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
A brand new dataset of more than 1200 images. A Keyword “NijiEX style”
Two Niji training merge to balance style and quality.
FAQ
Comments (55)
Wow NijiV6, exciting! Finally, some good news to cheer me up after the SD3 debacle 😎🙏👍
Very excited to hear that a Cascade version is coming too.
Yes. I've had my 3090 for 3 or 4 months, but it took me a while to achieve something satisfactory. I also spent a lot of time on Stable Cascade which is truly incredible for training!
@Stan_Katayama Better late than never. Welcome back, and I am sure V6 is worth the wait 🙏👍😎
@NowhereManGo I'm waiting for you to tell me what you think about it. He's more in control, but maybe he'll lose his joyfully chaotic side 😅
@NowhereManGo I checked "Use on Civitai generation service" Normally you should be able to generate images on Civitai with it. It does not work?
@Stan_Katayama Unfortunately, since Civitai switch from OctoML to their current GPU providers, many of the less popular models are no longer runnable on their generator.
When NijiV6 becomes more popular (via Civitai's model ranking system) it may become runnable, but that is not the case at the moment.
@NowhereManGo You will be able to generate images my friend :)
@Stan_Katayama Thank you Stan, this is a pleasant surprise 🙏. Just generated my first two NijiV6 images and they came out great 👍❤: https://civitai.com/posts/3682742
@Stan_Katayama I am both honored, and a bit embarrassed by your buzz, because I am supposed to be the one giving you the buzz 😅
Niji is the first SDXL fine-tune that I truly fell in love with, because it was a childhood dream come true: to be able to create my own anime/manga like images (I don't have enough talent and training for that).
With NijiV6, this dream has now become even sweeter 😎
@NowhereManGo If you want to see some images of my next model. Check it out on my Insta or my Twitter. I'll post more later!
@Stan_Katayama thank you, I'll definite check them out 🙏👍. I assume this is the Cascade version?
@NowhereManGo No, it's a model that will do both anime style and fantasy comic style. But it will be on SDXL
@Stan_Katayama I see. I look forward to its release then 👍🙏
Is ex nijiv6 style?
“NijiEX style” is for the “expressive style” of Niji5. But there are also images from Niji6 in the dataset. Maybe in the future I will make an exclusively Niji6 model.
Awesome!! i was eagerly waiting for your model !! thankyou so much ♥♥♥
Happy to please you! I hope you like this version! Don't forget the trigger word "NijiEX style" for a more "Niji" style 😊
what is the best way to prompt this? Tags like (masterpiece) like base SDXL?
The key word "NijiEX style" gives a more pronounced style. There are some indications of settings in the description. Otherwise try things, have fun
Like Stan said, no need for (masterpiece) or score_9, score_8, "best quality", etc. Look at the sample image to get a feel for what is needed.
But here is an explanation for why (masterpiece) score_9, score_8 are needed for some models.
Model such as Pony have a very large training image set, and some of those image are of poor quality. The reason these poor quality images are included is that for some characters or poses, only poor quality images exists. But a poor render is better than no render, so they are included.
But when you try to generate a "high quality" images that does not need those characters or poses, it is better to "exclude" them. So you tell the A.I. to only "include" images that have "score_9" or "masterpiece" in their captions.
Niji V6, on the other hand, is presumably trained only with high quality images generated using Midjourney. So it is already "biased" toward "high quality", so there is no need for such words in the prompt.
@NowhereManGo I think that on this version, we can put: best quality, masterpiece... :)
@Stan_Katayama Yes, I guess "best quality, masterpiece..." doesn't hurt either 😁
When will Cascade be available to the public?
I can't say when yet. I would really like to bring as many people to Cascade as possible because I don't understand its lack of success! But it shouldn't take very long. Actually, I already have a personal version. But it is not accomplished enough.
@Stan_Katayama I can't wait to use your model, good luck!
@adc25a137678477 I did 10% of the training. But since it's hot at home, I only do it at night
Fantastic checkpoint, my only complaint is on approximately every 10th generation, male models would get one very large nipple which tend to ruin a lot of great generations. Hope it'd get fixed on the next version. Thank you for your hard work. :)
Thank you for this feedback. I am very far from being a professional, I do not have a pro GPU, nor a multi GPU. I probably still make mistakes, but English is not my native language. I'm probably missing some things. I don't promise anything for the next one, but I'll do my best 😅
Valid criticism, but to be fair, flaws in A.I. image generated images are almost unavoidable with today's technology 😅.
That is why people still inpaint so much.
Good thing I am not really into nipples 😂
@NowhereManGo I don't particularly do nudes either. And then it's a Niji model and Niji is censored. So I can't change that.
@Stan_Katayama Yes, that's a good point. So most likely these weird male nipples are coming from SDXL base itself rather than the additional training Niji V6 has done (base SDXL itself produces bad nipple, IIRC).
The Midjourney images you've used probably has very few nipples (I know MJ shuns female nipple, but does MJ even produce male nipples? 😅), so very unlikely that the fine-tuning introduced them.
@NowhereManGo We are talking about male nipples here. What Cronoe does is magnificent. A bit of a Renaissance style.
In short, it is possible that my model is indeed a problem at this level.
The problem also comes from the fact that I use a lot of "expressive style" images from niji5 and that it makes a lot of mistakes
@Stan_Katayama @NowhereManGo I ran a test on just this checkpoint and found the same relic on some side view poses, but it's much rarer. I also found that the Jean Delville style LORA I was using forces the models to face sideway more often, thus giving me a lot more of the issue. It may as well be the main attributor to this problem. Overall, this is not a big complaint, just something I just want to bring to attention. Thanks again for making this checkpoint, have a nice day.
@Cronoe Thanks to you ! In any case, what you do is magnificent! 😊👏
it works 90 percent of the time, and you still complain..jeez..
@alexgevo781 He doesn't complain. It’s important for me to get feedback! As long as it is respectful, I take all criticism.
It would be interesting to know if it is with or without lora. Lora tends to create aberrations and the more you use, the more unexpected results you get.
How can I use a prompt to fix a style?
I'm not sure I understand?! You must read the instructions in description.
Look at the images people have posted, find a style you like, then look at the prompt. For example, use "Anime Artwork", "NijiEX style illustration", "style by Masamune Shirow", etc.
@NobodyButMeow Thank you for your help! My English is very modey and I sometimes have difficulty answering or understanding precisely 😅
I apologize for my poor English. It seems I didn't describe it clearly. What I want to ask is how can I generate a specific style?
@NobodyButMeow Thank you! This is very helpful. May I ask if there is a list that can find the styles of included artists?
@a5186587 Do not apologize. I'm not very good in English too 😅 There is the keyword "NijiEX style" to activate the style specific to this model.
Otherwise @NobodyButMeow explained everything to you better than I could 😅
@Stan_Katayama You are welcome. I am always happy to help others 😁
@a5186587 Most of these artistic styles are in fact build into SDXL and not specific to any particular model.
I've made a set of images to demonstrate some of my favorite artist's styles: https://civitai.com/posts/4474094 and https://civitai.com/posts/4530055
Here is a list of bookmarks I've collected, have fun 😁:
SD1.5 Artist Study (with images)
SD Artists Browser - a Hugging Face Space by mattthew
SDXL Artist Styles Studies MisterRuffian's Latent Artist & Modifier Encyclopedia
SDXL 1.0 Art Medium Study - 200 mediums : StableDiffusion
SDXL Artist Study | Weird Wonderful AI Art
Artist Studies | SDXL 1.0 Artistic Studies
Stable Diffusion: Trending on Art Station and other myths | by Adi | Medium
SD Styles list of artists for SD v1.4 A-C / D-I / J-N / O-Z
32 Art Styles on Midjourney V4 you must try! | by Exquisite Workers | MLearning.ai | Medium
Stable Diffusion Artist list - Style studies
Liste des styles - Stable Diffusion France
@NobodyButMeow HOLY,You are practically a saint, I'm almost falling in love with you dude
@Stan_Katayama yes yes,i saw it,thank u very much~~~
@a5186587 LOL, you are welcome 🙏😎
@a5186587 I've just made a set of images to demonstrate some of my favorite artist's styles: https://civitai.com/posts/4474094 and https://civitai.com/posts/4530055
Amazing Model!
Thank´s!
Thanks to you ! 💪
I have this problem:
RuntimeWarning: invalid value encountered in cast
img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))
It produces a black image for me during generation I see that the image is "forming", then before completing it becomes completely black
I'm using the model's VAE.
Trying a different VAE, I get the image but it's all distorted by pixels and artifacts
Do you have any suggestions on how I could resolve this?
Sorry,
I didn't see the message.
We are rather dealing with a software problem here. And I can't really help you on this point.
What I do in these cases is I copy/paste the error message into Google.
There is always someone who had the same problem.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.