add guide code

README.md

@@ -1,2 +1,5 @@
 # stable-diffusion-guide
 Here is a guide for Stable Diffusion
+
+
+Let's get start from `guide.ipynb`!

guide.py

@@ -0,0 +1,202 @@
+#!/usr/bin/env python
+# coding: utf-8
+# Author: Jiaojiao Ye
+
+# The Stable Diffusion Guide 
+
+# ## Prompt engineering
+
+# In[2]:
+
+
+model_id = "runwayml/stable-diffusion-v1-5"    
+
+
+# ## Speed Optimization
+
+# In[11]:
+
+
+from diffusers import StableDiffusionPipeline 
+import torch
+
+pipe = StableDiffusionPipeline.from_pretrained(model_id)
+
+
+# In[8]:
+
+
+prompt = "portrait photo of a old warrior chief"  
+
+
+# In[12]:
+
+
+pipe = pipe.to("cuda")  
+
+
+# In[13]:
+
+
+generator = torch.Generator("cuda").manual_seed(0)  
+
+
+# In[14]:
+
+
+image = pipe(prompt, generator=generator).images[0]                                                                                                                                                                                           
+image.save('./outputs/output_fp32.png')     
+
+
+# In[17]:
+
+
+import torch                                                                                                                                                                                                                                  
+
+pipe = StableDiffusionPipeline.from_pretrained(model_id,revision="fp16", torch_dtype=torch.float16)                                                                                                                                                           
+pipe = pipe.to("cuda") 
+
+
+# In[33]:
+
+
+generator = torch.Generator("cuda").manual_seed(0)                                                                                                                                                                                            
+
+with torch.autocast("cuda"):
+    image = pipe(prompt, generator=generator).images[0]                                                                                                                                                                                           
+image.save('./outputs/output_fp16.png')      
+
+
+# In[35]:
+
+
+pipe.scheduler.compatibles    
+
+
+# In[37]:
+
+
+from diffusers import DPMSolverMultistepScheduler                                                                                                                                                                                             
+
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)  
+
+
+# In[41]:
+
+
+generator = torch.Generator("cuda").manual_seed(0)                                                                                                                                                                                            
+
+with torch.autocast("cuda"):
+    image = pipe(prompt, generator=generator, num_inference_steps=20).images[0]                                                                                                                                                                   
+image.save('./outputs/output_fp16_20steps.png')                   
+
+
+# ## Memory Optimization
+
+# In[42]:
+
+
+def get_inputs(batch_size=1):                                                                                                                                                                                                                 
+  generator = [torch.Generator("cuda").manual_seed(i) for i in range(batch_size)]                                                                                                                                                             
+  prompts = batch_size * [prompt]                                                                                                                                                                                                             
+  num_inference_steps = 20                                                                                                                                                                                                                    
+
+  return {"prompt": prompts, "generator": generator, "num_inference_steps": num_inference_steps}     
+
+
+# In[43]:
+
+
+from PIL import Image                                                                                                                                                                                                                         
+
+def image_grid(imgs, rows=2, cols=2):                                                                                                                                                                                                         
+    w, h = imgs[0].size                                                                                                                                                                                                                       
+    grid = Image.new('RGB', size=(cols*w, rows*h))                                                                                                                                                                                            
+
+    for i, img in enumerate(imgs):                                                                                                                                                                                                            
+        grid.paste(img, box=(i%cols*w, i//cols*h))                                                                                                                                                                                            
+    return grid                                   
+
+
+# In[45]:
+
+
+with torch.autocast("cuda"):
+    images = pipe(**get_inputs(batch_size=4)).images                                                                                                                                                                                              
+image_grid(images)    
+image_grid(images).save('./outputs/output_bs4.png')
+
+# In[46]:
+
+
+pipe.enable_attention_slicing() 
+
+
+# In[49]:
+
+
+with torch.autocast('cuda'):
+    images = pipe(**get_inputs(batch_size=8)).images                                                                                                                                                                                              
+image_grid(images, rows=2, cols=4) 
+image_grid(images, rows=2, cols=4).save('./outputs/output_bs8.png')   
+
+# ## Quality Improvements
+
+# In[50]:
+
+
+from diffusers import AutoencoderKL                                                                                                                                                                                                           
+
+vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16).to("cuda")  
+
+
+# In[52]:
+
+
+pipe.vae = vae          
+
+
+# In[53]:
+
+
+with torch.autocast('cuda'):
+    images = pipe(**get_inputs(batch_size=8)).images                                                                                                                                                                                              
+image_grid(images, rows=2, cols=4)    
+
+# In[54]:
+
+
+prompt += ", tribal panther make up, blue on red, side profile, looking away, serious eyes"                                                                                                                                                   
+
+
+# In[55]:
+
+
+prompt += " 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta"                                                                                                                                         
+prompt                                                                                                                                                                                                                                        
+
+
+# In[57]:
+
+
+with torch.autocast('cuda'):
+    images = pipe(**get_inputs(batch_size=8)).images                                                                                                                                                                                              
+image_grid(images, rows=2, cols=4)                                                                                                                                                                                                            
+
+
+# In[59]:
+
+
+prompts = [                                                                                                                                                                                                                                   
+    "portrait photo of the oldest warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",                                                                                                                                                                                                                                                                   
+    "portrait photo of a old warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",                                                                                                                                                                                                                                                                        
+    "portrait photo of a warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",                                                                                                                                                                                                                                                                            
+    "portrait photo of a young warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3  --beta --upbeta",                                                                                                                                                                                                                                                                      
+]                                                                                                                                                                                                                                             
+
+generator = [torch.Generator("cuda").manual_seed(1) for _ in range(len(prompts))]  # 1 because we want the 2nd image                                                                                                                          
+
+with torch.autocast('cuda'):
+    images = pipe(prompt=prompts, generator=generator, num_inference_steps=25).images                                                                                                                                                             
+image_grid(images)                                                                                                                                                                                                                         
+image_grid(images).save('./outputs/output_bs8_improved.png')