What the noodle doin?

Alrighty! While I'm making progress on Helena and v10, in today's post I finally explain what exactly it is I'm doing when I'm working on my model. I've made this step-by-step outline, here we go!

The first step is deciding what the LoRa needs. It can be anything. Absolutely anything. A character I need to generate consistently, clothing, a location, a sex position, anything. I collect a huge amount of images whenever I think "hey that looks neat", but one thing I try to do is have each image pull double duty. An image of Bluebell would improve the AI's ability to generate her, and the model kinda sucks at 2koma comics, so let's do a 2-panel comic involving bluebell. And Yu too, why not.

In this case we'll just take some screenshots from an old dragon ball fighting game and slap them into a white border:

Next, we'll draw over it manually so we can inpaint it to look nice. The most important aspect at this stage is getting the general color outline, I cannot stress enough that this stage takes basically zero artistic skill:

Now we need a prompt. Here's what I wrote:

meme face, 2koma, head only, v-shaped eyebrows, white border, smirk, mouth closed, looking to the side, duo, leaning forward, outdoors, blue sky, animal crossing, trees, clouds, daytime,

1boy, cute boy, femboy, brown hair, short hair, t-shirt, red shirt, short-sleeved shirt,

sy-shop, animal crossing, furry female, shortstack, domestic cat, feline, blue fur, monotone body, dark paws, anthro, pricked ears, dark ears, light blue hair, hair tuft, thin eyebrows, red eyes, pink nose, grin, closed mouth, cleavage, red apron, medium breasts, sideboob, naked apron

All of this is covered in my img2img section in my stable diffusion guide, but basically I use part of the prompt, inpaint the top at varying settings until I think it looks nice, then inpaint the bottom the same way. Here's what it looks like when we're done inpainting:

Looks nice right? But unfortunately we now get to the painful part. Look closer.

Closer.

You see how blurry it is? Everything in our training data will influence the model, and this may look kinda blurry now, but if we train our model with blurry pictures the model will carry over that blurriness on top of its own natural blurriness. From there things will spiral out of control and we'll end up with slop. So, let's upscale it:

Much sharper! Except look closer:

Yeah, that's not ideal either. Thankfully we can use img2img to improve the quality of the lines. I have a multi-step process for all of my upscaled images to get them at the quality we need for training. First things first we'll be doing an img2img pass at denoising strength 0.3 on the NTRmix checkpoint. Here's the output:

Yep, it's nearly exactly the same. The changes are downright microscopic, but this stage is important because we'll need to manually polish the image every step of the way after this. This first pass is just to make the lineart slightly less bad without potentially breaking the image.

Alright, that was the last easy step out of the way. Next we'll be doing an img2img pass at denoising strength 0.5 on the NTRmix checkpoint:

And you can already see something's been screwed up, but there's often hundreds of smaller errors you won't notice at first too. So, I overlay the 0.3 with the 0.5 and go over every single inch in excruciating detail, erasing the areas of the 0.3 where the 0.5 looks better, ending up with a fusion of the two. This gives me the line sharpness and detail of the 0.5 without any of the AI jank:

We are now in the polishing stage. This is an extremely time-consuming process. I can lose hours of the day polishing up a handful of images.

I'll keep doing 0.5 ntrmix passes until I feel the image is just about perfect, at which point I need to make the image less shiny. Remember that small AI quirks go unnoticed, but when they start compounding they become a big problem. NTR mix has a habit of making skin and fur look very glossy, almost greasy. It also tends to give us low-saturation output. So we need to counteract that by doing yet another img2img pass, this time with another checkpoint. I use mistoon for this, as it's much less shiny and much more saturated. So, let's do a pass with mistoon and do the exact same thing as earlier. Overlay, erase, polish:

I also need to keep track of how far along each image is. Doing an 0.5 ntrmix pass before the 0.3 one will result in some blurry lines slipping through, and the mistoon pass must be done last. It can be helpful to keep different batches in different folders for this reason.

On top of all that, if I notice a major anatomy error at any stage, it can't easily be fixed. I need to reduce the size of the image and go aaaaaall the way back to the inpainting step, and do everything over again. And each polishing pass has the chance of some new anatomy error slipping through.

But we still aren't done. There's four major areas that are so important that I have to handle them manually at this stage. 3 of the usual suspects aren't issues with this image, thankfully. These are:

-1: Light outlines. Sometimes, right around the black outlines of our characters there will be a lighter outline where the color of the background meets the black of our outlines. Here's what this would look like if it were present in our image:

-2: Stray hairs. Small messy hairs will interfere with out ability to remove backgrounds, so I try to get rid of them. Here's an example:

-3: Bad hands. Self explanatory. Training your model with images with bad hands will, shocker, make your output images have bad hands too.

-4: Bad eyes. Finding a consistent eye style for the model is extremely important for quality. So, for every single image, I often just erase them and draw in new ones myself. Sometimes the eyes already look fine but I don't want to risk it. I'll also take this final opportunity to redraw anything else I feel looks a little jank, like sharpening the character outlines, teeth, hairline, fixing their eye direction, etc.

Note: you don't need it to look better than the AI, your final drawing can look a lot worse, but so long as it's your own weirdness and not AI weirdness, the final model will turn out better. A mistake in 1 out of 500 images won't have a noticable impact, but the AI doesn't make a mistake just once, it makes the same mistakes over and over, just in barely noticable ways.

After redrawing the eyes I then sharpen up the outline, etc. Whatever it takes to fix as much AI jank as I can find and finally, we are STILL not done. Because now the image needs to be retagged. For example we were making the image the 'femboy' tag was helping us get Yu's iconic look, but the final image doesn't look all that feminine. We need to go through and find every single tag that is no longer accurate and remove it. Then we need to go through the tag database and find everything missing, like 'areola slip'. On top of that, anything we don't want in our final model needs to be tagged or manually fixed. Skin still shiny? It must be tagged. Still some stray hairs? Tag it. And I'll remove any redundant tags like 't-shirt' when we already have 'short-sleeved shirt' too. Here's our cleaned-up taglist:

meme face, 2koma, upper body only, v-shaped eyebrows, white border, smirk, looking to the side, duo, leaning forward, outdoors, blue sky, animal crossing, trees, clouds, daytime,

1boy, cute boy, femboy, brown hair, short hair, red shirt, short-sleeved shirt,

sy-shop, animal crossing, furry female, shortstack, domestic cat, feline, blue fur, monotone body, dark paws, anthro, pricked ears, dark ears, light blue hair, hair tuft, thin eyebrows, red eyes, pink nose, grin, cleavage, areola slip, red apron, naked apron, sideboob

On top of that, when I have multiple different characters in one shot, I'll usually remove any tags that are core to their design and outfit. That makes group shots a lot easier in the long run. So for example bluebell should always have her natural hair color, so I don't need the 'light blue hair' tag.

Okay, we are finally done. One image down. Just...

A lot more to go.

And if you're wondering how long I've spent on the database as a whole, in my finished LoRa training collection, I currently have 1135 fully polished, cleaned, and retagged images. Now also keep in mind that I need to train the model, then test it thoroughly to find what's been mistagged, test if the training data is too shiny, not saturated enough, etc.

I'm sure this all sounds quite boring, but I actually find it very relaxing. I used to draw a lot, but for health reasons I can only draw in short bursts these days. I'm very grateful I have time to spend messing around with all this, thank you! And sorry for all the times I've fallen down the rabbit hole losing entire days to polishing images :P

VIP (Very Important Poll)

Salutatioooooooons~

How's it going everyone? I was sick for a bit, did I miss anything?

Well, I won't dally too long, I've been very productive so we've got a lot to cover!

Very important poll

As you all know, I've been working on the non-furry version, but I've reached an essential branching point. There's no way to satisfy everyone, so we'll just have to settle this with a poll. Perhaps the most important poll I've ever run.

For characters like Helena, and the Twins, who have short hair...

Should they have human ears, in addition to their fluffy ears?

Please choose your answer with the utmost care. The fate of Syrup Town's success, the chance that it will eventually achieve mega-stardom and become super popular may very well rest in your hands.

v10 Progress

Progress on v10 is going swimmingly! A bunch of images are finished, many more are nearly done. v10's new scenes will include:

• 3 new scenes where you train Helena to last longer/be more dominant on the road to eventually knotting you.
• 2 new scenes for Khanna, one being the cut scene from v9, the other being her wall scene (you're the one stuck in the wall this time!)
• 6-ish new scenes of the deer family's home invitation and repeatables, including petting.
• 3 petting scenes for Bluebell, things get just a liiiittle bit crazy...

Here's some previews of these:

Jiggies

Subscriber batch 4 is just about complete. I still have a few requests left, though I may end up leaving those for subscriber batch 5 in the next release. All 38 new jiggies are being tested on the PTR, though I haven't come up with names for any of them yet.

The final meaty jiggy is of Yu, so I made 5 additional variants so the jiggy would reflect your gender and skin color. Please look forward to it!

Subscriber requests

Howdy guys! I finished the next prototype of my model. The jiggies aren't great representations of it since I polish them up manually so much, but here's a few test shots:

https://x.com/NoodleJacuzzi/status/2011694689452507278

(Just some test shots so it's okay for me to be lazy right? :P)

Surprisingly, it works best with the WaiillustriousSDXL and Mistoon checkpoints. The main reason I trained it is because I really want it to get eyes right, that way I can use it to fix the eyes of my remaining unpolished images instead of needing to do it all manually.

But the real meat is that I've been able to use this new model to do a bunch more subscriber requests! These will be added with v10 as the next subscriber batch. Speaking of that, I need to do some polls soon.

Anyways, I still have a bunch in the oven, but here's what's cooked:

[Syrup Town] v9.1 Bugfixes

Hallo hallo~

As always, I found some bugs. It would be nice to put out a bug-free release, but I'm just glad they work at all.

Here's a v9.1 bugfix patch:

https://mega.nz/file/LqBXgZKJ#Z72ueErkeiCdUPDErm-bs_ojhwmYMu_EYIWJu1MLLHc

Misc changes:

• Replaced the old single achievement for collecting jiggies with ones for each batch.
• Completing the every jiggy in the Syrup category or any subscriber batches will unlock the random and custom jiggy that used to require every single jiggy in the game to be beaten.
• Individual jiggy categories will not have completion trophies, instead completed categories will now have white text and a crown.

Bugfixes:

• Fixed a bug where a number of characters would still have their repeatable petting scene locked after asking to pet them.
• Fixed a bug where petting Helena or Marlow would prevent you from petting Mary-Lou.
• Fixed a broken image for an old Sorbet outfit
• Fixed many many bugs and typos found by users such as R08013r and OrangeMaestro

I have already updated the links in the v9 release post, but here's all the download links again:

Syrup Town v9.1 Full Download (webp version): https://mega.nz/file/H3R1xBKC#U03GNVYzEgSDs_l1jZvOigSUyG2AKb7aqW10Kl-MbtA

Upgrade Pack from v8 to v9.1 (webp version): https://mega.nz/file/G74RDBZZ#I_9ZgwBSl6XGMhoDexWY5GRFQipizyZRZUnr-R7urGE

.apk Version for Android Devices: https://mega.nz/file/vuxRgJ4Y#653dMAbObktQK9DmQpAhC_intYzUtw2HZpMnaJlMY5U

Please accept some Sorbet for your troubles:

[Syrup Town] [Release] - Syrup Town v9

Hallo hallo my friends! Happy birthday to meeee~

Play Syrup Town online here: https://noodlejacuzzi.neocities.org/syrup-town/

Play the mobile version online here: https://noodlejacuzzi.neocities.org/syrup-town/mobile

Syrup Town v9.1 Full Download (webp version): https://mega.nz/file/H3R1xBKC#U03GNVYzEgSDs_l1jZvOigSUyG2AKb7aqW10Kl-MbtA

Upgrade Pack from v8 to v9.1 (webp version): https://mega.nz/file/G74RDBZZ#I_9ZgwBSl6XGMhoDexWY5GRFQipizyZRZUnr-R7urGE

.apk Version for Android Devices: https://mega.nz/file/vuxRgJ4Y#653dMAbObktQK9DmQpAhC_intYzUtw2HZpMnaJlMY5U

EDIT: As always, I found some bugs. Here's a v9.1 bugfix patch: https://mega.nz/file/LqBXgZKJ#Z72ueErkeiCdUPDErm-bs_ojhwmYMu_EYIWJu1MLLHc

All of the above links and the PTR have been updated to v9.1 as well.

Upgrade packs for the PNG version will be uploaded later. I'll likely be retiring it in the future now that I know the non-furry version and the femc routes will need to use different script files, and I don't have the manpower to keep track of / test all these different versions.

Content:

• New Character pair: Nutmeg & Cinnamon, the mother daughter duo
• Added a repeatable petting scene for Angelica, Cayenne, Sorbet, Mary-Lou, Marlow, Riley, and Helena
• You can now visit Khanna's church
• Added the DQ and Holofuta magazine images as jiggies

Other changes:

• Reworked the jiggy game to make more interesting shapes
• Misc jiggies are no longer required for the jiggy completion achievement
• Changed the way the collection room's code functions, so it should autosave after completing jiggies.

Bugfixes:

• Replaced an image in Bluebell's meaty magazine where she had a pusspuss
• Fixed the missing bar jiggy

Upcoming:

• With all core characters being added and using their new avatars, I can finally get the ball rolling on the non-furry version.
• I'm several drafts into v9 of my generative model, so expect an update to the guides in the technology wing.
• A new tellyvision show starring an emo girl investigating weird critters.
• Multiple scenes and jiggies are already finished, but I didn't have time to finish/test them. So expect another update soon.
• More monster fucker permit usage.
• Adding the Plap Pal cards to the game proper.

Oh man, so much had to be left on the cutting room floor! Scenes and jiggies with finished images had to be dropped, and there are a few causes like my internet going out for that whole week, etc. But the more interesting delays were:

• Nutmeg and Cinnamon, oh boy. Getting their scenes done was 100x harder than expected. Not one, not two, but three threesome scenes where only one character is a human were headbendingly difficult. Even their non-threesome scenes all had at least one element that made them harder to generate too...
• I'm several drafts into a finished version 9 of my generative AI model. Even the janky prototypes made the new scenes a lot more bearable to work on, and it'll make the non-furry version so much easier it's unreal!
• Finagling the jiggy engine into making better shapes was harder than I budgetted for. Hopefully it's a lot more fun like this.

But I really wanted to release today! It's my birthday, but this is my present to you. I hope you'll enjoy!

Thankfully I'm way too lazy to let it all go to waste, so I'll definitely finish all the unfinished stuff soon. Like this followup to the confessional scene:

And a magazine of the non-furry cast:

And new jiggies I didn't have time to add (I was genuinely working on this to the last hour):

One last note: all the petting scenes can be triggered just by talking to the character (I'll add Bluebell and the deer later), except for Angelica. Hers can only be triggered if she isn't horny, so just do her repeatable and then talk to her again after.