CivArchive
    SamDoesArts (Sam Yang) Style LoRA - offset
    NSFW
    Preview 73379
    Preview 126879
    Preview 126881
    Preview 73494
    Preview 76023
    Preview 73362
    Preview 73360
    Preview 708301
    Preview 73361
    Preview 73358
    Preview 73356
    Preview 73363
    Preview 73408
    Preview 126880

    SamDoesArts (Sam Yang) style LoRA

    Making models can be expensive. Do you like what I do? Consider supporting me on Patreon 🅿️ or feel free to buy me a coffee ☕

    There are already countless models for this artstyle, but I still wanted to test out how LoRA's can handle it. I also wanted to experiment with AnyPastel (or Anime Pastel Dream - Soft). This LoRA didn't come out super strong, so it has the right flexibility to get a style that's closer to Sam Yang's on AnyPastel, while still reminding of him on other models. Check the examples (mainly the model hash) to get what I mean. Triggers with sam yang and I suggest AnyPastel as base model (or, if you don't care about having a more matching artstyle, AnyLoRA or AbyssOrangeMix2). Use weight=1 for the offset version or 0.65 for the original (check "original" tab for a comparison).

    The pic of Melina is also using my Melina LoRA.

    Pic 3 and last 4 ones (all the realistic ones) have been made by kaynite using NED.


    How to use LoRA's in auto1111:

    • Update webui (use git pull like here or redownload it)

    • Copy the file to stable-diffusion-webui/models/lora

    • Select your LoRA like in this video

    • Make sure to change the weight according to the instructions (by default it's :1)

    Description

    FAQ

    Comments (35)

    1nnocenceFeb 4, 2023
    CivitAI

    Is that a Melina from Elden Ring LoRa? If so, could you perhaps release it as well?

    Great stuff as always!

    Lykon
    Author
    Feb 4, 2023· 1 reaction

    nope, I simply described Melina a bit. I'll do a Melina LoRA tho.

    1nnocenceFeb 4, 2023

    @Lykon Okay, great news! Looking forward to it.

    thms2307Feb 4, 2023
    CivitAI

    Your styles are some of the best I've seen so far. Just wanted to suggest SunsetRiders7 style since I haven't seen anything like that yet. Good stuff though

    Lykon
    Author
    Feb 4, 2023· 1 reaction

    It seems that he's replicating the WB cartoon style. I'd rather do that directly.

    thms2307Feb 4, 2023· 1 reaction

    @Lykon Yeah honestly that's cool too. It's basically the same

    brunogcarFeb 4, 2023
    CivitAI

    outstanding job, really well done not just in this one, but overall, specially in the descriptions, telling checkpoints used, example prompts etc

    since you redid this, IMO better than other attempts, may I suggest what can be considered by now, classic models like: archer, arcane, disney classic and modern, etc

    https://huggingface.co/nitrosocke

    not a request, just a suggestion,

    keep up with the excellent job, lots of pleople appreciate what you do, thanks

    Lykon
    Author
    Feb 4, 2023· 1 reaction

    I might do Arcane in the future. It would be interesting as a LoRA as they're more easily combined on top of other models and styles.

    168550Feb 4, 2023
    CivitAI

    I love your stuff and I've been inspired to try to create my own embeddings.

    With that in mind, how can I make sure it only takes the artist's style and not the characters in the training photos?

    Lykon
    Author
    Feb 4, 2023· 3 reactions

    different characters + 1 style -> gonna learn the style
    different styles + 1 character -> gonna learn the character
    1 character + 1 style -> gonna learn both (or none)

    168550Feb 4, 2023

    @Lykon I see, thank you!

    I'll upload it if i manage to get something I'm satisfied with.

    168550Feb 5, 2023· 2 reactions

    @Lykon It worked like a charm. Thank you very much!

    ORIHARA Lora coming soon 👍

    JustAiGuyMay 27, 2023

    @Lykon sorry for bothering you, but how do you caption the style? Are there any rules? I did LORA of a few characters and now want to try the style. Thank you for your great work!

    P.s. also how many pictures do I need? If there are a thousand works of an artist, for example?

    Lykon
    Author
    May 27, 2023

    @JustAiGuy training a style should not be much different than training a character. You just need a dataset that's consistently of 1 style and has multiple subjects. For the dataset size, the more you have the better.

    JustAiGuyMay 27, 2023

    @Lykon @Lykon Thank you!

    But what about captions(tags)? Do i need to do captions (tags) in the same way, as with Lora for a character? 1 male, black suit, dark room, etc.? And also i heard that lycoris is better for styles. There is any reason to train lycoris instead of lora?..

    Lykon
    Author
    May 28, 2023

    @JustAiGuy lycoris is still lora, just using unconventional methods for parametrization. Yeah tag all the things or use wd tagger and some manual control

    JustAiGuyMay 28, 2023

    @Lykon hm...I did both. Lycoris seems different to me. Lycoris is somehow stronger than lora and more aggressive even with low weight. Also, it always does a strange stylization, and strange colors, with the same dataset.

    Lykon
    Author
    May 28, 2023

    @JustAiGuy they're hard to compare in general. Weights can be offset easily and training results change even with the same type of lora if you try multiple executions.

    A1988Feb 5, 2023
    CivitAI

    Nice model.

    Do you have any tutorial to do Lora training?

    168550Feb 5, 2023

    I just followed AItrepeneur's tutorial on youtube and it worked like a charm.

    ElThunasFBI12Feb 6, 2023
    CivitAI

    Very good VAE, could you do a VAE of Highschool dxd?

    Lykon
    Author
    Feb 6, 2023

    You probably meant LoRA. Trauter has made Rias and another couple of characters from that

    ElThunasFBI12Feb 6, 2023

    @Lykon Where did I upload it? I can't find it in Civitai or Huggingface

    Lykon
    Author
    Feb 6, 2023

    @ElThunasFBI12 no idea but if he uploaded it then it's on his HF.

    VulcanFeb 7, 2023
    CivitAI

    hi guys, wich ckpt model are you using this Lora with ? It didnt work for me on those models :

    -SD1.5- Anything v4.5 - Abyssorangemix - Anypastel - PastelMix -

    Lykon
    Author
    Feb 7, 2023

    Should work on all of them, especially AnyPastel.

    VulcanFeb 8, 2023

    @Lykon yes yes it does work, there is definitely a difference, but not in the style of samdoesart, can someone plz tell the name of the models used in the example images above ?

    VulcanFeb 8, 2023

    @Lykon I am using a google colab notebook, I Can add the files to the lora folder and it shows in my Lora tab in A1111, but cant tell if that's working with trigger words

    Lykon
    Author
    Feb 8, 2023

    @Vulcan you should see the model hash in the generation data of the example images. The first one is using AnyPastel. The trigger words do something but not much when it comes to style. You're probably missing my negative prompt embeddings, but that's not super relavant, you should still et the correct style. By any chance, are you using the lora filename from my examples? Because civitai changes the file name on download, so you'll have to use yours.

    VulcanFeb 8, 2023· 1 reaction

    @Lykon thank you all of what you said was correct, though my biggest mistake was that the lora file did not upload fully on the google colab hardware, I managed to write some additional code on the notebook to transfer the Lora file it directly from google drive servers to google colab ==> it works now perfectly, even though I dont understand how you get so crisp color images : mine are more pastel like palette, pale and shy colors , but the style is applied coorectly

    allehopFeb 11, 2023
    CivitAI

    Seriously love the work, but I have a pretty noobish question. Which upscaler did you use and where can I get it? I am trying to replicate just the first image, but AnyPastel is a little too present and the colours look all washed out

    Lykon
    Author
    Feb 12, 2023· 1 reaction

    Anynupscqler that's not Latent will get the job done. If you don't see the upscaler in the generation data it means I just used img2img at higher resolution and 0.4 denoise or lower (which is basically nearest upscaler but without having to regenerate the image)

    imoralFeb 13, 2023
    CivitAI

    The art looks amazing and i rlly want to try it out but after i downloaded the file i put the file in the models folder in the stable diffusion folder, after doing that i launched stable diffusion and chose samdoesartsSamYang_normal.safetensors form the model list and it doest work i used the generation data and it didnt work it just gave me a picture of a realistic looking bird, can someone please help.

    Lykon
    Author
    Feb 13, 2023

    it's a LoRA, so you have to use it from the models/Lora folder and select it with a weight (don't use the default 1). There is a short guide I added to the description. Basically this is a "pluggable" fine tune that you can use on any model. I advise you use this on AnyPastel for accuracy, but you can use it on most models and it will do something.

    imoralFeb 13, 2023· 2 reactions

    @Lykon  oh ok thanks it worked results are amazing btw.

    LORA
    SD 1.5
    by Lykon
    Download (Beta) View on CivitAI
    girls
    artstyle
    anime
    girl
    style
    digital illustration
    women
    samdoesarts
    sam yang
    art style
    woman

    Details

    Downloads
    71,882
    Platform
    CivitAI
    Platform Status
    Available
    Created
    2/4/2023
    Updated
    5/17/2026
    Deleted
    -
    Trigger Words:
    sam yang

    Files

    sam_yang_offset_right_filesize.safetensors

    Size:
    144.11 MB
    SHA256:
    b67fff7a42e85faf612bc6635c3915645112ba3692a8124d7f4e822c97947657

    Mirrors

    HuggingFace (63 mirrors)
    CivitAI (1 mirrors)
    CivitasBay (1 mirrors)
    TensorArt (1 mirrors)
    TensorHub (1 mirrors)

    Available On (1 platform)

    Same model published on other platforms. May have additional downloads or version variants.

    CivitasBay
    SamDoesArts (Sam Yang) Style LoRA - offset