<h1 id="melita:-a4lj2qqr5">Melita:</h1>Doing some simple basic testing with a character turnaround. Melita's design is really simple so I'm not exactly looking for high consistency with her and wanted to have some sort of baseline for loopback testing. Unfortunately, I wasn't very successful so it's not really worthwhile writing and article for. I'll probably need to design a character with more complex details later for more stress testing. It seems for Illustrious knowing the proper prompts/artist keywords seem to help with the posing. The dataset is only a limited turnaround so the posing is very stiff. The anytest CN can help with adding dimensionality to the character which I think is only possible as her design is really straightforward.<h1 id="prisca-gn7kz5ycv">Prisca</h1>A complex character design. Images are very cherrypicked on the basis of the outfit but the LoRA performs very well in terms of inpainting. More overtrained compared to the other LoRAs due to loopback usage.<h2 id="elris-c7wz4aw9f">Elris</h2>A rather simple character design this time to test out a couple of things regarding accessory training, and clustered objects along with experimenting with loopback.3/4/24 Update:Blind Test with SDXL training. I'll write an article about this later. Read 'About this Model' for more details.Kie recommended weight:1I am probably going to need to change the title of the page now since I am using txt2img ControlNets to augment the dataset for style. (I did use a LoRA to loopback but I used weight of .1~.2 and outd to weaken the effect. I did another round without the LoRA and the difference wasn't very noticable). Read below if you want see the specifics of the augmentation.<ul><li>Enna LoRA focuses on a simpler design to see what mistakes in the turnaround make an impact on the LoRA design</li></ul><h1 id="about-6b4n7kyn5">About</h1>This is a place where I keep my tests with single image OC LoRAs. The goal is to try and create a consistent character LoRA with a minimum amount of effort. If you're willing to spend time, you can create a consistent character in this <a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=iAhqMzgiHVw">tutorial here</a> (not mine). Keep in mind that these experiments are more of a learning effort rather than producing an accurate LoRA. Due to the limited dataset, these LoRA are highly likely overfit in weird scenarios that I haven't tested yet. I'll probably write an article about the process in the future.<h1 id="articles-on-workflow-bcfig50l6">Articles on Workflow</h1><ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/articles/903/training-a-lora-with-a-single-base-image-lora-part-1">Part 1</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/articles/1058/single-image-lora-part-2">Part 2</a></li></ul>Images are of-course cherry picked with adetailer and hi-res fix with (AnimeSharp4x\UltraSharp4x) at 0.2 ~ 0.5 denoising strengths.Isabella's original image is the second image in V1. (Base prompt might be useful)<h1 id="usage-dq8pgxzxc">Usage</h1>Recommended Weight for V4: (0.5~0.8) Block Weight can help but isn't completely necessary.Recommended Weight: 0.6 ~ 0.8 + (LoRA Block Weight with OUTD preset recommended; For version 3, the block weight extension is highly recommended).For Enna: Recommended Weight: 0.8 ~ 1 (Lower if encounter concept bleeding)Use terms such as: flat_chest,petite to get character proportions correct.You can get away with 1 weight with the block weight extension but will need to lower the weight if you want do custom outfits. Requires long prompts that describe that character and the environment to work well.It should work well enough with anime-based models, I haven't tried anything realistic or semi-realistic. Probably best used with an inpainting workflow.<h1 id="problems-rl3cmgjas">Problems</h1><ul><li>Inconsistency with coloring and outfit</li><li>facial expression is overfit</li><li>Hairline is occasionally flipped</li><li>V2 is probably better than V3; I ran into a roadblock but decided to publish my results for now</li></ul><h1 id="training-7m9mf3vbm">Training</h1>V1<ul><li>1 Full body shot +1 crop image of face (same image)</li><li>Trained on NovelAi</li></ul>V2<ul><li>Reused dataset from V1</li><li>Added upperbody,lowerbody, and leggings crop</li><li>lowered repeats</li><li>more captioning</li></ul>V3<ul><li>removed flip augmentation</li><li>reused data set from v1</li><li>added 3/4 images of the head and skirt, boots to neck,close-up of bangs</li><li>more captioning</li></ul>Others<ul><li>I likely hit a wall with Isabella so I might not be able to improve the LoRA any further. There are a few more things I could try with kohyaSS trainer but I'm not expecting much. The main limitation is that the base image isn't very high quality and has image artifacts on zoom up.</li></ul>V4<ul><li>reused data from v3</li><li>moved regularization over into the main dataset and added trigger words for style</li><li>added style images using txt2img with canny and lineart using regional prompter to augment the dataset for style, tagged as 'alternativecostume'<ul><li>selected images with closest colors</li></ul></li><li>added style and outfit triggerwords</li><li>n tokens increased to 4</li><li>removed block weight during training</li></ul><h1 id="enna-03soqmjds">Enna</h1>V1<ul><li>Based on a character turnaround</li><li>Uses same concepts from V4</li></ul>V1.1<ul><li>minor update with more captioning and lower training rates by 1/10th</li></ul>V2<ul><li>Added full turnaround of swimsuit and underwear</li></ul>

OC LoRA Tests

This is a rather simple character to only test how LoRA training affects headwear, neckwear and clustered objects. Recommended Strength:<ul><li>0.8 ~ 1 for general usage</li><li>0.5 or use fewer keywords if you want to change the bikini type</li></ul>Dataset:<ul><li>Original base image at 4096 x 11040 resolution<ul><li>image cropped into multiple smaller images for more detail</li><li>used faceless/face censor augmentation</li></ul></li><li>Style Augmentation for reducing style bleed</li><li>Loopback training to help with outfit swapping</li></ul>Observations:<ul><li>Necklace and number pattern design was earliest than expected</li><li>Getting the flower next to the frills was harder than initially expected and doesn't generate all of the time</li><li>Face trains quickly</li></ul>Notable Changes<ul><li>Trained with prodigy optimizer</li><li>Doesn't require a trigger word<ul><li>Trigger word for Elris is included along with captions for numberbikini and alternativeoutfit exist. These keywords don't seem to have much of an impact but numberbikini can help with the success rate of the flower appearing</li></ul></li><li>Loopback images included<ul><li>This is mostly to help with outfit removability. Added multiple outfits + nude images</li></ul></li><li>Requires specific negatives such as head_out_of_frame, faceless, censored, identity_censor,face_censor in negative prompt</li></ul>Issues/Difficulty:<ul><li>Posing is stiff and/or requires prompt emphasis</li><li>Having separate wings is difficult</li><li>LoRA likes to place a lot of extra moles</li><li>Style Bleeding but can be mitigated with artist keywords</li><li>Face size is overtrained</li></ul>

Files with SHA256: 3f92ab8be217831af382e66e233f6dfc8c65a8da96a6e17e92ea890eab519f43