LoRA trained on EDG Zoe from League of Legends
requested by: @yoooo_
tags: edg zoe, 1girl, heterochromia, short dress, black dress, pantyhose, very long hair, fingerless gloves
models used
anime: MeinaPastel - V5, YesMix - v2.0
2,5D: PastelDiffusedMix - V2.1, LusterMix - v1.5, NijiDiffusedMix - v2.0, The Agenda Mix - 2.0
realistic: Paragon V1.0
better hands LoRA: EnvyBetterHands LoCon - beta2
negative embeddings: Fast Negative Embedding (+ FastNegativeV2) - v2
Recommended settings:
Make sure you are aware on the usage instructions of LORA
VAE: vae-ft-mse-840000-ema-pruned (for realistic models & 2,5D), kl-f8-anime2, blessed2.vae (for anime models)
Sampler: DPM++ SDE Karras or DPM++ 2M Karras(euler a for img2img upscaling) from time to time im also using DPM++ 2M alt Karras
Steps: 20 to 35
Clipskip: 2
Upscaler : GitHub - pkuliyi2015/multidiffusion-upscaler-for-automatic1111 4x-UltraSharp(for 2,5D & realistic), 4xAnimesharp(for anime)
CFG Scale : 5 to 9 (mostly 7)
LORA weight for txt2img: 0.5-1 (for the examples i mostly used 0.8 and 1)
Denoise strength upscaling: 0.3 to 0.55
If you want to autocomplete the tags & trigger im using, I recommend using Booru style tag autocompletion for AUTOMATIC1111's Stable Diffusion web UI
For full body shots i recommend using: ComfyUI Impact Pack
or the a1111 alternative GitHub - Bing-su/adetailer
Description
FAQ
Comments (17)
She's cute :3
can you do yuichiro & shinoa from ONS?
Sure, there is one Shinoa on this site, but it's a bit overcooked, so I'll do both :)
looks great, great job! <3
Thanks! Glad you like it :)
OMFGSIDOJFSOIDJFSDIOFJS bows to your greatness You recommended one of my models XD
Of course! It's a great model for creating cute characters. It's quite unique as well. Most of the time, I can easily identify the models that were used for the merge, but not on yours somehow xD
@AhriMain Mine's a ... well LAUGHS I've been learning lora smooshing and Agenda mix was that - It had Niji Vkei, Niji Slime and Niji Pride in it plus a couple other stableizations - actually it was more that I had a MODEL i have on HF I made with those three loras XD that's FUN to keep but not quite sure it's -- I might upload it anyways even if it's not lora compatible (I have a friend who yells if my shit isnt' lora compatible XD)
Supeeer cute model <3! I wanted to ask you, In upscaling, do you use Hires. fix or img2img?, and how many steps? i read that i should use half number steps of sampling steps and upscale by : x1.5 cuz x2 it is not supported by the majority of the models :c
Ty! Img2img with Tiled Diffusion & 4xUltra-sharp i upscale it by 2x. Typically, I apply similar steps to the img2img process. When using the hiresfix technique, I suggest using 15 hiressteps along with an upscaler such as "esrgan anime+6b" or "4xultra-sharp
@AhriMain Thanks a lot! I really love this model, but i have two more things , about the results in Hires. fix and img 2 img. Are they the same in quality or slightly different? , by that what method of upscale do you recommend? (PD: i have low VRAM around 2gb's and 20 gb's of RAM) and 2) i was using the model now for a long time, I don't have the generation results that you have :c, moreover, most of the eyes are generated poorly drawn or malformed, even so I used negative prompts like : poorly drawn face, poorly drawn eyes, malformed eyes... among other similar (including the negative embeddings you suggested to download, However, when I use one of your seeds from your images that you have posted, I do have more beautiful results.)
@gabriel1196 I think both can achieve pretty much the same. Img2img with Tiled Diffusion should be more budget friendly though, cause it tiles the image into 8 pieces. Here are some example settings for your gpu. Make sure to enable the Tiled VAE as well, this will save some VRAM (you don't have to adjust the settings, its automatically adjusted to your graphics card). When it comes to faces, there's a simple solution. In txt2img, I use Adetailer, which results in a perfect face 99% of the time. Especially useful for full body or cowboy shot images. If the Tiled Diffusion settings is still to heavy for your gpu maybe use this
@AhriMain Thank you so much!! sorry, I'm kinda new to this, also if I want to get more details in my image, I should use upscale, right? Whatever it is I think, to be more specific my graphics card is an NVIDIA 940mx, and what if I use the option "resize by" and using the multiplier by 1.5x and the upscaler factor at the same value? instead of using the option "resize to" and about tiled VAE, im using the VAE's (vae-ft-mse-840000-ema-pruned) you recommended, it changes something in them? (im using img2img upscale, it seems more easy ) Thank you again...
@gabriel1196 Np, Tiled VAE is the tap under Tiled Diffusion, its used to tile the VAE decoder process (the process that comes right after you have 100%) into 8 pieces. But yea this VAE also helps, cause its only 326mb. Upscaling is useful for enhancing, if you go up with the denoise strength you also get more details, but then the image gets different from the orginal. Therefore, it is important to find a balance and for me thats 0.35-0.5 like in the example i showed. You dont need to change the size of the image, because the Tiled Diffusion setting already does that for you, just leave it as is 512x768 or what ever you use in the situation :)
@AhriMain Thaaaaanks a lot! well i was trying to upscale but i always have the error : CUDAoutofmemory, i know it is caused by my graphics card, well i was learning now that i should use the argument"--xformers" (im already using the arg: --lowvram) to improve the speed and VRAM usage, but i think that using --xformers will cause a noticeable variation on quality, prompts, and generation of zoe edg lora ;-; , and i wont have some amazing results like your post....or well that's what i know, also will be more inexactly if i use tiling option in img2img upscale?, do you recommend me using that?
@gabriel1196 I'm using xformers as well. Xformers doesnt really reduce the quality from my understanding. The only downside of xformers is that it's more challenging to replicate something created by others. 2GB VRAM is really though, try this ultra low setting, maybe this will work, if not then unlucky
Thanks a lot for this amazing model, hope you are fine, I want to ask you if you plan to train a Zoe EDG model in the same way, but in Stable Diffusion XL?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.