💡 Newest Version - Walnut (Euler or Euler a sampler recommended)
🌿 Community Favorites - Flax, Hemp II, and Walnut
Model Series: multi-series
Version: Walnut (ymk3.0.0)
Created on: ~March/14/25ℹ Notice - Do not reprint this model. This model can be merged and shared, however derivatives cannot be used for commercial purposes (generative services, commissioned models, SeaArt, PixAI, etc.)
Info
A multi-series model page to showcase various 'flavors' of similar or continuations of related illustrious/noob based model merges. This page hosts various styles. Go through and decide on your favorite(s). Some versions may be more 'lora heavy' in comparison to the others.
view more on Hugging Face
Recommended Use
< See example images >
• Euler or Euler a or HeunPP2 or Restart sampler, 3 or ~3-6 cfg, ~28 steps
Some versions may also work well with:
• Euler a CFG++ sampler, SGM Uniform scheduler, ~2 cfg, ~28 steps
You can find additional sampler and schedulers via the ForgeUI/a1111 extension here.
---
Euler, 3 cfg, 28 steps is a good starting point for most versions
Hires fix - This is optional of course, and I recommend adjusting any use to your context and application. The example images utilize this both on and off.
Disclaimer
This model(s) may output NSFW content unintentionally depending on parameters used. Make sure you tailor your prompts accordingly. For example "nsfw" in the negative prompt.
The purpose of sharing this model is not to showcase obscene material in a public forum. The use of this learning model is entirely at the discretion of the user, and they have the freedom to choose whether or not to create SFW or NSFW content. The decision of whether to engage with SFW or NSFW content lies with the user and their own personal preferences. The AI model(s) do not contain explicit visual content that can be accessed easily.Version Log
Walnut 🌿 (Euler or Euler a sampler recommended)
Model Series: ymk
Version: Walnut (ymk3.0.0)
Created on: ~March/14/25Hemp II 🌿 (Euler or Euler a sampler recommended)
Model Series: ldbr
Version: Vanilla (ldbr4.1.1)
Created on: ~January/30/25Vanilla 🌿 (Euler or Euler a sampler, 3-6 cfg recommended)
Model Series: epg
Version: Hemp (epg1.0.1)
Created on: ~March/4/25Hemp 🌿 (Euler or HeunPP2 sampler recommended)
Model Series: ldbr
Version: Hemp (ldbr4.0.1)
Created on: ~January/30/25Coconut 🌿 (Euler A or DDPM sampler recommended)
Model Series: dmai
Version: Coconut (dmai1.1.0)
Created on: ~December/27/24Almond 🌿 (Euler sampler recommended)
Model Series: pfi
Version: Almond (pfi3.0.0)
Created on: ~March/5/25Flax 🌿
Model Series: ldbr
Version: Flax (ldbr3.0.0)
Created on: ~January/24/25Oat 🌿
Model Series: GRK
Version: Oat (GRK3.0.0)
Created on: ~January/15/25Additional version information to be added.
License & Use
SDXL - CreativeML Open RAIL++-M
Illustrious - https://freedevproject.org/faipl-1.0-sd/
NoobAI-XL - https://civarchive.com/models/license/1140829
Terms of use
- You are solely responsible for any legal liability resulting from unethical use of this model(s)
- If you use any of these models for merging, please state what steps you took to do so and clearly indicate where modifications have been made.Note
If you see any conflicts or corrections to be made, please let me know.Description
Hemp Milk II
recommended to try Euler and Euler a samplers first. (3-5 cfg, ~28 steps)
a bit of a 'refined' version of the original Hemp. Still creative and a bit messy - not necessarily better than the original Hemp version depending on use and preference
try some uncommon image ratios
lora, character tags, artist tags, etc. may provide more structure in various ways if needed
hi-res fix may not do well with this version (nerfs details/cohesiveness of images a bit imo)
FAQ
Comments (252)
Showing latest 150 of 252.
Amazing 😍 🤩
不多说了，神中神
Impressive checkpoint! Walnut, Hemp II and Flax...amazing!
Those seem to be the favorites, yep - thanks for the feedback!
@Ocean3 any chance of getting almond, coconut and vanilla on civit?
@leonel22 that's determined by user's 'bidding' for a specific version each week
@Ocean3 understood, will be rallying bids for them 😊
What a great checkpoint! So much fun to play with. Very crisp and responsive 10/10 A+
ngl this chcekpoint is too op
What is the difference between versions?
This is scary close to the last image checkpoint we'll ever need.
It really does just about everything SO well.
You are doing God's work out here
I've tried so many models and it's the only checkpoint that isn't working so far. Images are blurred or aawful. I don't get what am I doing wrong
Which version are you using and what are the parameters you're using? Are you using a local ui? If so which one? Or are you using a generation service? Are you using any lora models?
@Ocean3 hi. thank you for your message. I tried your model locally in fooocus. I don't know the model version I just pushed download in the model page. walnut milk? I tried the model with the "add detail lora" (strenght 0,3) then I tried without any lora. I didn't change any parameter. usually I just set 768x1280 and speed performance. I never mess with advanced settings. did I have to? were your suggested parameters mandatory?. the images I got were very blurred (at best) but most of the times the image was "corrupted" like when you use too many loras or set way too much strenght on a lora. Maybe I'm not explaining myself well. I have tried so many models and this is the first time it happens. probably it's my fault somehow
@zultan3 I believe I get what you mean. Maybe try to re-download the model or that the file is the correct file and loading properly. I'm unfamiliar with foocus. The model shouldn't be outputting that way. Try Euler sampler, 28 steps, and 3 cfg as a base. These parameters are just a base and aren't required. Then make sure there are no lora conflicting just to test things and make sure. There may be something off with something you have set or with how foocus is doing something. The name of the model should be displayed via download name as well as in image metadata btw. So you can also use that to double check.
@Ocean3 Hi. thank you so much for your reply. I couldn't find it before. Today I downloaded the model again (walnut), again no loras, I just set euler and the model is generating images. maybe this parameter was really important in focus, I don't know. Thank you again for helping me. Good work. The images look good!
Oat Milk has been my go-to for training my Illustrious Celeb LoRas(on Tensor). The ease of mixing realism into the anime aesthetic is amazing
Thanks for the feedback man
very stylish and fun to use, thanks!
This model is peak, only issue I have is whenever I generate, at least one image will always be in black and white, what’s the issue? It still does it even when I put monochrome in the neg. Prompt
@Ocean3 Thanks I guess. It’s working spiffy now.
@M1Carbinemann Neat!
This model (coconut) could be the best model I've ever used, and I've used a lot of models. All Plant ed milk models are great imo, but this one gets the best results for me. Thanks for an awesome suite!
thanks for this great checkpoint!
walnut is such a creative beast
.
Good checkpoint. Could you tell me the training base model for these models? Because I find when I use latent upscale, the image quality will deteriorate. I don't know how to solve it. By the way, I tryed three models.
Amazing model, like I said before, I just have one question, what’s the difference between each “plant milk”? Just curious.
Now, I'm not a lawyer, but I do believe the licenses for these models do not extend to outputs. So while understandibly, you shouldn't and can't repost a model for commercial purposes, the outputs made with it have the standard limitations for commercial use etc correct?
Hello, ocean3! Any plans on new models? I know, walnut is already almost perfect, but there are no limits
Absolutely love this!
Just wanted to say, this model really surprised me. For having 18 female samples (to two non-female), I thought it was gonna for sure twist all my males into being female, but the prompt adherence on this model is just amazing. I get the most consistently good generations on this one.
It's not top notch for pure realism, but in everything short of that, it does quite well. Especially in terms of creating near-real versions of anime/cartoon characters. Like, exact likenesses though, it doesn't blend in realistic style, it keeps the style of the characters in whatever lora you're using, and makes them nearly real.
In a pipeline with another model for polishing, this could probably even make them fully realistic.
walnut has got to be the best model I have ever used. It does a variety of styles extremely well and has great prompt adherence.
WOW!
As the civitai spokesperson journalist for models (lie) I can confirm that this is THE model. From furry to human, abstract or anime, from detailed to sketching (most of this I haven’t done.) this is one of the best, I’d argue to say THE best. Keep doing what ya do, Ocean3.
12 Burbonist France’s out of 10 Coughing Babies
After trying many different chkpts, I always love coming back to Flax. Walnut is clean and sharp, Hemp is creative and messy, but Flax is just right.
Geeze, damn, i love this Model, well done work. I will create more stuff in the near future with this one.
Thanks for creating one of, if not the current best Illustrious model on the site!
Just one thing, I realise that for Walnut Euler and Euler a samplers are recommended, but I was wondering if this is still the consensus given that there isn't a comprehensive breakdown of this version compared to the other ones. Cheers!
This model (Walnut) is working really nicely for anime and flat color generations. Thanks!
Very nice, keep up the good work!
Did you removed plant milk model from tensor art? i was just using it yesterday and now it is gone
back plant milk to TENSOR ART!!!
Please make it visible again on Tensor.
Great model, highly recommend
非常好的模型，它使我的图片充满魔力！爱来自广袤的东部❤
awesome!
I have merged this model with CyberIllustrious | CyberRealistic .. it's better than i hoped thanks for this model <3 btw can i upload it to civitai ?
Walnut is an incredible model. Very smart with prompt recognition, understands and replicates character very well. I also appreciate the added contrast the style has, makes semi-realistic styles much more believable.
first i wanna thank you for this amazing model its one of the best i've tried but i have a question would the noobai control nets work on this one?
Respect to whoever promoted Hemp II to 2nd featured this week - thanks for showing your support and interest 👍
Lacking art knowledge, I had no idea how to tell the difference between models.
It's a really nice model and gives good results, but it has its flaws. For example, it only draws Pokémon characters like Machamp in 3D, even when I try to specify otherwise in the negative prompt. It also seems to have a preference for a certain type of lighting. I've been testing the model for three days now, and I'm using fairly simple prompts, but I probably doing something wrong. Do you have any suggestions?
This is my favourite model.
Such a creative and excellent model. It would be hard to top this one for the stylistic choices and imaginative imagery it creates. Excellent work. If you're reading this, just get it now.
This model always gives beautiful results, by itself or to refine other anime models.
In addition to that, it gives amazing/different results by changing the clip :
https://civitai.com/articles/17854
Very good job !
this latest illustrious model is said to be noob merged. is this v pred or epsilon type?
It's an amazing model! Especially when it comes to anime characters' skin and hair, it's absolutely top-notch.
Really pleased with this model!!!
Images are created in a monochrome color scheme, how can I fix this?
Evaluate your positive/negative prompt, change your parameters etc. I don't know what you're doing or the context.
Can you upload an example? Sometimes yes, i got that effect too, when image that shouldn't be monochrome actually IS, because of some combination of prompts
add " black and white, b&w " in negative prompt, it worked for me.
"black and white", "greyscale' for negative prompt and dont put too much weight on any tags in positive. For me this sometimes happens if i put more than 1.3 weight for any tags in positive
Such cool generations with this
Love this Model! :) <3
Thank you bro, my favorite model, flexible and beautiful :3
I LOVE this
so so so so so gooooood！！！！！！！！！！
Good model, thank you.
This model seems to be pretty consistent with different loras and prompts, which can't be said about some other popular Illus
Walnut is one of my most used Illustrious model, such good stuff and high compatibility with LoRA's. Can't wait for the next tune <3
Love it!
Flax is unavailable, is there an effort to restore it?
I do what I can, they limited the threshold from 200 to 100. I have the main versions that people use on auto-bids. The amounts that people bid on other models at the end varies and can sometimes push models out of the threshold, even with 2 minutes remaining for that period. Feel free to bid on the versions you want to see each week as well. This will help them have the ability for them to be used on-site here.
Flax was great what happened to it?
I do what I can, they limited the threshold from 200 to 100. I have the main versions that people use on auto-bids. The amounts that people bid on other models at the end varies and can sometimes push models out of the threshold, even with 2 minutes remaining for that period. Feel free to bid on the versions you want to see each week as well. This will help them have the ability for them to be used on-site here.
@Ocean3 I bid on Vanilla days before the cycle ended, much time for bids - never intended to push Flax out, didn't even knew it would happen. How should I knew? I didn't give 2 sh.... about Flax or the other 2 Versions. Anyway, I leave you all alone with your precious Flax, I cannot get a hang of it. Bye
PS EDIT: I wouldn't have bid more again, I just bid and left and waited, I was more caring about WAI anyway, and would have spent a million on it if I would. So, don't worry. Never again being a culprit here.
Reminder to those that want to see specific versions available to use with the on-site generator - the threshold for available models for Civitai's 'auction' system has been lowered from 200 to 100. This means that more people will likely bid on the models they want to see more competitively, especially within the last few minutes of a cycle (1 week). I have auto-bids for Walnut, Hemp II, and Flax from myself currently. This may not be able to get a model through one week or another at times, depending on the influx of people bidding on other models. I'm also not available to check the final minutes most of the time during the auction's phase. If you'd like to see a model maintain consistency with uptime, or see another version become available, feel free to bid on them each week 👍
For reference:
Vanilla was bid on by other users this previous week with an amount of 30,100 buzz and ranked at 32nd 'place' and became available to use.
Meanwhile Flax was bid with an amount of 23,000 buzz from myself and another user(s) and ranked 119th 'place' and is not available to be used this week due to being above that 100 limit threshold.
This is a good example of just how tight the last rankings can get, especially on the final days, hours, minutes, etc.
Oops, sorry, I'm the culprit (not that I kept it a secret, quite the opposite). In addition to 30k for One Obsessions and WAI NSFW v15.0 with 50,000 buzz, I just felt like bidding on this with mentioned 31k. But it was long before the deadline, easily 4 or 3 days... you could have easily scraped together 10k. Oh well, I tried it, you can have Flax back. I don't like plant milk, and I never will. To each their own.
@ToshioRiko you're not the reason flax wasn't in this week's. I don't really know what you're sayin here, but gg. If you don't enjoy Vanilla or the other versions, no sweat man.
We really need Flax, sad they changed their auction system.
I really like Flax. Works surprisingly well for both anime and western cartoon styles.
What's the minimum image size? Or optimal image size? Don't think I saw it listed anywhere.
typical for sxdl/illustrious or look at example images
Yeah, the images are mostly huge so I went with Illustrious. Thanks.
@eponymouse777487 some of them are upscaled/hi-res'd. Typically I use 768x1152 or 896x1280 as the base.
Use classic PONY/SDXL sizes, they work fine. To be honest, it works with basically everything.
These usually work (you can swap dimensions if you'd like):
- 896*1600
- 1024*1536
- 720*1280
- 832*1216
- 768*1344
- 640*1536
- 1344*1856
You can go higher as well, but your bound to get more artifacts and less stable results depending on the model used and the data it's trained on. I would also recommend not to go higher than x1.5 on the highres fix and keep the denoise in the range of 0.3 - 0.5.
Generally you want the dimensions divisible by 64, and total pixels between 1 and 1.5M.
Bad resolutions will result in unprompted twins, unprompted multiple views, weird focus, faces that look like flounder or cattle, etc. If those issues aren't showing up often, your dimensions are fine.
@EbenezerDanglewood In case of ILLU, usually not. SD15, yes that would be major issue.
You can trigger those issues only if you get over size limit what model can do (altho its more like CLIP limit really).
As for size limits, power of 8 is enough, much like for SDXL. Altho in most cases models will work even with really odd sizes. Just not good results.
I really liked Hemp II, it’s sad to see it up for auction. 😞
What happened now? How much time it's unavailable? Thx!
@Sanjinho They’ve reduced the on-site generation checkpoint slots to 100. If you’d like Hemp II to be available for on-site generation, you’ll need to bid on it so that it secures one of those slots. Until then, it won’t be available for on-site generation.
@IceB2 Ah okay! Thanks for the info! 🙏🏻
@IceB2 For Hemp II to be available in the Civitai generator, what position does it need to be in the Bid? Thank you.
Does the model suite walnut version come with it's own vae already packaged? if not, which do you recommend?
@artist271 VAE is baked into the ckpt, but if you'd like to use another one, you can try any sdxl vae.
Walnut has quickly become my favorite model. Incredible.
As for portrait images what resolution is better to use - 1024x1360 or 832x1216?
@nambell Based on what I've tried, these are the best resolutions for portraits:
- 896*1152
- 832*1216 (recommended)
- 768*1344
- 640*1536 (recommended)
- 720*1280
- 896*1600
- 1024*1536
If you have decent hardware, you can go higher; however, you'll definitely encounter more artifacts in your image. Hope this helps, cheers !!
I never comment, but Walnut an incredible model. Quickly became my first choice.
rubs chin Great checkpoint. Only issue is Pink haired girl photobombing EVERYTHING.
sounds like a you issue bro
I will make so many monsterfemales with this Walnut.
I love it! ❤️🎉
Do you ever plan to make a new version with Illustrious v2 or you're going to wait for something else?
ILL v2 is a downgrade
perfect and beautiful model...
N°1
No.1
These models had such an affect on model merging, big time. Dying to see a new model uploaded here
why "Hemp II' is disabled?, I think is the best from all,. please enable again
Do I need VAE?
it's back <3 (glad some more model slots opened up again, 25k to qualify was bonkers)
how much slots opened? thx
@DavyXebec 50 more, so there's now 150
Hemp II is godly
I love your sketch images made with Hemp II, I use them as a base to create my own, thank you so much for your creations!
This model is MUCH better at realism and 3D than WAI!
Love the model (using Walnut mostly) - ty!
What's the exact merge versions of IL/NAI? Helps determine valid prompt tags and other details for optimizing prompts and other nodes.
This is officially my favourite anime model. It doesn't overdo things, it looks clean and the hands/feet almost always look good.
I don't support Tensor.art. Please stop asking me. I asked the guys that uploaded there to remove them. One did, which I respect a lot. Prior to that, the models themselves were up for a time (after I had to claim them), then they were privatized on my account (that I didn't want). These companies promoting and fueling ripping and auto-uploading models are no bueno. Seaart even goes as far as 'creating an account for you' as if the person uploaded resources to their site themselves. I do not support those sites, their tactics, or their teams. Thanks.
As far as this website, I haven't been messing with AI stuff at all for months. I'm not really interested. I see too much of themes I don't care for from people and don't care to keep seeing it all the time. There are other versions, no tba on uploading them.
I see it as a sad attitude, even more so now that Civitai has been prohibiting the use of NSFW for common users, where even images that are not very vulgar are censored by the site's generator. Tensor Art was the best site to use its models (which are very good and varied), but unfortunately, like Civitai, we are now censored from using this great model in the way it was made. That's my opinion, really sad.
You have talent, Walnut is my favorite model. I wish it got updated with newer characters. If you have talent you should not waste it, I would like to donate to support you if you need it, and I'm sure others would too, so please continue your work.
@DavyXebec "...in the way it was made." I have no idea what you're talking about dude.
@DeviantApeArt just model merging. Not sure some of you realize this. I'm not training models in any specific way. I merge mixes/base models and lora together in various ways for different effects. A lot of the time, lora are used with both positive and negative values which do different things to the base/merge models. Over time, these 'series' (or versions) I do build up over time into what they are here. Most of the version I have done were for fun/tests and aren't shown anywhere. These ones stood out for various reasons that are mentioned. That's why you see stuff like 'series GRK, or NOM' etc. I've done more experimenting with merges than image generation.
@Ocean3 That's ok, if your merges create something better than the original individual model than you have created value and deserve credit. What base models/loras were use to create Walnut if you don't mind me asking?
Hemp ll is my favorite by far.
No way no one bids in Flax its the best model come on guys
Stable finger & toes, Good style with artist prompts, Incredible detailed background for anime model. Love walnut ♥️~
I feel like no matter what i try i cannot gen feet.
it keeps making them kneel, hiding the feet. or just having them stand and its cropped at the thighs. any tips?
Do you have an example of your outputs? Would help to see your positive and negative prompts to pinpoint where the issue is coming from. Also, which model are you using from the ones above?
try adding 'full body' or 'standing', and if that doesn't help maybe try a different resolution
^ Solid advice. Also try 'foot focus' or 'feet focus' in the positive prompt if you really want to try forcing things. I also found, personally, (toes, five toes:1.1) is a solid addition to any positive prompt if you intend to show toes, since reminding the model to keep the toe count consistent is always useful.
Without it I tend to get four toes, or six toes more often.
Also try to avoid any sitting positions that could cause their feet to be hidden under their bodies, so 'legs up', 'spread legs'. 'Presenting foot' is also good if you want the view to be focused on their foot.
try ((show feet:1.3))
I have and like all the Plant Milk's, but Walnut is in a league of its own. If I could have only one art model it would be Walnut.
Amazing model, which doesn't need many extra loras.
Honestly been massively enjoying this model since I picked it up. Works great even with cross-model LoRA and generally gives me the styles I want with little prompt input.
Nice ee
Noob question: What are your observed positives (and maybe negatives) compared to other Illustrious models, like NovaFurry / NovaAnime? In the preview images here, I see no major differences. Maybe someone who has experience using both can chime in. Thanks!
The main difference lies in the consistency of walnut, for example; it has better color absorption, and the loras don't disrupt checkpoint generation at all, unlike many other checkpoints. Add to that good prompts, and walnut is unstoppable.
Still by far the best checkpoint
it's perfect bro
This is a serious mode, and the fact that can also handle ultrawide is exceptional, with this I managed to create some super desktop background, and for the rest it will do anything you will ask. Truly superb, few are this good. Please keep it up.
I've really enjoyed using the Plant Milk Model and the Flax checkpoint over at moescape ai but last few days I've noticed how everything is coming out with a strange glossy pastel or plastic look with bold inking style rather than the clean images it used to generate before. Wondering if anyone or the creators know how to rectify this. Much appreciated and love the model.
Best model! More pls?
Are these models not meant to generate nsfw? I am having a bit of difficult time to get the characters do what I want.
Was looking for a great not completely NSFW driven model and was skeptical at first, but wow the detail and responsiveness is beautiful. Truly amazing work!!
I absolutely love this model! I do almost all of my work in it — huge thanks to the author!
Details
Files
plantMilkModelSuite_hempII.safetensors
Mirrors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
plantMilkModelSuite_hempII.safetensors
Available On (5 platforms)
Same model published on other platforms. May have additional downloads or version variants.