Preface: Sometime last year, whilst I was still using Midjourney, I managed to create this effect and fell in love with it. It was rather scribbly in comparison to what you see here, but the general premise was a grungy cyberpunk angel type effect. I've used 52 of my previously created images to train this LoRA. I have absolutely no idea what I am doing at this stage, so training settings used in the creation were basically the default settings (I trained here at civitai).<h3 id="heading-165">Support Me</h3>If you like what I do, want to keep me awake at night, or feel like generally showing support to a fellow creator, click the link below (should open in a new tab):<a target="_blank" rel="ugc" href="https://ko-fi.com/ai_ryan_ai">Feed me Caffeine! @ Kofi</a><h3 id="heading-3795">This LoRA applies a grungy cyber angel effect.</h3>I've tested it on 5 different models, ranging from realistic, semi realistic, cartoony, anime.<h3 id="heading-2589">---Start Changelog---</h3>Time to add a changelog. Starting today, 27/10/2023, I will add notes regarding the progress of this LoRA.Previous details (below this changelog section) refer to V1.0a. Anything new, related to the upcoming versions will be noted in this changelog.<ul><li>I have migrated to ComfyUI at this stage, and am still in the process of figuring it out, however it seems to be allowing a modicum of more control over the generation. That being said, previous information (below) was based on StableDiffusion WebUI. I'd imagine a percentage of interchangeability between different UIs when it comes to settings, however I've noted that things work slightly different in ComfyUI and therefore some things will need to be adjusted to suit your specific UI. (27/10/2023)</li></ul><ul><li>V1.0b (27/10/2023)<ul><li>This version is unreleased at this stage</li><li>Currently in the process of generating images to train the new version.<ul><li>Generating poses: Standing/Sitting/Kneeling<ul><li>With the current LoRA, it seems I cannot generate other poses (probably because original version was trained on images that were almost exclusively in the standing pose, with arms by side)</li></ul></li><li>Generating camera angles: Front view, Side view<ul><li>Also seems, with the current LoRA, that I cannot generate other camera angles (also likely due to original limited training data)</li></ul></li><li>Generating on top of a variety of models, for realism, semi realism, anime, etc.<ul><li>Current models generated (on top of): (see Credits at bottom)<ul><li>Animerge V2.2 (currently trained at strength 1.5 aka "mid_strength")<ul><li>Portrait aka Face only (currently none, but intend to do 10x of each as per below)</li><li>Close-up (10x Standing Front View, 10x Standing Side View, 10x Sitting Front View, 10x Sitting Side View, 10x Kneeling Front View, 10x Kneeling Side View)</li><li>Full Body (10x Standing Front View, 10x Standing Side View, 10x Sitting Front View, 10x Sitting Side View, 10x Kneeling Front View, 10x Kneeling Side View)</li></ul></li><li>RealCartoon-Realistic V8 (currently trained at strength 1.5 aka "mid_strength")<ul><li>Portrait aka Face only (currently none, but intend to do 10x of each as per below)</li><li>Close-up (10x Standing Front View, 10x Standing Side View, 10x Sitting Front View, 10x Sitting Side View, 10x Kneeling Front View, 10x Kneeling Side View)</li><li>Full Body (10x Standing Front View, 10x Standing Side View, 10x Sitting Front View, 10x Sitting Side View, 10x Kneeling Front View, 10x Kneeling Side View)</li></ul></li><li>no others yet, will add more to this list as I generate training dataset samples</li></ul></li></ul></li></ul></li><li>Goal of this new version of this LoRA is to be able to generate more reliably on top of differing styles. Also removing unwanted artifacts (hopefully) such as extra faces/heads/multiple torsos/multiple sets of breasts/etc.</li><li>I intend to try also generate varying strengths of my training images (tagging as "low_strength", "medium_strength", and "high_strength") on top of each model used for sample image generation. Whilst LoRA strength is controllable by adjusting the strength value, I also want to be able to control it a touch more. Unsure as to whether this will actually work or not. We'll see.</li></ul></li><li>V2.0 (27/10/2023)<ul><li>This version does not exist yet, and is but a glimmer in my eye...</li><li>This version is intended to be the first "proper" release, after the alpha and beta releases of V1.0</li><li>Intended additions:<ul><li>Allow for different ethnicity</li><li>Allow for different hair colour</li><li>Allow for different base body type/shape</li><li>We'll see... Any other suggestions, appreciated... just drop a comment.</li></ul></li></ul></li></ul><h3 id="heading-2590">---End Changelog---</h3><h3 id="heading-3796">CFG Scale</h3>Cfg scale 5-7 seems to be fine, minor differences (as shown by the images).<h3 id="heading-3797">LoRA Strength</h3>LoRA strength values between 0.5-2.0 were tested. My testing suggests that the lower you go, the more of the original model style is visible. The higher you go, the more of this effect and overriding of the original model style. (See example images).<h3 id="heading-3798">Keyword &amp; Prompt</h3>I would suggest at the bare minimum the keyword of CyberAngel. All my example images were tested on the prompt:<ul><li>1girl, full body, CyberAngel, {standing|sitting}</li><li>LoRA strength values range from 0.5-2.0 for my sample images</li></ul><h3 id="heading-3799">Upscaler</h3>I am using 4xUltrasharp_4xUltrasharpV10 upscaler, with denoise strength of 0.7 (default) on all sample images (see <a target="_blank" rel="ugc" href="https://civarchive.com/models/116225">https://civarchive.com/models/116225</a>).NOTE: Be wary with your Denoise Value on 4x Ultrasharp when combining multiple LoRA together.<ul><li>I have been testing with this LoRA + 3 others &amp; found that Denoise value above 0.5 breaks the results. With just 2 other LoRA applied, Denoise value above 0.6 breaks the results. So it seems (just a theory at this stage) that for every LoRA I add on top the Denoise value needs to decrease. I'm guessing that all the "noise" in the image generated by these extra LoRA I am using are causing the Upscaler to want to really "clean" up the image.<ul><li>I have now tested with up to 6 LoRA applied to the same seed based image. No issues with Denoise Value of 0.5</li><li>I suggest, for just this LoRA applied, a Denoise Value of 0.7 is suitable.</li><li>For 2 LoRA applied, Denoise Value of 0.5-0.6 (may depend on your use case)</li><li>For 3 or more LoRA applied, best to be safe and use Denoise Value of 0.5.</li><li>Any observations or notes, let me know in the comments.</li></ul></li></ul><h3 id="heading-3800">Negative Inversions</h3><ul><li>negative_hand-neg easynegative<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/56519/negativehand-negative-embedding">https://civarchive.com/models/56519/negativehand-negative-embedding</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/7808?modelVersionId=9208">https://civarchive.com/models/7808?modelVersionId=9208</a></li></ul></li></ul><h3 id="heading-3801">VAE</h3>No VAE set at all in my SD webui.<h3 id="heading-3802">Version</h3>This LoRA was built on SD1.5 in CivitAi training section. I decided to use this to hopefully allow it to be more universal in its use case. Not sure if this was a wise choice or not. This is currently v1.0a. I intend to improve upon this LoRA over time, although may go straight to SDXL for the next version (advice on whether this is a good option or not is welcome).I have not yet tested in conjunction with other LoRA. Please upload if you test this and let us all see what happens.Any and all feedback is highly appreciated. Feel free to ask questions and share tips. If you like my work, please give it a ❤️ and I'm happy to take Buzz tips (so I can use to create v1.0b). If there are any suggestions for anything extra to add to this information page, let me know in the comments.<h3 id="heading-3803">Credits</h3>Models used to create my samples and tests were as follows:<ul><li>Animerge V2.2 <a target="_blank" rel="ugc" href="https://civarchive.com/models/144249/animerge">https://civarchive.com/models/144249/animerge</a></li><li>Flat-2D Animerge V3.0 <a target="_blank" rel="ugc" href="https://civarchive.com/models/35960/flat-2d-animerge">https://civarchive.com/models/35960/flat-2d-animerge</a></li><li>epiCRealism Natural Sin RC1 VAE <a target="_blank" rel="ugc" href="https://civarchive.com/models/25694?modelVersionId=143906">https://civarchive.com/models/25694</a></li><li>RealCartoon-Realistic V8 <a target="_blank" rel="ugc" href="https://civarchive.com/models/97744?modelVersionId=161244">https://civarchive.com/models/97744</a></li><li>SeekYou - Mikan <a target="_blank" rel="ugc" href="https://civarchive.com/models/60572?modelVersionId=74057">https://civarchive.com/models/60572</a></li></ul>

Cyber Angel

V1.0a. Totally in production stage. Just testing to learn about making LoRA and how to get it to do what I want.Applies a grungy cyber angel effect.

Files with SHA256: 3b5eea966fdb8375626ccae21c199f44246221b8b24c890ffa960a83f8d86bec