Good day folks,I'm here to present my initial work on creating a LoRA model for stable diffusion, focusing on Sarah Miller, a character from the video game "The Last of Us".<h3 id="heading-1882">Who is Sarah Miller ?</h3>Sarah Miller is a character from the video game "The Last of Us", developed by Naughty Dog. She is the 12-year-old daughter of the game's protagonist, Joel. Sarah appears in the early stages of the game.<h2 id="heading-312">How to use</h2>In all versions is it possible to change the age of the character, ( example <a target="_blank" rel="ugc" href="/posts/563577">here</a> ) so please specify the correct age in the prompt.It should also be possible to change various physical aspects, such as length and color of hair, eye color etc ... so be specific in the prompt if the result does not correspond to your expectations.The prompt suggested to have a character as similar as possible to the one in the video game<pre><code>1girl, Sara Miller, 12 years old, blonde, short hairs, blue eyes, bracelet, necklaces, wearing pajamas, double long sleeves, long pants, </code></pre>Sometimes it is necessary to add<pre><code>double long sleeves, sleeves</code></pre>to the negative prompt to remove some artifacts<h2 id="heading-1883"><a target="_blank" rel="ugc" href="/models/103126?modelVersionId=110390">SD1.5 v1.0</a></h2>This is my first foray into this kind of work for stable diffusion 1.5, and I'm aware of the challenges and the steep learning curve involved. A significant hurdle was the limited amount of visual references available for the character, which added complexity to the task. While I am conscious that the current model may not be up to the mark, I consider this as a stepping stone towards better and more accurate models. I am committed to continuously improving and refining this model, and I'm looking forward to sharing updated versions in the future. I appreciate your understanding and support as I navigate this new territory.It only works with Stable Diffusion 1.5 and all derived models (which are most of the ones available here on Civitai)<h2 id="heading-281"><a target="_blank" rel="ugc" href="/models/103126?modelVersionId=158962">SD1.5 v2.3</a></h2>Improved version of v1.0.Lots of work still needs to be done to achieve a good result.It only works with Stable Diffusion 1.5 and all derived models (which are most of the ones available here on Civitai)<h2 id="heading-1884"><a target="_blank" rel="ugc" href="/models/103126?modelVersionId=151299">SDXL v1.0</a></h2>Much better version of v1.0 for stable diffusion 1.5 thanks to the capabilities of Stable Diffusion XLIt only works with Stable Diffusion XL and all derived models<h2 id="heading-1885"><a target="_blank" rel="ugc" href="/models/103126?modelVersionId=153608">SDXL v3.0</a></h2>slightly improved version with the addition of images with better qualityIt only works with Stable Diffusion XL and all derived models<h2 id="heading-1885"><a target="_blank" rel="ugc" href="/models/103126?modelVersionId=158845">SDXL v3.1 ( 80 epochs version of SDXLv3.0 )</a></h2>This is the same training batch of SDXLv3.0, but with a much higher epoch ( v3.0 is 18 epochs v3.1 is 80 epochs ).Suggested weight: 0.8It only works with Stable Diffusion XL and all derived models<h2 id="heading-1886">Future releases for stable diffusion 1.5?</h2>Yes, I'm working on a improved version of v1.0. You can see the firsts results <a target="_blank" rel="ugc" href="/posts/569151">here</a><h2 id="heading-148">Dataset source ?</h2>You can download the dataset of images used to train the models <a rel="ugc" href="/api/download/training-data/158962">here ( sd1.5 versions )</a> and <a rel="ugc" href="/api/download/training-data/158845">here ( SDXL versions )</a>

Sarah Miller - The Last of Us ( SD XL + SD 1.5 )

Stronger version - 88 epoch version - same training batch of v3.0

dataset.zip

sarah_tlou_sdxl_v3-000088.safetensors

Files with SHA256: 01cee469a86942ac84374e76fc8afe0f8b13364d64c78816bd17de71ba171d8f