Swan Songs By Xenos

Early in January (2025), I came out of my txt2txt LLM haze for a moment and noticed that a lot of my favorite model creators had, over the holidays, released final versions of their SD1.5 based checkpoints before moving on to bigger and better things in XL, Flux, and the plethora of other great new stuff out there. URPM, CyberRealistic, and Galena Blend (though adopted by Alexds9, I'm not sure it got an update, or if he just reposted after taking over maintenance) were some of the key SD 1.5 Swan Songs released. And I also stumbled upon FellDude's "HD Porn Finetune" (Toon Edition) and found it to be a really well done piece of work.

Feeling a bit nostalgic, I dove into my old A1111 merging install and spent nearly a week trying to cobble together a working version. (Nightmare! Shoutout to DucHaiten who has maintained an older version with Supermerger still capable of merging LoRas! Thank you!) Finally, with two barely working WebUI's, I was ready to have a little fun.

I started out with my Kitsch-N-Sync model, since it's already the mutt of my releases - the one where I had just shoved in every possible checkpoint and let the chips fall where they may. I experimented with some of the new merging methods on my up-to-date install and combined that with my various sanity LoRas and was really just expecting garbage to come out - something I might get one or two generations from and become frustrated.

Surprisingly, that's not what happened. This actually came out really awesome (for the most part, anyway).

THE VERSIONS

SwanSongsByXenos_V1: This is the main checkpoint. It runs GREAT with simple prompts - with plenty of whimsy, fun, and even the occassional surprise. As a "mutt" of a checkpoint, though, certain concepts run hotter and blow up a lot more easily - especially with heavy prompts.

SwanSongsByXenos_V1FINE: In this version, I used some enhanced interpolation methods and (lightly) merged the OG SD1.5 base model and had it look for hugely modified blocks and smoothed them back toward the base. I knew it wouldn't affect new training concepts introduced by various newly trained models, but I hoped it would serve to calm down common problems like hands blowing up (as easily) on overcooked concepts or special effects suddenly becoming weird.

The V1FINE version is considerably more stable, but also less creative and whimsical. You can use it as a regular checkpoint and it'll do fine, but I made it to serve as a "refiner" model for the V1 Mutt version. It's especially helpful for those shots that just don't quite "finish" nicely - and it also helps sharpen art style type backgrounds a bit - in most cases anyway.

Ultimately, I haven't had to resort to using the refiner very often unless I had a very complex set of things I really wanted to get working all together - such as the cover image with all sorts of random "swan" references. Feel free to use these how you want - just remember "V1FINE" is a bit more stable but less creative, "V1" is a bit more volitile but brings in much of the "fun".

GENERATION TIPS:

This is optimized (as best I could do with half broken tools) for under powered rigs. A little goes a long way with many tokens much closer together than they would be in a standard model. That doesn't translate to "fewer steps" needed, but it does tend to make "more stuff come out" with less need to prompt for it.

I've been having really good luck with Euler A but you should be able to get decent results with other samplers. For those of you still using SD1.5 out of hardware limitations, that's probably where I'd start.

Steps for simple prompts can be around 20, with more complex ones, you can pump that up to around 30 or even 40 and still get improvements. This is sort of a balance between "Slow Machine - How Long is This &$^#& Going to Take!?!?!" and "I made 40 minute Firefly renders in DAZ Studio for years, I can wait 15-20 minutes for a good generation!" People on better rigs - up the steps a bit.

For upscaling, I typically run that at half of whatever my main steps are. So 20 steps would be 10 upscale. 30 would be 15, and so on. I do "Latent" with a 0.4 or 0.5 denoise strength. Going higher CAN sometimes help fix issues - but other times, it can exaccerbate them. Starting at around 0.4 or 0.5 seems to give a nice balance that works most of the time, but you may need to tweak that and experiment with extra tricky stuff.

If you're using ADetailer - set it to save a copy before detailing. In some cases (especially with non and not-quite humans) the detailer can hurt.

NEGATIVE PROMPT

Hands, as with any mutt model, are sometimes fine, and sometimes.... not. There are lots of embeds for this, but most really change the image quite a bit. I've found that BadhandV4 does a pretty good job without changing too much. I also found a few other tokens that help (in the area of weird fingernails) but didn't want to create a new embedding because using them in the positive prompt (and actually doing things like painting and manicuring them) can actually work pretty well. It's just that whomever did that original "fingernails" merge did it a little too strongly and I haven't found the model used in order to back it out, some.

So, my default tends to be:

badhandv4, long fingernails, bad fingernails, deformed hands, unfinished hands,

Also... Pink Hair is often useful.... when you start going toward more sexy or cartoony things, the AI has a tendency to make blonde hair pink. So, putting "pink hair" into the negative helps prevent that AND also bleeds back a bit into the rest of the picture and will tone down it's tendency to make everything pink without totally removing it from the color palette.

So...

pink hair, badhandv4, long fingernails, bad fingernails, deformed hands, unfinished hands,

This is a good alternative, just remember it's there if you ever ARE wanting to have someone with pink hair in your shot.

---

Learn More/Do More

Please don't forget to Follow Me here on Civitai so you will be informed of new things.

Need help prompting, finding tools and techniques to accomplish specific things, or generally want to learn more about SD, Generation, Making your own LoRas, Checkpoints, TI's and more - join the Xeno Engine Discord - we're a growing group of creators from all over the world with various, though often a bit deviant, interests. (NSFW).

Please take a moment to like, collect, and review any of my models that you enjoy. Each of those things helps the model be more visible to the community. The more visible, the more likes and reviews and downloads, which leads to more buzz and more rankings. Eventually, I hope that my earning pennies for each dollar put into buzz will get me to the point where I'm making an extra $10 a week to cover my Friday PBR Six-Pack. I know... high hopes, but... I can dream, right?

The Only Dumb Question is the Unasked Question. This is an important rule that I live by. You might feel shy about asking for help with something that is probably obvious to most people who have been doing this for a while, but here's a secret... we ALL were at that "total newb" stage at one point - and heck, when it comes to prompting especially, I still consider myself a "somewhat skilled novice". If you want some help on something - it's only dumb if you don't ask. If you do ask - on site, on discord or wherever - you'll never be looked down upon by me (or any of the people on the discord, lest they be shut down like an illegal gambling operation.