The LHC (Large Heap o' Chuubas) is aiming to be the model for all of your VTuber needs. There are other minor goals, like improving aesthetics, backgrounds and anatomy, but the main goal is to offer a lora-less option for generating VTubers.<h1 id="alpha-v0.5">Alpha V0.5</h1>LHC v-pred v0.5 is a custom finetune of NoobAI v-pred 1.0. It has improved image quality and retains nearly all the inherent artist knowledge of NoobAI, while massively expanding it's knowledge of vtubers.Not all vtubers are known to the same level, some need additional tags, or don't work well at all from my testing. A full list of characters, with notes on tagging and usage, can be found here: <a target="_blank" rel="ugc" href="https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha_v05/vtubers_valpha05.txt">https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha_v05/vtubers_valpha05.txt</a> . Additionally, an album of example images of every single vtuber (including ComfyUI metadata) has been prepared here: <a target="_blank" rel="ugc" href="https://catbox.moe/c/pjfwt1">https://catbox.moe/c/pjfwt1</a> .<h2 id="dataset-details:">Dataset Details:</h2>All vtubers were seperated and their image counts normalized between 80 and 120, aiming for 100 whenever possible. Additionally, 1000 images of multiple vtubers were included as well to train multicharacter capabilities. This resulted in a dataset of around 16000 images. Additionally, all images were upscaled to at least 1MP, and jpegs were additionally cleaned using upscalers specialized on removing compression artifacts from webtoons.Some vtubers required large amounts of data that didn't have associated tags yet. Here, ai-tagging models were used to assist. I would have liked to completely tag every image by hand, however the size of the dataset makes this a larger undertaking than I am able to do at this point.<h2 id="training-details">Training Details</h2>LHC v 0.5 has been trained for 102 epochs at varying learning rates and batch sizes due to uncertainty with parameters after the major restructuring of the dataset. In general, Unet Learning rates were between 1.5e-5 and 5e-5, and TE learning rates between 4e-6 and 8e-6. Cosine schedulers and batch sizes between 8 and 32 were used. Exact training logs for tensorboard can be found on huggingface.Training took over 400 hours and trained for over 1.6M samples.<h1 id="alpha-v0.4">Alpha V0.4</h1>As opposed to the previous versions which used the LoKR method, v0.4 is a full finetune of Noob V-Pred 0.6. At over 340,000 samples(4500 images over 80 epochs) seen, it took nearly 90 hours of training, which does not included several experiments. Despite this, the understanding of artists and concepts is still very close to the base model.A list of characters can be found here: <a target="_blank" rel="ugc" href="https://huggingface.co/Jyrrata/LHC_XL/blob/main/characters/alpha04.txt">https://huggingface.co/Jyrrata/LHC_XL/blob/main/characters/alpha04.txt</a> . Some of these will only need the character tag, some will need additional descriptors.And the Lora Extract here: <a target="_blank" rel="ugc" href="https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha/v04/lhc_04_extract.safetensors">https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha/v04/lhc_04_extract.safetensors</a><h3 id="training-details">Training Details</h3>Dataset:A dataset of ~3500 images(4500 including repeats) was used. This includes 3 artists with a total of ~350 images, ~500 images of multiple characters and the ~2650 images of the 100 included characters.Repeats were chosen so that each character has somewhere between 30 and 50 images per epoch. Whenever possible, high quality pngs with resolutions &gt;1MP were chosen. If this was not possible, then the images were upscaled and/or cleaned using upscaling Models designed to remove jpeg artifacts from images.<h1 id="alpha-v0.3.1">Alpha V0.3.1</h1>Due to some mistakes during the training of alpha v0.3, the model has diverged significantly from NoobAI. Nonetheless, it is a capable model with good understanding of most of the 79 trained vtubers and a passable one for the rest. For an overview refer to:<a target="_blank" rel="ugc" href="https://huggingface.co/Jyrrata/LHC_XL/blob/main/characters/alpha03.txt">https://huggingface.co/Jyrrata/LHC_XL/blob/main/characters/alpha03.txt</a>and <a target="_blank" rel="ugc" href="https://civarchive.com/posts/9579061">https://civarchive.com/posts/9579061</a> for a visual guide on basic character comprehension of the two v0.3 models. Many characters work with only their activation tag, though some require a little or a lot of additional tags to work.Alpha V0.3 and V0.3.1 were trained on NoobAI-XL V-Pred-0.6 version.A lora extracted version can be found here: <a target="_blank" rel="ugc" href="https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha/v03/lhc_v03_1_lora.safetensors">https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha/v03/lhc_v03_1_lora.safetensors</a>If you want to use V0.3, it can be found here: <a target="_blank" rel="ugc" href="https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha/v03/LHC_alphav03-vpred.safetensors">https://huggingface.co/Jyrrata/LHC_XL/blob/main/alpha/v03/LHC_alphav03-vpred.safetensors</a>Additionally, there is also an eps version and a version trained on rouwei-vpred of intermediate datasets in that huggingface repo. Refer to the character .txt files for overview of the v0.2.5 knowledge.<h1 id="alpha-v0.2">Alpha V0.2</h1>Same general approach as v0.1, however the dataset has been expanded by 10 additional vtubers for a total of 28 now, and the final two epochs include an experimental dataset of 1200 images covering a wide base of concepts intended to realign and improve the model aesthetically.Included vtubers this time are:<ul><li>aradia ravencroft</li><li>bon \(vtuber\)</li><li>coni confetti</li><li>dizzy dokuro</li><li>dooby \(vtuber\)</li><li>haruka karibu</li><li>juniper actias</li><li>kogenei niko</li><li>malpha ravencroft</li><li>mamarissa</li><li>michi mochievee</li><li>rindo chihaya</li><li>rin penrose</li><li>atlas anarchy</li><li>dr.nova\(e\)</li><li>eimi isami</li><li>isaki riona</li><li>jaiden animations</li><li>juna unagi</li><li>kikirara vivi</li><li>mizumiya su</li><li>tsukinoki tirol</li><li>alias nono</li><li>biscotti \(vtuber\)</li><li>mono monet</li><li>rem kanashibari</li><li>yumi the witch</li></ul>In addition to adding new ones, the datasets for some of the old ones have been redone, especially juna unagi and juniper actias. Juniper has also gotten two new tags, juniper actias \(new design\) and juniper actias \(old design\), which tries to seperate her models into two distinct phases. This is experimental and might not be carried forward to future versions.A showcase of the base character tag understanding is <a target="_blank" rel="ugc" href="https://civarchive.com/posts/9183473?returnUrl=%2Fmodels%2F940235%2Flhc%3FmodelVersionId%3D1064798">here</a>. Some vtubers don't work with only their character tag, instead you will need additional descriptive tags.<h1 id="alpha-v0.1">Alpha V0.1</h1>This model is currently still in alpha. The current state is not indicative of all future capabilities, but rather just a proof of concept.A basic test model, with nice results nonetheless. Trained on roughly 1000 images featuring mostly 18 vtubers that the base NoobAI model did not know well. This model is based on the NoobAIXL v-pred-0.5-version model.As a V-pred model, this model will not work in all WebUIs, but only those that have implemented vpred sampling. The necessary state dicts of the model have been set for UIs like Comfy and ReForge to set the required settings automatically. If not, it is necessary to activate v-pred sampling and it is recommended to turn on ztsnr as well.The newly added/enhanced vtubers are (listed by their trained tags):<ul><li>Aradia Ravencroft</li><li>Malpha Ravencroft</li><li>Mamarissa</li><li>Koganei Niko</li><li>Rindo Chihaya</li><li>Mizumiya Su</li><li>Isaki Riona</li><li>Kikirara Vivi</li><li>Coni Confetti</li><li>Dizzy Dokuro</li><li>Dooby (Vtuber)</li><li>Haruka Karibu</li><li>Juna Unagi</li><li>Juniper Actias</li><li>Michi Mochievee</li><li>Rin Penrose</li><li>Jaiden Animations</li></ul>Additionally included were especially Nerissa Ravencroft and Vienna (Vtuber), as well as many images featuring 2 or more characters at once.For a showcase of the base character comprehension, check out <a target="_blank" rel="ugc" href="https://civarchive.com/posts/9039071">this post</a>.<h3 id="recommended-settings:">Recommended Settings:</h3>Sampler: EulerCFG: 4-5Steps: 25+<h3 id="training-details:">Training Details:</h3>Trained as a full dimension LoKr, based on the methodology of the <a target="_blank" rel="ugc" href="https://civarchive.com/models/332076?modelVersionId=372033">KohakuXL </a>series, with the Lycoris settings found <a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf/Kohaku-XL-Delta/blob/main/docs/lyco_config.toml">here</a>.Specific parameters:<ul><li>Dataset: 1035 images</li><li>Batchsize: 2</li><li>Gradient Accumulation: 4</li><li>Training steps: ~6400</li><li>Training Epochs: ~50</li><li>Unet LR: 3e-5 (lowered to 2e-5 for the last 12 epochs)</li><li>TE LR: 2e-5 (lowered to 1e-5 for the last 12 epochs)</li><li>Optimizer: AdamW 8-bit</li><li>Constant scheduler</li></ul><h3 id="special-thanks:">Special Thanks:</h3>kblueleaf (Kohaku Blueleaf): for the Lycoris library and the resources on finetuning via LoKrOnomaAI &amp; Laxhar Dream Lab: for amazing base modelskohya-ss: for sd-scripts

Full finetune of Noob v-pred0.6. No alternate training methods like Lokr were used.

lhc_vPredAlpha04.safetensors

Files with SHA256: 79408c80ce4ceb65dec9318a7c92b34bba1adb388aa4526018e61f8460b54399