output_image_path
 
 
 

item_dict_gen.py

@@ -1,27 +1,25 @@
-from llama_cpp import Llama
+import replicate
 import ast
 import gc
 import torch
+import os
 
-model_path = "./models/starling-lm-7b-alpha.Q8_0.gguf"
+api_key = os.getenv('REPLICATE_API_TOKEN')
+
+
+model_path = "meta/meta-llama-3-70b-instruct"
 def load_llm(user_input):
-  llm = Llama(
-  model_path=model_path,
-  n_ctx=8192,  # The max sequence length to use - note that longer sequence lengths require much more resources
-  n_threads=8,            # The number of CPU threads to use, tailor to your system and the resulting performance
-  n_gpu_layers=32 # The number of layers to offload to GPU, if you have GPU acceleration available
-  )
-  return llm(
-  f"GPT4 User: {prompt_instructions}  the item is {user_input}: <|end_of_turn|>GPT4 Assistant:", # Prompt
-  max_tokens=768,  # Generate up to 512 tokens
-  stop=["</s>"],   # Example stop token - not necessarily correct for this specific model! Please check before using.
-  echo=False        # Whether to echo the prompt
-  )
+  input = {"prompt" : f" {prompt_instructions} the item is {user_input}"}
+  output = replicate.run(model_path,
+    input=input
+    )
+  return output
+  
 
 def call_llm_and_cleanup(user_input):
     # Call the LLM and store its output
-    llm_output = load_llm(user_input)
-    print(llm_output['choices'][0]['text'])
+    llm_output = "".join(load_llm(user_input))
+    print("".join(llm_output))
     gc.collect()
     if torch.cuda.is_available():
         torch.cuda.empty_cache()  # Clear VRAM allocated by PyTorch
@@ -55,11 +53,24 @@ def convert_to_dict(string):
 
 # Instructions past 4 are not time tested and may need to be removed.
 ### Meta prompted : 
-prompt_instructions = """ **Purpose**: Generate a structured inventory entry for a specific item as a hashmap. Follow the format provided in the examples below.
+prompt_instructions = """ **Purpose**: ONLY Generate a structured inventory entry for a specific item as a hashmap. Do NOT reply with anything other than a hashmap.
 
 **Instructions**:
 1. Replace `{item}` with the name of the user item, DO NOT CHANGE THE USER ITEM NAME enclosed in single quotes (e.g., `'Magic Wand'`).
-2. Ensure your request is formatted as a hashmap. 
+2. **Dictionary Structure**:
+    
+    {"{item}": {
+    'Name': "{item name}",
+    'Type': '{item type}',
+    'Rarity': '{item rarity},
+    'Value': '{item value}',
+    'Properties': ["{property1}", "{property2}", ...],
+    'Damage': '{damage formula} , '{damage type}',
+    'Weight': '{weight}',
+    'Description': "{item description}",
+    'Quote': "{item quote}",
+    'SD Prompt': "{special description for the item}"
+    } } 
 3. Weapons MUST have a key 'Damage' 
 4. The description should be brief and puncy, or concise and thoughtful.
 5. The quote and SD Prompt MUST be inside double quotations ie " ".
@@ -111,22 +122,6 @@ Imitative Predators. Mimics can alter their outward texture to resemble wood, st
 When it changes shape, a mimic excretes an adhesive that helps it seize prey and weapons that touch it. The adhesive is absorbed when the mimic assumes its amorphous form and on parts the mimic uses to move itself.
 Cunning Hunters. Mimics live and hunt alone, though they occasionally share their feeding grounds with other creatures. Although most mimics have only predatory intelligence, a rare few evolve greater cunning and the ability to carry on simple conversations in Common or Undercommon. Such mimics might allow safe passage through their domains or provide useful information in exchange for food.
 
-11. 
-**Format Example**:
-- **Dictionary Structure**:
-    
-    {"{item}": {
-    'Name': "{item name}",
-    'Type': '{item type}',
-    'Rarity': '{item rarity},
-    'Value': '{item value}',
-    'Properties': ["{property1}", "{property2}", ...],
-    'Damage': '{damage formula} , '{damage type}',
-    'Weight': '{weight}',
-    'Description': "{item description}",
-    'Quote': "{item quote}",
-    'SD Prompt': "{special description for the item}"
-    } }
     
 - **Input Placeholder**:
     - "{item}": Replace with the item name, ensuring it's wrapped in single quotes.

main.py

@@ -1,5 +1,6 @@
+import img2img
+from img2img import img_generator
 
-import img2img 
 import card_generator as card
 import utilities as u
 import ctypes
@@ -111,11 +112,11 @@ with gr.Blocks() as demo:
     # Called when pressing button to generate image, updates gallery by returning the list of image URLs
     def generate_image_update_gallery(num_img, sd_prompt,item_name, built_template):
         delete_temp_images()
-        print(type(built_template))
         image_list = []
-        img_gen, prompt = img2img.load_img_gen(sd_prompt, item_name)
+        image_generator = img_generator()
+        img_gen, prompt = image_generator.load_img_gen(sd_prompt, item_name)
         for x in range(num_img):
-            preview = img2img.preview_and_generate_image(x,img_gen, prompt, built_template, item_name)
+            preview = image_generator.preview_and_generate_image(x,img_gen, prompt, built_template, item_name)
             image_list.append(preview)
             yield image_list
             #generate_gallery.change(image_list)

template_builder.py

@@ -40,8 +40,6 @@ def paste_image_and_resize(base_image,sticker, x_position, y_position,img_width,
     return base_image
 
 def build_card_template(selected_border, selected_seed_image):
-    print(selected_seed_image)
-    print(type(selected_seed_image))
     selected_border = u.open_image_from_url(selected_border)
     if type(selected_seed_image) == str:      
         print(f"String : {selected_seed_image}")      
@@ -65,6 +63,8 @@ def build_card_template(selected_border, selected_seed_image):
     canvas = paste_image_and_resize(canvas, selected_seed_image,seed_x,seed_y, seed_width, seed_height)
     
     canvas.paste(selected_border,(0,0), mask = mask)
+    print(f"Canvas is : {canvas}")
+    print(f"Canvas is : {type(canvas)}")
 
     image_list.append(canvas)
 

user_input.py

@@ -21,7 +21,8 @@ def index_image_paths(repo_name,directory_path):
     files = []
     for content_file in contents:
         if content_file.type == "file":
-            files.append(content_file.download_url)  # Or content_file.path for just the path
+            media_url = content_file.download_url.replace("raw.githubusercontent.com", "media.githubusercontent.com/media")
+            files.append(media_url)  # Or content_file.path for just the path
     
     return files
 
@@ -51,7 +52,7 @@ def user_pick_item(user_prompt,list_of_items):
 def call_llm(user_input):
     # Process the query and get the response
     llm_call = igen.call_llm_and_cleanup(user_input)
-    response = llm_call['choices'][0]['text']
+    response = llm_call
         
         # Find the index of the phrase
     index = response.find(end_phrase)