For business inquires, commercial licensing and consultation contact me under [email protected], for custom LoRAs visit my ko-fi page!
Follow me on Twitter: @YamerOfficial Discord: yamer_ai
If you have enjoyed this checkpoint:
Introducing ᑎIᕼIᒪᗰᗩᑎIᗩ: A fusion of 𝐔𝐧𝐬𝐭𝐚𝐛𝐥𝐞 𝐃𝐢𝐟𝐟𝐮𝐬𝐞𝐫𝐬 and DataVoid's 𝐏𝐫𝐨𝐭𝐞𝐮𝐬-𝐑𝐮𝐧𝐃𝐢𝐟𝐟𝐮𝐬𝐢𝐨𝐧 checkpoint. This extraordinary model was born from the fusion of Yamer's SDXL Unstable Diffusers Version 11 and RunDiffusion's Proteus model. This fusion brings together the boundless creativity and unpredictability of Unstable Diffusers with the enhanced versatility and style-unlocking capabilities of Proteus.
At the heart of NihilMania lies the innovative Proteus CLIP architecture, which enables the model to transcend the limitations of specific genres and styles. This means that NihilMania can effortlessly generate art ranging from vibrant anime to stunningly photorealistic images, all while maintaining the unique charm and edge of the Unstable Diffusers series.
V12 NihilMania Update Note:
The model's improved handling of CFG scaling, with a range from 5 to 50, ensures that users can fine-tune their creations to their exact preferences. One of the key features of NihilMania is its ability to interpret and understand prompts in ways that surpass the original boundaries of style and genre of SDXL base. This allows users to experiment with a wide array of tags and push the limits of their imagination, Pony Diffusers trigger words can be used but they are very experimental.
The model encourages creative exploration and supports users in bringing their unique visions to life. It's important to note that working with NihilMania may feel different from other AI art models you've used before. Its unique design and response style require an open mind and a willingness to adapt your approach. As you embark on your journey with NihilMania, take the time to familiarize yourself with its capabilities and quirks, and don't be afraid to experiment.
This fusion is not just a technical achievement but a testament to the endless possibilities that arise when great minds come together. Prepare to unleash your creativity like never before with NihilMania. Whether you're a seasoned AI art enthusiast or a curious newcomer, this model will take you on an unforgettable journey through the realms of imagination.
𝓦𝓱𝓪𝓽 𝓲𝓼 𝓪𝓷 𝓤𝓷𝓼𝓽𝓪𝓫𝓵𝓮 𝓓𝓲𝓯𝓯𝓾𝓼𝓮𝓻?
It's a collection of checkpoints where limitations cease to exist and creativity knows no limits. This checkpoint defies restrictions and empowers you to generate anything your imagination desires; 𝐘𝐨𝐮 𝐡𝐚𝐯𝐞 𝐭𝐡𝐞 𝐩𝐨𝐰𝐞𝐫 𝐭𝐨 𝐬𝐡𝐨𝐰 𝐛𝐞𝐚𝐮𝐭𝐢𝐟𝐮𝐥 𝐚𝐫𝐭! 🖌️
As the name suggests, the results obtained from this checkpoint exhibit an inherent unpredictability. You may stumble upon something that can be defined only as perfection with a specific prompt or encounter a delightful chaos of visuals. Embrace the adventure and discover for yourself what works best. Good luck in your artistic endeavors fellow creator!
This checkpoint was created with the intent to be unstable, with that I mean that it was created using a huge variety of models, the Unstable Diffusers models are both trained and merged, trained + merged are the models that will usually give you the best results in any type of style because of the huge quantity of styles it can reproduce, Unstable models are the best with abstract and unique work, they can generate anything that comes to mind and heart!
--> REDESTRIBUTION OF THIS MODEL TO PLATFORMS OUTSIDE CIVITAI AND SELLING IMAGES WITHOUT MY WRITTEN CONSENT IS FORBIDDEN! <---
You can contact me on Discord: (yamer_ai), Twitter: @YamerOfficial, email: [email protected] or directly using the Civitai chat system for my approval.
ᐯ11 ヤメールの帝国 Update Note:
Time to go back in time! This version is a follow up on V8 Heaven's Wrath unique style, I consider that version as one of the greatest and because of it this new version will pick up on that style and build on it.
Unstable V11 has been merged with RunDiffusion Photo to bring some amazing photo realistic properties to the model! This is a 30/70 merge (RD/Yamer) and if you want to play with the full version of RunDiffusion Photo it’s available on their website at https://rundiffusion.com. Unstable V11 is also available there as well, check it out and make some beautiful art!
Find this checkpoint on hugging face: https://huggingface.co/Yamer-AI/SDXL_Unstable_Diffusers
The name includes japanese to reflect a change in the way some models in the future will be released, instead of changing completely the style, from time to time a model will build on previous styles and future collaborations with RunDiffusion.
Creator note:
Hi, I'm an Italian creator on a mission to spread the joy of using AI to generate images and the progress it has made in the past few years. I craft checkpoints and prompts that anyone can use to generate their own art and share it. My ultimate aim is to be a good creator, sharing the models I create as a hobby with others.
The primary goal of this checkpoint is to be multi use, good with most styles and that can give you the creator, a good starting point to create your AI generated images and start your own career in the AI world.
All the various versions follow the general idea of an unstable model, but they individually showcase distinct styles. V2 for instance, isn't a direct improvement of V1 and the same applies to all the other versions. Every version introduces its own unique style, and with each iteration I try my best to enhance the aesthetics and visuals, ensuring each model stands out with its own distinct identity! 😸Why is this checkpoint better than regular SDXL1.0:
No embedding needed
Better variety of style
The refiner is not needed
Less AI generated look to the image. (When creating realistic images for example)
No face fix needed to generate good faces
Better photography composition
Better realistic details in all generations
Reflections, lighting, skin detail, hair, etc look more natural
Better saturation, overall quality, texture and tonality
Less blurry images
No adetailer required (Just use hires fix)
Style is stronger. (More defined and unique)
Better with SFW and NSFW. (A nsfw LoRA is recommended for specific nsfw results)
Better coloration of the images.
More variability.
Better close ups.
Less image artifacts.
Sharper and darker lines.
A better look to abstract concepts.
Liberty to try various styles and adaptability to all settings
^w^ It works best with:
Restore Face: It is not needed to generate high quality images (Not used in the current showcase.)
Clip skip: 1-2 (The model works well with clip skip set to both.)
SDXL Refiner: Not used with my models!
1) Reality is often mischievous, create your own.
2) Imagination will take you a million light years from home. But will it bring you back?
3) Be chaotic, be creative, never stop improving and never put a limit to yourself.
-Yamer GrimlockP͓̽e͓̽r͓̽s͓̽o͓̽n͓̽a͓̽l͓̽ ͓̽R͓̽a͓̽t͓̽i͓̽n͓̽g͓̽:͓̽
Quality: 9/10 (for version 1 to 9 - 11) | 7.5/10 (Turbo)
Stability (Variety): 6/10, the lower = the more variety with one prompt
Flexibility: 8.5/10
CFG resistance: 5/50 (for version 1 to 9 - 11) | 3-7 (Turbo)
Style strength: 7.5/10
Style Variety: 9/10 (for version 1 to 9 - 11) | 8/10 (Turbo)
Recommended settings (Non turbo):
Image Quality: 1024x1024 (Standard for SDXL), 16:9, 4:3, 6:13
Steps: 35-150 (under 30 steps some artifact may appear and/or weird saturation, for ex: images may look more gritty and less colorful).
Hires Upscaler: 4x_foolhardy_Remacri or 4xUltraSharp
Hires upscale: The only limit is your GPU (I upscale 2,5 times the base image, 576x1024)
VAE: SDXL VAE
To achieve a specific nsfw result I recommend to use a SDXL LoRA, this is not a nsfw focused model but it can generate some NSFW content. (SD 1.5 LoRAs doesn't work with SDXL models)
V10 Turbo Settings:
Steps: 3 to 15
CFG: 3 to 7 (The model works best with low CFG)
Image Quality: 1024x1024 (Standard for SDXL), 16:9, 4:3, 6:13
Hires and VAE are the same as the SDXL 1.0 versions
Description
V9 Heaven's Wrath VAE was merged with SDXL Turbo and SDXL Turbo Plus to make this model!
FAQ
Comments (51)
v10 works best at cfg 1-2 is what I've found, with euler a @ 6+ steps
Thank you for sharing your experience, If you want to share images too please do so that everyone can see what is capable at CFG 1-2
This is an amazing model. Please keep developing the turbo version!!! It works great for me on Easy Diffusion.
For sure mate! 😸
What is the difference between this and your 𝐒𝐃𝐗𝐋 𝐓𝐮𝐫𝐛𝐨 𝐏𝐋𝐔𝐒?
I'm struggling w/ the V10 Turbo model compared to "SDXL TURBO PLUS - RED TEAM MODEL" ... the sweetest spot I've found (to get more illustration vs. comic like results) is CFG 1.5 / Euler a / 5 steps, but Turbo Plus gives, for my prompts, far superior results. But, if I wanted comic illustrations for sure this one would rock.
This is the first Unstable Turbo models, future turbo version will improve the quality of the images! Thanks for the feedback :3
@Yamer I can't tell you how much I appreciate all your work! Thanks.
I agree the sweetest spot is CFG 1.5 / Euler a / 5 step.
So, does this new version already have the VAE baked in?
Yes, V10 Turbo Edition has the SDXL VAE already baked in.
I absolutely love v8!!! Thank you for making this amazing checkpoint! When I tried v9, the outputs have terrible quality, anyone have an idea why? I'm using automatic1111 on a 4090
Thank you for the feedback!
I haven't seen anyone mention quality output problems using V9 :/
I downloaded the model but I can't load it. What are the requirements to use this model?
From: https://gofind.ai/stable-diffusion/sdxl-system-requirements#:~:text=16GB%20VRAM%20can%20guarantee%20you,it'll%20be%20even%20better.&text=24GB%20VRAM%20is%20enough%20for,training%2C%20according%20to%20our%20tests.
12GB VRAM – this is the recommended VRAM for working with SDXL. The generation is fast and takes about 20 seconds per 1024×1024 image with the refiner. Some users report that they’ve been able to train a 1024×1024 LoRA model on 12 VRAM.
16GB VRAM can guarantee you comfortable 1024×1024 image generation using the SDXL model with the refiner. It'll be faster than 12GB VRAM, and if you generate in batches, it'll be even better. 24GB VRAM is enough for comfortable model fine-tuning and LoRA training, according to our tests.
If that didn't solve the problem and you have less than 12Gb of Vram you should try to use the --medvram settings by changing the settings of the weubi.bat file, id you don't know how I will show it to you when I am back home.
@Yamer thanks for the answer, but I have GPU - AMD Radeon RX 6800 VRAM - 16368 MB and 16gb of ram and I still can't load the model, ¿maybe the model doesn't work with amd gpus?
@deann11 no, amd should be fine, try to updste your current automatic1111 to the latest version
I've tried two turbo checkpoint models and multiple SD1.5 Lora's and it seems that SDXL turbo can run with SD1.5 Lora's, at least in ComfyUI. Not 100% sure the lora's actually do what they were trained for, the image is a bit off and differs, the combination SD1.5+SDXL does not throw an error like before. Also it seems to be possible to scale an SD1.5 latent to a turbo latent without ecode/encode. Weird.
Trying with/without Lora's for identical seed does yield different images. Using the CandyMagic Lora, the turbo SDXL with the CandyMagic linked is not the most CandyMagic like. Both do produce something of an image in 5 steps.
@pico0obyte442 hmm interesting, SD 1.5 LoRAs should not work with SDXL models, please let me know if you find anything interesting.
If you can please upload some images just to see what the differences are with and without the LoRA, thank you for informing me about this! 😸
sure: https://civitai.com/posts/978677 (Edit: link is broken by now)
all the filenames have "turbo" in it now... have they all been merged and converted to turbo? I don't like the turbo stuff
Yeah the recent model is a Turbo one now but you can still use the old SDXL 1.0 to make your images.
@Yamer that wasn't entirely clear... so, are V9 and V8 now turbo-merged models or no? or is that only V10?
@shapeshifter83 only the last V10 Turbo is a merged turbo model, all the other ones are SDXL 1.0.
@Yamer excellent, thanks
The future versions of this model will all be SDXL 1.0, I won't release any future Turbo model, probably you are happy to hear this because you said you didn't like turbo stuff, and I personally am not the biggest fan of it. :)
@Yamer Can you tell us why you dont like Turbo, for me it affords me the abbility to generate image on colab. Will be sad if the creators of all the excellent models wont be investing their time with it going forward, But it is what it is . Like usual I will find some way to make images
@lukeovermind because most users either use SDXL 1.0 or SD 1.5, not a lot use turbo checkpoints and that is the main reason why I don't make them, another reason is that I just prefer SDXL 1.0, just by using 20 steps you can get great looking images.
I'm downloading the Unyielding Grimoire +vae version because I've tried the newer versions after V8 & I don't like which way it's going. I think you're over mixing what used to be a good thing, imho, but what do I know?
Old version have a simpler style while the newer ones have more creativity Wich can be both good and bad, that is why old version are still up and downloadable, of someone likes the old style they can download older versions, let me know how your experience with unyielding goes! ❤️😸
Thanks for the feedback!
I have to agree, I always find myself going back to v8
I get duplicates of figures or parts of figures 80% of the time in 1:1 image aspect ratio with this model. Image overall looks very nice and detailed after highres fix, I would love it, but those duplicates...
Tested without any Loras, Euler A. sampler with normal scheduling, CFG 1.5-3, steps 5-15, many different prompts (usually simple). For image ratios other than 1:1 it's worse.
Thank you for the feedback, I will try to fix it in future versions
This Model is labeled as SD 1.5. You might want to fix that.
The latest version (V10) is labeled correctly as a SDXL Turbo model, no error was made, if you are referring to another version let me know
I think he means that when you filter your search for models that are SD1.5 it shows up, I'm seeing the same thing.
@Bdwonk older version of this model are SD 1.5, that is why it shows up as 1.5
this is a dumb question I know but I have used a few of your models and I like the output. I have a workflow that makes stickers. I see I can sell the images but how do I credit you if I am selling images locally or do I only need to credit a model if I use it as a base to train off of? also i know i cant use this one for commercial purposes since its a turbo model.
You can credit me with a link to my Civitai profile/Link to the model used/Link to my twitter profile.
If you want to sell stickers using my models that is fantastic, if you have a profile on platforms such as Facebook, Youtube, Twitter etc... you can credit me there with a link to the original model or my profile :)
Thank you for asking before selling, I really appreciate that!
@Yamer Thank you for getting back to me. Some of these I have made die sublimated bookmarks with and stickers with. I have sold a few in my area but only used the SDXL MergeHeaven 🔝 𝓣𝓱𝓮 𝓑𝓮𝓼𝓽 𝓞𝓷𝓮𝓼! since it has the "Use the model without crediting the creator" checked in green. I mostly sell at small salons and book stores in my local area. I do Alot of IT work for small businesses so finding places that I could sell locally was easy. I don't have an online store setup yet. but for those local places if I used other models that require credit how would you like me to credit you?
@technoshaman if possible when you sell something one of your stickers, you can write on a piece of paper something along the lines of "mentions" and mention my twitter profile name @YamerOfficial as the checkpoint creator that was used as base for the sticker to be made.
@Yamer will do. Thank you again for clarification. I Appreciate your work.
@Yamer I uploaded a few pics of the bookmarks on twitter and tagged you as well. Thanks again for your time. https://twitter.com/Technoshamon/status/1742117172959428771
@technoshaman perfect friend, I have seen you tweet, if you want to help me I have a ko-fi page for donation but you are not obligate to donate, my models will always stay free to use and download. :)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.