- Oct 16, 2023
- 18
- 3
HTML Syrup Town [v1.5] [NoodleJacuzzi]
- Thread starter NoodleJacuzzi
- Start date
- May 20, 2019
- 707
- 1,632
More posts! I would really have loved to have responded right away, but for some inane reason gmail decided to move all emails for things like subscribestar and f95zone to a new tab and stopped notifying me of them. Thanks google!
So I've been working on v2, I have people testing sharly right now actually, but the long delay was caused by (among a few other personal life matters) me finally creating an SDXL LoRa for syrup town!
It would have been a lot easier to just straight up use my non-furry lora but I really like the softer nature of syrup town's current art and didn't want the shift to be -too- jarring. The ideal would be literally syrup town's current artstyle but with sharper lines and fewer anatomical errors.
So I went through every single image in the game and many other ones I found to be the best from my unused batches, cleaned them up heavily by hand, re-tagged them so they were more accurate, and then used Img2Img with various models at low denoising to get shots that I think still match the general vibe of syrup town but are much higher quality:
Some direct before/after shots:
Then I realized that the huge amount of Angelica, Riley, and Sorbet in particular would throw off the model by overfitting it to their details so I did the same thing with every one of my pokemon images and the tarot cards:
And the model can do even more pocketmanz designs as well:
You can find a link to the model and all the tags I used for it (and plan to include in the training data in the future) here, along with a link to all the training data as well:
In total making the lora required cleaning and retagging about 850 images, and while I'd really like to do a thousand more to get even better quality I called it here for now so I can get back to work on the game.
Notably the lora has trouble with Bluebell's hair, it likes to turn Marlow's ears and tail brown, it messes with Angelica's fur giving her a white fur pattern a lot, and I'll need to completely redesign Mary-Lou in the future as her current design feels really weird in the newer style.
I don't think it's a perfect improvement. There are some areas where I prefer the lighting style and more abstract backgrounds of easyfluff, but my main goal was to create a model that would let me:
- Generate the characters reliably and consistently, as when using easyfluff there are substantial detail and anatomy errors in around 60% of images.
- Use plugins like regional prompting so I can do threesome scenes.
- Generate profile shots with sharper edges so that I can use rembg to remove backgrounds instead of needing to remove them by hand.
- Take advantage of sdxl's larger database so that I can generate syrup characters doing things like footjobs, which just wasn't happening with the easyfluff model.
And the fact that I now can do all that is a big win in my book! So while it was a lot of work I think it was worth all the effort.
With every CG editted to match the new style I've spent the last few days generating new expression sheets for every character. A lot of the older ones had pretty notable errors that stood out a lot when you compared one costume to another in the museum. All 17 characters (dick vs no-dick system chars are treated as separate characters), 13 expressions each (I added a new 'blush' expression I felt was lacking while writing Sharly), and every character has at least two costumes.
At each stage of the process I used the scripts I made for making expressions for the non-furry cast, so really the only hard parts were:
- I generated 48 total base images for every character/outfit, and had to narrow it down to just one
- After generating the variants of the base image to use for the expressions I had to clean up details like breath clouds, floating hearts, and make sure details like clothing strap colors, hair styles, and tail patterns were consistent. This was a lot of manually cleaning things up.
As I write this I'm upscaling all of the images before I use remBG to remove the backgrounds, so the last bit of hard work will be going back in afterwards to clean up any areas remBG missed. Because the new model works so consistently I was able to img2img the whole character instead of just the face without it ruining the shots, so characters now leak precum in their ahegao portraits or sweat in their excited ones, for example.
I was actually able to try out lots of different new costumes as well:
And once all of that is finished I'll re-sort everything into its proper place, downscale it to reduce the massive filesize (SDXL hires images are almost 4 megabytes a pop) and put it all on the PTR for testing before getting back to work.
Sharly's main story like I mentioned is totally finished and all the framework for the rest is done, it's just scenewriting for the rest of the fluffies left now.
By the way while I was searching through my files I found a bunch of the very first shots I used of the syrup cast and made a little image comparing the oldest style to the newest one (sorry about the weird spacing I've used a lot of different image sizes over time):
And another to try and show off some of the game's major styles (fluffyrock with a generic lora, fluffyrock with the first lora I trained, fluffyrock with a vastly improved lora, an older version of easyfluff, and finally the latest SDXL lora)
Good gravy technology certainly has come a long way! I'm very grateful to have a patient community, if I'd stopped trying to constantly improve and better understand the tech I never would have gotten this far.
While there is a character who theoretically could do the swap, Helena, the one who actually owns the pocketmanz cards in the first place, I think trading the variants manually sounds tedious, especially because the only purpose for getting all the cards is to bask in a full collection.
While the new model might be able to generate decent feminine versions of specific mons some definitely still don't work, incineroar always wants pecs and a masculine frame. And I still have no idea how I'd make a dickgirl version of eevee among many other ferals.
Unironically I had a reference to this as one of the bug-hunting encounters planned:
But I scrapped it as I couldn't find a good way for easyfluff to do dinosaurs.
Please for the love of goodness do not hold me to expanding on the terrible ideas I have while writing at 3am. We barely avoided the timeline where one of the prisoners in anomaly vault was an american-flag bikini wearing bombshell who turned people into hamburgers:
To be honest every time I look at the oldest images I generated for the game I wonder how I ever thought "yeah, this could work for a game". I promised I'd eventually make a non-furry version to go alongside it so that people would stick around while I worked on the most cartoony goofy nonsense I don't understand how I thought I was cooking:
And while it can take me quite a while I try my best to stick to my word eventually, so the non-furry version is something I'm sticking with.
Plus I can use the models I create for the non-furry version on other games in the future, like rainy dayz:
Or argent science:
So I've been working on v2, I have people testing sharly right now actually, but the long delay was caused by (among a few other personal life matters) me finally creating an SDXL LoRa for syrup town!
It would have been a lot easier to just straight up use my non-furry lora but I really like the softer nature of syrup town's current art and didn't want the shift to be -too- jarring. The ideal would be literally syrup town's current artstyle but with sharper lines and fewer anatomical errors.
So I went through every single image in the game and many other ones I found to be the best from my unused batches, cleaned them up heavily by hand, re-tagged them so they were more accurate, and then used Img2Img with various models at low denoising to get shots that I think still match the general vibe of syrup town but are much higher quality:
Some direct before/after shots:
Then I realized that the huge amount of Angelica, Riley, and Sorbet in particular would throw off the model by overfitting it to their details so I did the same thing with every one of my pokemon images and the tarot cards:
And the model can do even more pocketmanz designs as well:
You can find a link to the model and all the tags I used for it (and plan to include in the training data in the future) here, along with a link to all the training data as well:
You must be registered to see the links
In total making the lora required cleaning and retagging about 850 images, and while I'd really like to do a thousand more to get even better quality I called it here for now so I can get back to work on the game.
Notably the lora has trouble with Bluebell's hair, it likes to turn Marlow's ears and tail brown, it messes with Angelica's fur giving her a white fur pattern a lot, and I'll need to completely redesign Mary-Lou in the future as her current design feels really weird in the newer style.
I don't think it's a perfect improvement. There are some areas where I prefer the lighting style and more abstract backgrounds of easyfluff, but my main goal was to create a model that would let me:
- Generate the characters reliably and consistently, as when using easyfluff there are substantial detail and anatomy errors in around 60% of images.
- Use plugins like regional prompting so I can do threesome scenes.
- Generate profile shots with sharper edges so that I can use rembg to remove backgrounds instead of needing to remove them by hand.
- Take advantage of sdxl's larger database so that I can generate syrup characters doing things like footjobs, which just wasn't happening with the easyfluff model.
And the fact that I now can do all that is a big win in my book! So while it was a lot of work I think it was worth all the effort.
With every CG editted to match the new style I've spent the last few days generating new expression sheets for every character. A lot of the older ones had pretty notable errors that stood out a lot when you compared one costume to another in the museum. All 17 characters (dick vs no-dick system chars are treated as separate characters), 13 expressions each (I added a new 'blush' expression I felt was lacking while writing Sharly), and every character has at least two costumes.
At each stage of the process I used the scripts I made for making expressions for the non-furry cast, so really the only hard parts were:
- I generated 48 total base images for every character/outfit, and had to narrow it down to just one
- After generating the variants of the base image to use for the expressions I had to clean up details like breath clouds, floating hearts, and make sure details like clothing strap colors, hair styles, and tail patterns were consistent. This was a lot of manually cleaning things up.
As I write this I'm upscaling all of the images before I use remBG to remove the backgrounds, so the last bit of hard work will be going back in afterwards to clean up any areas remBG missed. Because the new model works so consistently I was able to img2img the whole character instead of just the face without it ruining the shots, so characters now leak precum in their ahegao portraits or sweat in their excited ones, for example.
I was actually able to try out lots of different new costumes as well:
And once all of that is finished I'll re-sort everything into its proper place, downscale it to reduce the massive filesize (SDXL hires images are almost 4 megabytes a pop) and put it all on the PTR for testing before getting back to work.
Sharly's main story like I mentioned is totally finished and all the framework for the rest is done, it's just scenewriting for the rest of the fluffies left now.
By the way while I was searching through my files I found a bunch of the very first shots I used of the syrup cast and made a little image comparing the oldest style to the newest one (sorry about the weird spacing I've used a lot of different image sizes over time):
And another to try and show off some of the game's major styles (fluffyrock with a generic lora, fluffyrock with the first lora I trained, fluffyrock with a vastly improved lora, an older version of easyfluff, and finally the latest SDXL lora)
Good gravy technology certainly has come a long way! I'm very grateful to have a patient community, if I'd stopped trying to constantly improve and better understand the tech I never would have gotten this far.
I think your first sentence there got cut off.
While there is a character who theoretically could do the swap, Helena, the one who actually owns the pocketmanz cards in the first place, I think trading the variants manually sounds tedious, especially because the only purpose for getting all the cards is to bask in a full collection.
While the new model might be able to generate decent feminine versions of specific mons some definitely still don't work, incineroar always wants pecs and a masculine frame. And I still have no idea how I'd make a dickgirl version of eevee among many other ferals.
Probably not, I'll probably run out of ideas for porn scenes long before I run out of characters who need content. The random encounters are just a space to do random things for one-off ideas, gags, or references.
Unironically I had a reference to this as one of the bug-hunting encounters planned:
But I scrapped it as I couldn't find a good way for easyfluff to do dinosaurs.
Please for the love of goodness do not hold me to expanding on the terrible ideas I have while writing at 3am. We barely avoided the timeline where one of the prisoners in anomaly vault was an american-flag bikini wearing bombshell who turned people into hamburgers:
Yes but I believe they were specifically talking about certain vibes not appealing to certain people, so trying to make a version of eevee that appeals to folks who would disable male content. Trying to decide if such a mode should disable characters like nutmeg, riley, and cayenne is hard given their very androgynous nature, it's even harder when you're working with ferals like eevee.
Yes someone coming into animal crossing rule 34 would probably not be interested in the non-furry version, but there's plenty of people who come to syrup town from my other games and if they just plain dislike furry art then the non-furry version gives them a way to enjoy the game.
To be honest every time I look at the oldest images I generated for the game I wonder how I ever thought "yeah, this could work for a game". I promised I'd eventually make a non-furry version to go alongside it so that people would stick around while I worked on the most cartoony goofy nonsense I don't understand how I thought I was cooking:
And while it can take me quite a while I try my best to stick to my word eventually, so the non-furry version is something I'm sticking with.
Plus I can use the models I create for the non-furry version on other games in the future, like rainy dayz:
Or argent science:
Rainy dayz sounds like a resident evil genre and argent science can be any form of science fiction of all types.
- May 20, 2019
- 707
- 1,632
I think what you're saying is that games inspired by resident evil and science fiction would have very different tones from the non-furry version of syrup town? Apologies if I'm misunderstanding you.
That is true, but training an AI model revolves around two things:
1. Exposing the AI to known concepts, which is basically saying to the AI "You know X? This is how you do X in this style." This teaches the AI how to draw the things it already knows in the desired style.
2. Exposing the AI to new concepts, saying "This is Y in this style". This teaches the AI how to interpret the new concept and expands the total amount of concepts it can create, teaching it to draw new things.
SDXL knows natively how to draw anime-styled hair, shortstacks, shirts, and so-forth, so including those things in the training data teaches it how to draw in the style I like. But in order to do that I need multiple images in that style and I need to take the time to clean up any errors in that image. Making Syrup Town's non-furry version means I have a large trove of varied images that I've already spent a lot of time cleaning and that multiple others have looked at, giving me more eyes to find flaws.
From there I apply that style to new things teaching it bit by bit. The model now knows how what I mean when I ask for a penis, standing shot, and sneakers, so when I need to teach the AI how to do vore for rainy dayz I give it images where only needs to learn how to draw one unfamiliar concept, the rest is already in the training data, and so with only a bit of work I get:
And since my non-furry syrup lora has so many images I was able use it at a super low denoising level to clean up the furry version's images. There would be no furry SDXL model if not for the work I did making the non-furry version. For one example this means if the non-furry version of the game didn't exist, making a threesome scene with jasper and garnet would be unfeasible (because I'd need to use regional prompting, an extension that doesn't work with easyfluff but does work with SDXL)
Beyond the training are the scripts I've written for the non-furry version. Easyfluff had very messy outlines so there wasn't any point in learning how to use RemBG to automatically remove backgrounds, but SDXL models have much crisper lines so I learned how to use it, created scripts to upscale things to SDXL resolutions, scripts to generate full expression sheets, scripts to remove and re-inject PNG data, scripts to convert e621 tags to danbooru tags, etc. Among other things I used these scripts for generating the new syrup expression sheets, and hopefully once I've made them less jank and shared them, others will use these tools and the guides I've written to make their own games (and avoid the pitfalls I've already stumbled into)
In short I work on a variety of things, my understanding of the technology improves, I apply that understanding to make more training data, I get a better general use model, it forms the foundation for other more specific ideas. Just like how working on Syrup Town gave me more javascript knowledge that I'll apply to my other games.
Last edited:
Those images from both types still looking good and it wasn't a bad idea to use some from RD in non-furry if this will make your job easier. But since there was that big delay, no one can be sure what will be next. And I know I asked that before, but anyways, What's the probable date of release? Like in 1-2 months or longer maybe?
Indeed, you have misunderstood me.
I made those guesses because the former reminds me of DayZ, which originated from a mod from Arma 2 and the latter is a pretty basic science fiction title. It's more of a basic observation than any relation to syurp town.
I do appreciate the non-sequitur SDXL explaination.