CivArchive
    Preview 2291448
    Preview 2271991
    Preview 2282111
    Preview 2271984
    Preview 2271979
    Preview 2282112
    Preview 2271982
    Preview 2271981
    Preview 2271980
    Preview 2271988
    Preview 2271989
    Preview 2271985
    Preview 2271986
    Preview 2298957

    All models are exclusive to Civitai ! Anyone who publishes my models without my consent will be reported!

    Hi all!

    SDXL_Niji_Seven is available!

    I have incredible results with my LoRa model SDXL_Niji_V6_DLC_LoRa_ v2 at strength 0.5!

    https://civarchive.com/models/541460?modelVersionId=601994

    I haven't tested v4 yet

    What's new?

    -This version is very different because it is based on NijiV6 To be more precise, around 1000 images from NijiV5 and 1600 from v6 (compared to a total of 1300 images in v6).

    -I chose not to use the "expressive" style of NijiV5 this time. But maybe I will add it in a 7.5 version in the future.

    -There is no longer a trigger word

    -Version 6 is very different so I still recommend it.

    Some tips for use:

    -Don't hesitate to play with things like: realistic, hyper realistic, anime studio, digital artwork, illustration...

    -Clip skip -2 or -3 is recommended (-3 better overall)

    -A minimum of 26 steps

    -cfg scale: between 3.5 and 6

    -Sampling method: DPMPP_SDE Karras (recommended, best quality), Euler_Ancestral simple (2nd best), DPMPP_2M_SDE Karras or ddim simple (average quality for both).

    Have fun ! 😊

    Description

    V5 : Completely new dataset for this model. 3 model that I have merged. The first at 498 images for 18000 steps, the second at 311 images for 11000 steps and the last at 237 images for 8500 steps
    V5.1 : Just a better merge

    FAQ

    Comments (25)

    sevenof9247Aug 30, 2023· 1 reaction
    CivitAI

    Hey there,

    the next person i talk to about the neg prompt being important ;)

    ok first, i don't have the pre-relase 1.6 but it seems the one is better - but you can test this prompt if you want to have realistic photos:

    painting, drawing, sketch, cartoon, anime, manga, render, CG, 3d,watermark, signature, label

    below i tested it with v1.5.2

    Stan_Katayama
    Author
    Aug 30, 2023

    Thanks for the advice. Your negative prompt looks good for photography. :)

    wuliang19862014Aug 30, 2023· 2 reactions
    CivitAI

    The V5 is too bright, brighter than the V3, and the gamma has to be pulled down a bit.

    Stan_Katayama
    Author
    Aug 30, 2023

    This is also why I released version 5.1 quite quickly. It's a little less bright.

    wuliang19862014Aug 30, 2023

    @Stan_Katayama It's not that I'm playing in 5.1, the gamma in 5.1 is so high that I can't even pull it back in post, there's no way to separate the colour levels.

    Stan_Katayama
    Author
    Aug 30, 2023

    @wuliang19862014 I think then that the problem does not come from the model. Just see the few images that others have generated so they don't have this problem.

    wuliang19862014Aug 30, 2023

    @Stan_Katayama The problem is that it exists on the model, it has something to do with your fine-tuning, gamma colour inaccuracies lead to this ~ and it's a lot of images that this is the case, I've tested a lot of them ~ they're all a bit overexposed and there's overflow of colours, it's either that or it's something to do with the samples you've been training on ~~ your training samples are probably all high saturation and high brightness images

    Translated with DeepL

    Stan_Katayama
    Author
    Aug 31, 2023

    @wuliang19862014 Have you tried without VAE? I deactivate the VAE to generate my images. I'm having problems with my 1080ti otherwise. So I do all my tests without VAE


    TwnsthAug 30, 2023· 2 reactions
    CivitAI

    Thank you for the sophisticated model upload. Looks awesome. Do keep up the good work!

    Stan_Katayama
    Author
    Aug 30, 2023· 1 reaction

    Thanks a lot 🙏

    olivettyAug 30, 2023· 1 reaction
    CivitAI

    Since 5.1 has entirely new dataset, are you training from scratch on the base SDXL 1.0 or what are you starting from? Thanks for the model btw, looks sick!! <3

    Stan_Katayama
    Author
    Aug 30, 2023

    Thank you so much ! Yes I'm starting from the base model 1.0. I actually do several training sessions and then I merge. This is the longest part! I spend several days there testing the best combinations

    olivettyAug 30, 2023

    @Stan_Katayama Sweet! Thanks for the info! Takes long but it's worth it!

    walidfluxAug 31, 2023· 3 reactions
    CivitAI

    this is the best model for sdxl i don't why people not using it more

    Stan_Katayama
    Author
    Aug 31, 2023· 2 reactions

    Thank you so much ! I don't think it's the best model. It is far from perfect! But I'm working on it 😅

    Honestly its because it was a little overcooked, had to be used on very low CFG. It looks to have improved, but the first couple of versions were less capable that the base model at producing a variety of 2D art styles, seemingly only able to produce garish overly colourful scenes with massively thick undetailed and messy linework.

    Like I said though. It has improved a lot from what I can see and I will definitely be redownloading it again and trying it out soon.

    Saying that. I still liked the original, 2nd and 3rd versions of the model. Just couldn't use them as effectively as I would have liked.

    NowhereManGoSep 16, 2023

    This is my favorite model by far, but I like to create a particular type of images, i.e., anime inspired illustrations, which is precisely what SDXL_niji_SE is designed for. I think most people outside of Asia who are not anime fans are more interested in "realistic" photo style images, and there are better models for that. Even for people who are into Anime, from what I've seen here, their taste runs more toward "lolita/moe" if not outright porn, so they will be sticking with their favorite SD1.5 anime models.

    So for anime and other type of illustrations where one wants that dynamic, somewhat exaggerated "anime/manga" feel, this is the best SDXL model by far, leaving all the other "Niji" models in the dust.

    We as fans of this model should just generate lots of quality images and hopefully that will attract more people to our camp. 😁

    The ratio of quality images I've seen here is higher than any other model page, so we got some very talented people here, and I'd love to see more!

    @NowhereManGo I'm actually an incredibly huge anime snob, who spends all of his free time watching anime nearly 24/7. If I had the money to support the industry, I'd call myself an Otaku. But unfortunately I am just an anime enthused NEET.

    I watch anime in every genre from Shounen to Shoujo, to Seinen, to Isekai, Ecchi, harem etc. etc.

    You name a genre, I will be able to recommend at least 5 shows in that genre that I have watched from memory. And I have watched thousands upon thousands of hours of anime and 10's of thousands of episodes of said anime.

    I haven't updated this page properly in nearly a decade:

    https://myanimelist.net/profile/Inuyaruto

    But here is my old anime list profile.

    @NowhereManGo I completely and 100% agree with you that - at this point in time - this is the absolute best model for genuine anime styles, which fit the actual character designs. Just as a consequence of feeding it so much data, it suffers in other areas which impact that data also.

    I am REALLY looking forwards to the next version, and subsequent versions also. However.

    NowhereManGoSep 18, 2023

    @Triple_Headed_Monkey 

    You sure watched a lot of anime. Way more than I ever did :-)

    It should be pretty obvious from my postings that I used to be a big Anime/manga fan when I was a young many years ago in the late 80s. I watched lots of anime, read lots of manga, and bought many art book. But like music and many other art form, artists needs to change and come up with new material, so trend moves into new directions. Yet, people tend to get stuck on what they enjoyed when they are younger. So today's anime has gone in a direction that I've lost touch with. My taste is pretty much stuck with older masters like Miyazaki, Shirow, Otomo, Kawajiri, etc.

    That Niji may be weaker in some area due to it specializing in Anime is quite acceptable to me. Of course, a real master like Shirow has nothing to fear from A.I. (yet) because even a model as brilliant as SDXL_niji will not fool a true Shirow fan into believing that the image is from “Intron Depot”. But what one can generate using Niji is already better than even a fairly accomplished amateur can muster when they try to emulate Shirow. That Niji is working hard behind the scene is quite clear when you compare what Niji is producing using an art style prompt such as “art by Masamune Shirow” to what base SDXL or even other “niji” can produce. As you said, only SDXL_niji looks and feels like “real anime”.

    I am certainly looking forward to the next release of SDXL_Niji as well.

    @NowhereManGo I just released my own anime merge model, made from 13 different models, with a bit more of a variety of styles. If you like Niji, you may also like that.

    It isn't anywhere near as stylized as Niji is. And not to mention that without the training that Stan has done on the niji model(s), it is likely that none of the existing anime models, my own included, would be anywhere near where they are today.

    That said, I've been watching anime religiously since 2004 when youtube came around and some of my favourite anime are the oldest and least pristine. I love Things like Yu Yu Hakusho, Ranma 1/2, DragonBall, and Ultimate Muscle. And Uresai Yatsura, the original of course, is one of the greatest pieces of animated comedy in history.

    I agree that it's a 2 fold issue. Digital is too clean, and things are drawn using tablets and mistakes etc are easily fixed.

    What needs to happen is some kind of analoguification of digital media in order to make it as palatable as older content across the board. And AI is hopefully the force to bring this back.

    NowhereManGoSep 18, 2023

    @Triple_Headed_Monkey

    I am also a big fan of UY. UY2: Beautiful Dream is still one of my favorite Anime, and one of Mamoru Oshii's most underrated masterpiece.

    Anime certainly lost some of its quirky charm when the industry moved from hand drawn and painted cells to digital, but that is inevitable. I never thought of it, but you are quite right, A.I. is now capable to emulate that hand-drawn quality, so retro style 80's and 90's anime just might make a comeback!

    walidfluxSep 2, 2023
    CivitAI

    Does the last version combines all versions? i am asking because some of the workflow of the previous versions doesn't work will with the newer once

    Stan_Katayama
    Author
    Sep 2, 2023· 2 reactions

    I wasn't happy with version 5, version 5.1 has the same basics. I just merge it in a different way. For each model, I make several models with different number of images with different number of steps.

    In reality, it seems to me that I used the same models. But I was so tired, maybe I just didn't put the models in the same order. In short, I am not very happy with either V5 or 5.1 I just made a much better version 6. I still have a lot of tests to do. But it should be here soon.

    PS: I'm not really answering your question, I spent so much time on V5. Perhaps I was no longer thinking clearly at the end. I am very disorganized in my work. I can't always explain it very well. 😅

    Checkpoint
    SDXL 1.0

    Details

    Downloads
    2,608
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/30/2023
    Updated
    5/12/2026
    Deleted
    -

    Available On (1 platform)

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