<h2 id="heading-6"><span style="color:#868e96">Pre training</span></h2><p><span style="color:#868e96">I gathered 23 images of the prompt “Lestat from queen of the damned” with Stewart Townsend in them. I generated multiple images with the prompts Vampire, Damned Queen, Queen of the Damned, Lestat from Queen of the Damned, Interview with the vampire. I used the RealisticVision model because that is the model I will be using for final composition. I extracted the faces using </span><a target="_blank" rel="ugc" href="http://Faceswap.dev"><span style="color:#868e96">Faceswap.dev</span></a><span style="color:#868e96"> and then processed the images using Blip captions. I edited the texts to reduce bad prompts.</span></p><h3 id="heading-7"><span style="color:#868e96">Pre Training Notes</span></h3><p><span style="color:#868e96">I could have used the actual video clips of Stewart Townsend’s face and gone through the best images but who has time for that. I could have included full body images.</span></p><h2 id="heading-8"><span style="color:#868e96">Training</span></h2><p><span style="color:#868e96">I used the learning rate 0.005:100:0.0025:250,0.001:500,0.0005:1000,0.00025 for training with a batch size of 1 and Gradient accumulation step of 2. I switched to SD 1.5 EMA Only model for training. I am using a 1xRTX 3080 Ti on </span><a target="_blank" rel="ugc" href="http://Vast.ai"><span style="color:#868e96">Vast.ai</span></a><span style="color:#868e96"> cloud. Training used 11 Gb out of the total 12 Gb. I was paying $0.278 an hour.</span></p><h3 id="heading-9"><span style="color:#868e96">Training Notes</span></h3><p><span style="color:#868e96">I don’t know why but on occasion there will be a hallucination in the training images and it will display a morphed looking abstract image that seems to resemble a kaleidoscope. Majority of images have very distinct features such as blue eyes, gaunt face, pale skin, sunken red darkish eyelids, and long brown curly hair.</span></p><h3 id="heading-10"><span style="color:#868e96">Things that I could have done better</span></h3><p><span style="color:#868e96">I could have upscaled the images before extracting the faces so I could reduce blur.</span></p><p><span style="color:#868e96"><br /></span></p>

Files with SHA256: 79b0b4c739a3f44e71b1331c05166ee9267770ce0164b7b3a555f15fe509a691