Stable Diffusion Generation Dump

Reader mode
汤圆Mix
  • Threadmarks
    • Continuing the naming trend, here's 汤圆Mix:


    huggingface.co

    Jemnite/SomethingSweetAndSticky · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    I don't consider it part of the ShuimohuaAnimeProject because it's very clearly not even pretending to be a watercolor model anymore. As far as general models go though, I really like it. It's inspired by a mix off Wonton called Five Nuts Mixed 五仁月饼Mix - v1.0 | Stable Diffusion Checkpoint | Civitai and the mix of that mix called Dark Sushi Mix 大颗寿司Mix - Brighter Pruned | Stable Diffusion Checkpoint | Civitai
     
  • Threadmarks
    • Quality work as always. Hope you keep it up.
     
    汤圆MixV2
  • Threadmarks
    • Jemnite said:
    Continuing the naming trend, here's 汤圆Mix:


    huggingface.co

    Jemnite/SomethingSweetAndSticky · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    I don't consider it part of the ShuimohuaAnimeProject because it's very clearly not even pretending to be a watercolor model anymore. As far as general models go though, I really like it. It's inspired by a mix off Wonton called Five Nuts Mixed 五仁月饼Mix - v1.0 | Stable Diffusion Checkpoint | Civitai and the mix of that mix called Dark Sushi Mix 大颗寿司Mix - Brighter Pruned | Stable Diffusion Checkpoint | Civitai
    Click to expand...
    Click to shrink...
    Updated it to V2. here's what the same prompts look like in comparison.

    IMO this is actually a flat upgrade.
     
  • Threadmarks
    • Wagashi 和菓子 Collection - 鯛焼きMixV1
  • Threadmarks
    • Incidentally since I haven been unusually productive both in Stable Diffusion and work (but not my personal life lmao), here's Wagashi 和菓子 Collection:


    The specific goal in mind with this one was to upgrade one of my favorite models, Mix-Pro

    huggingface.co

    GIMG/AIChan_Model · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    And upgrade the backgrounds. It's semi successful. I'll go ahead and copy the post I made on Civitai on how the results went.
    I said:
    This model is an attempt to strike the middle balance between the expert character model that Mix-Pro is and give it some more ability to work in stunning landscapes besides the excellent character art. It... sort of accomplishes that, but at the cost of being neither the best landscape model, nor the best character model, but splitting the difference. It retains the style of Mix-Pro pretty well but mainly tends to zoom out a bit from the character, which makes them less prominent.

    Give it a shot if you're interested. I'll probably keep tinkering with this on and off.
    Click to expand...
    Click to shrink...
    This one is probably just going to be continuous sidegrades with all the tinkering. Kind of a fun side project. If you feel like checking it out here's the link.

    huggingface.co

    Jemnite/WagaShikashi · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    And a bonus treat for the Americans:
     
    Last edited:
  • Threadmarks
    • tankdrop24 said:
    Impressive levels of smug face.
    Click to expand...
    Click to shrink...
    She's one of the two main characters of a 2014 anime about candy. I named the repo after her as a joke (since Dagashi and Wagashi are both considered to be traditional Edo period sweets except Dagashi is cheap and for the plebs and Wagashi are each handcrafted with the artisan skill from doing nothing but shaping tiny mochi into pumpkins for rich feudal aristocrats of upper echelon Japanese society for 30 years)
    en.wikipedia.org

    Dagashi Kashi - Wikipedia

    en.wikipedia.org en.wikipedia.org
    IMO, she was the OG crazy swirly eyes waifu, and Makima from Chainsawman is just a Johnny-come-lately.
     
    餃子Mix3
  • Threadmarks


    • GyozaMix3. Probably going to be the last GyozaMix unless I discovered some new secret shit. At this point the model has basically been pushed as far as it goes as far as retaining mousymix textures and still having good proportions and backgrounds goes.



    Essentially the idea is that we're splitting the model mixing process into two parts: a picture composition/landscape arrangement component and a textural component. The former we can make by reverse-cosine Block Weight Merging any sort of highly detailed anime model against Silicon-29. I like to use dpep4 variants. Then, after we tune up the textual component to get the exact sort of texture we want, we cosine Block Weight Merge the textual component against our compositional component. After that, finetune for taste and we're ready. This, as you might figure, makes heavy use of Block Weight Merging, but it's pretty good for getting your desired effect without the huge amounts of trial and error you'd otherwise need with Block Weight Merges.

    As always, link.
    huggingface.co

    Jemnite/GyoZanMix · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co
     
  • Threadmarks
    • MilitantBird said:
    Whenever I open this thread, it starts jumping like crazy as it loads in all the pictures. How do I fix this plz halp
    Click to expand...
    Click to shrink...
    Broken25 said:
    Same here.
    Jemnite, maybe you should put the images in spoilers?
    Click to expand...
    Click to shrink...

    It is certainly an issue across the forum, though I think art threads are the only subtopic that can get away with this (except for Screencaps from Last Night). Even so, Jemnite has to use the vertical height of each image to showcase the full extent of any particular character model. The input dimensions can further affect both the pose and quality of facial expression as the AI must on the fly come up with a reason to simulate a scene fully within the canvas space.
     
    Kingcrusader said:
    These images are so damn colorful, holy shit.
    Click to expand...
    Click to shrink...
    I use Wonton+Colorbox instead of basil-mix/chillout-mix (like everyone else does) to do textural enhancement. Shadow work gets less prominent but color vibrancy is hugely enhanced. Generally speaking you're probably better off if you use basil if you want to do 2.5d or semi-realistic but since Gyoza is meant to be stylized.

    waito_x said:
    It is certainly an issue across the forum, though I think art threads are the only subtopic that can get away with this (except for Screencaps from Last Night). Even so, Jemnite has to use the vertical height of each image to showcase the full extent of any particular character model. The input dimensions can further affect both the pose and quality of facial expression as the AI must on the fly come up with a reason to simulate a scene fully within the canvas space.
    Click to expand...
    Click to shrink...
    Because of the way random noise is generated if you change the dimensions of the initial generation the output will be extremely different. If you do weird aspect that are ever so slightly off vertically, it actually fucks up the proportions, though I've never seen that work the same way horizontally. There's actually a model that is pretty cool that basically can only generate at 1024 x 512 ratio or above.

    huggingface.co

    ashen-sensored/mzpikas_tmnd_enhanced · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    It's pretty constricted, but it's great at what it does.
     
    Jemnite said:
    I use Wonton+Colorbox instead of basil-mix/chillout-mix (like everyone else does) to do textural enhancement. Shadow work gets less prominent but color vibrancy is hugely enhanced. Generally speaking you're probably better off if you use basil if you want to do 2.5d or semi-realistic but since Gyoza is meant to be stylized.
    Click to expand...
    Click to shrink...
    I have no idea what those are to be honest (I don't know how AI art works, so forgive for not having the knowledge of it). I'm more of a paper, and basic digital artist type of guy. But I have to admit, some of these do look like they are possible to make when using a digital software such as krita, the others are almost video game like.
     
    餃子Mix3.2
  • Threadmarks
    • I haven't updated this in a long while lol. Uh, so what's new? A bunch of stuff. I'll segment them into separate posts. First up: I released GyozaMixV3.2 without much fanfare, even though it represented a very substantial improvement upon the last GyozaMix.



    As always you can download it here:
    huggingface.co

    Jemnite/GyoZanMix · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    Or here:
    civitai.com

    Store Bought Gyoza (餃子Mix) - WinterHoliday2023Edition | Stable Diffusion Checkpoint | Civitai

    饺子皮是工场做的可是我答应质量很好。比手工还好吃！ Description: A stylized anime model. This model is the result of attempts to get the textures from MouseyMix , a model th...
    civitai.com
     
    Last edited:
  • Threadmarks
    • Wagashi 和菓子 Collection - 金平糖Mix
  • Threadmarks
    • I also made a new experimental mix that does some very cute and fluffy stuff. As an experimental mix, it's considered part of Wagashi Collection, so I gave it a Wagashi name: Kompeito


    But mostly I just used it to generate a bunch of cute hololives.


    You can see the full collection here:

    www.pixiv.net

    CaWFee

    https://huggingface.co/Jemnite/WagaShikashi
    www.pixiv.net www.pixiv.net

    Waga Shikashi

    huggingface.co

    Jemnite/WagaShikashi · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    Kompeito

    civitai.com

    Kompeito (金平糖Mix) - v1.0 | Stable Diffusion Checkpoint | Civitai

    This mix is light and fluffy. It is called kompeito. Partially an experiment in mitigating the downsides of mzpikas_tmnd_enhanced (as it's trained ...
    civitai.com

    Incidentally one of the merge materials output pretty good stuff, so I went ahead and dropped that on Civitai as well.

    It has very stringent use conditions, so you really want to read the description before trying to use it.

    civitai.com

    MBW-COS Aurora_N dpep5mzpikaz+tmnd - v1.0 | Stable Diffusion Checkpoint | Civitai

    Update: There is a better version of this checkpoint here: https://civitai.com/models/93427 Given the stringent use conditions of making this model...
    civitai.com
     
  • Threadmarks
    • 月光Mix - Lua's Light
  • Threadmarks
    • This one... well it's an experimental mix that doesn't feel like an experimental mix. This one I mostly did to test out how to best improve lighting and such and such. It's a very lighting effects focused model. The reason I say effects is that generally it tends to be pretty dark. It's just that when you prompt bokeh or lens flare or chiaroscuro or caustics the results really explode (in a good way).



    Here's the HuggingFace page with all the mix methodology and such and such:

    huggingface.co

    Jemnite/ALunarLight · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    And here's the civitai page with the pretty pictures that other people upload:

    Lunar Radiance(月光) Mix - Lua's Light | Stable Diffusion Checkpoint | Civitai

    Dream... not of what you are, but of what you want to be. A mix targeted at enhancing lighting effects. The base checkpoint is also CLIP expanded u...
    civitai.com
     
  • Threadmarks
    • 汤圆MixV3
  • Threadmarks
    • And finally, I backported what I learned doing Lua's Light back to 汤圆Mix. With it came some more lighting improvements (obviously) and some very targeted block adjustment with a 3d polygon LoRA in order to smooth out certain anatomical features and textures.

    As a point of comparison, here's TangYuanV2 vs. TangYuanV3.


    As you can see, big improvements across the board. Now I upload the pretty picture:


    huggingface.co

    Jemnite/SomethingSweetAndSticky · Hugging Face

    We’re on a journey to advance and democratize artificial intelligence through open source and open science.
    huggingface.co

    civitai.com

    Tang Yuan (汤圆Mix) - v3.0 | Stable Diffusion Checkpoint | Civitai

    团圆已经过了。还想吃一点汤圆吗？ I tried out Dark Sushi 大颗寿司Mix (aka Dark Souls mix because it's pronounced dàkēshòusī in Chinese lol) and thought it was fantastic...
    civitai.com

    Finally, I have displaced Counterfeit again in Baughn's Stable Diffusion bot :V :V :V

    From here on out I think what I'm going to do next is going to be experimenting with lora block alteration. Some people did all the critical research already so I only have to coast along on their work:
     
  • Threadmarks
    • I'm still amazed by what AI could do when it comes to image making, though i still perfer to make such images myself by hand.

    Jemnite, are you a computer scientist by chance?
     
    Reader mode
    You must login or register to reply here.
    Back
    Top