Hello...
This is a LoRA for Tingyun character from Honkai Star Rail, trained in Anythingv4.5 for 20 epochs with 20 images * 10 repeats,
its match the character pretty well with model like anything or any anime models, for more information you can see the image previews.
for the weight i recommended using around 0.6-0.8.
as for the trigger words, i think you can use
tingyun hsr, brown hair, ponytail, fox ears, green eyes, dress, animal_ears, earrings, detached_sleeves, bracelet, animal_ear_fluff, fox_ears, sandalsi guess that is all, thanks you....
Edit:
So i forgot some images that i use for the preview using control net when looking for pose or propotions, Here some guide to use ControlNet
Go towards extensions tabs and install this extension, https://github.com/Mikubill/sd-webui-controlnet
Download ControlNet model (https://huggingface.co/lllyasviel/ControlNet/tree/main/models) (recomended the openpose one) and place it inside
sd-webui-controlnet/modelsextension folderAfter installing, a tab named Control Net should be appeared in txt2img or img2img
Open the tab and use this image as a guidelines, pretty much use openpose preprocessor and openpose model then place the image that you want to use inside the tab.
Sometimes this method not working, you can use openpose stickman instead, here my collection of openpose stickman (set preprocessor to none like this)
Then just generate images like usual, just leave the ControlNet enabled. the image should resemble the pose
ps im just gonna leave some of this here :>
Description
FAQ
Comments (13)
Which model did you used for the preview images? The style is very pretty in terms of the colors
andite latte model, hmm i think you can find that in https://huggingface.co/andite/yohan-diffusion/blob/main/Latte.safetensors
@Zipsy I am honestly curious how you got those results the LoRa doesn't seem to work on my end neither when I used the metadata of your images did you do anything specific here?
@Freshii wait which image do you mean? Some of it i used control net for the propotions. like the landscapes one and the 1st image.
@Freshii update some guidelines for using ControlNet, and i just realize some of the img metadata using img2img metadata for fixing hand. uploaded some original image from txt2img
Images just seem impossible to replicate, definitely not an easy to use Lora. Not going to review poorly under the assumption that It's my fault but yeah... This Lora definitely doesn't work without Img2Img or other tricks and even then i struggle to create anything good looking.
Yeah sorry, i used controlnet and img2img (for hand fixing) a lot for these img... And this img is old, im also sometime has difficulty to reproduce the same img nowadays because of SD extension updates
I wanted to ask doesn't work as you cant reproduce the same img? Or you just cant generate tingyun (the character) at all? Because if you really want 1:1 copy of the images, yes it difficult now :)
@Zipsy @Zipsy I could produce a character resembling Tingyun but it would pretty much always be with an outfit that had the waist revealed, and the quality of the image would often be very low. I didn't expect to replicate the images here, theres usually deviations between set ups and versions but it was just nowhere near it even with the suggested Checkpoints.
@Ixalia hmm im also kinda forgot how i produce these img tbh xD. as far i remember im just using control net for some images because i want to those specific pose. other than that, probably a lot of inpainting for fixing hands and maybe its VAE since im using WD VAE.
Very rigid LoRA, extreme difficulty to make anything that is not her normal outfit and hard to do a lot of poses even with other LoRAs trying to help it along. 20 similar images just aren't enough for a high quality LoRA
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.