Magical Working Girls, season one fin

Following last week’s cliffhangers, Our Magical Trio explores the limits of bullet-hell boss fights in a show that really isn’t set up for dynamic combat but gives it their best shot. TL/DR, the Power Of Three saves the day, and we montage our way through the rest of Guest Gal’s training visit, wrapping up on a happy note.

Until the after-credits teaser for next season, which raises the stakes on Boss-chan’s Secret History with His Old Rival.

Verdict: this was fun; lots of budget-cutting shortcuts, but they managed to make the big final battle look good and occasionally frantic (as opposed to some earlier episodes where the monsters just stood there waiting for their puzzle to be solved). Looking forward to more.

GGO 2, fin

And with the resolution of the Totally Unpredictable Plot Twist That Everyone Figured Out Weeks Ago, we reach the best part of the show: a montage showing all the characters interacting in real life, showing off all the personality and character growth that was basically absent in-game.

Verdict: yeah, it would take a lot for me to watch a third season of this. Mostly because Shigusawa is too much in love with his creation. And I mean Squad Jam, not Karen or Llenn.

Stick a fork bullet in it, it’s done.

Promo video for Yandere Dark Elf

Just in case we have another weak season in the Spring. I mean, they went to all the trouble of making her giant breasts bounce all over the place, so they seem to be major characters.

“Hey, ChatGPT…”

Generate a set of tropes for a new light novel series where the protagonist is a human female college student who acquires a harem of females, one per book. First create the main tropes for the series, then a set for a new girl introduced in each book. There are ten books total.

The results were surprisingly plausible, coherent, and consistent, and followup questions adequately fleshed out the characters, college, town and complications, both for the series as a whole and each book in it, and it created perfectly-reasonable first and last scenes that were consistent with all of the above. It refused to write a final harem sex scene, however, until I caved:

Write an alternate final scene where all characters have sex together, leaving out any elements which might violate your terms of use or usage policies.

Is it any good? Not really; the explicit biases and forced limits pretty much make it shallow and obvious tripe that stresses “diversity”, avoids non-approved viewpoints and facts, and follows a completely generic modern novel structure. More importantly, this sort of thing is the easy part of writing a novel; helpful for people who lack imagination, determination, or a way to keep their thoughts organized, but it bears the same resemblance to good prose fiction as my occasional doggerel bears to professional songwriting and poetry.

However, you can at least tell it to do things like “rewrite the opening scene in the style of Roger Zelazny’s ‘Amber’ novels”, which I wouldn’t call a huge success but was definitely less generic.

Going meta…

I started with the simple prompt “Create a title for a series of isekai novels”, iterated on it a few times, and then asked ChatGPT to consolidate all my instructions into a single new prompt:

Create 10 long, detailed isekai novel titles, each incorporating at least five anime tropes, where the protagonist originates from a completely different isekai with its own distinct set of tropes. Each title should vividly describe the premise, character dynamics, and the clash between the protagonist’s previous world experience and their new one in a single run-on sentence.

Here’s what happened:

I’ve seen worse. I’ve even read worse. Masochistically.

Speaking of which…

---

Write a back-cover blurb in the style of Doc Smith for an isekai light novel with the title: After Reigning as the All-Powerful Demon King in a High-Magic Isekai, I’m Summoned to a Mecha Sci-Fi World Where My Magic-Based Tactics, a Genius Loli Mechanic, and My Frustration With Advanced Tech Become Humanity’s Last Hope.

Stable Diffusion cover art:

Stable Diffusion struggles with differentiating between subjects in the prompt, blending all of the descriptive elements into one characer, often randomly generating additional characters who possess some of these traits, whether you want them or not. Meanwhile, most checkpoint models are overtrained by their makers with a specific style or type of content in mind, making them prone to devolve into a pile of randomly-connected body parts. But if you want to produce uncensored and/or anime-style content, you need to pick one and learn its quirks (I’ve tried about two dozen, and Citron Anime Treasure has been very versatile and reliable, despite some weird random objects appearing if you do not put “censored” into your negative prompt have the word “award-winning” anywhere in your prompt).

And the bias, homuhghod the bias. The people creating checkpoints first try to remove perceived (sometimes actual) biases in the base model (like diversity, lack of sexytimes, etc) and then inject things like the danbooru tags database and tons of manually tagged images that reflect their fetishes. And, frankly, most of them aren’t good at it. Scratch the appropriate itch enough of the time and they get an upvote on CivitAI, but lurking in the shadows are tags that can summon hideous monstrosities from an explosion in a wax museum. Or CSAM.

Example of tagging-induced bias in the official model: ‘diverse’ is strongly correlated with ‘black’, and ‘black’ is frequently tagged on images of people with African skin tones and facial features. So, putting “a flowing black cape” on a demon lord will make him visibly African or African-American, or more likely her if the checkpoint was trained by a guy with a stash of porn. And it will bleed into any other human-ish figures in the scene.

Example of something that often happens with checkpoints trained primarily on pictures of naked women: a group of muscular men who all had detailed hairy pussies between their legs instead of sausage. I managed to trip over this on several different models when trying to produce a crowd of men looking at a female character (not gang-bangs, but things like trying to do a non-anime version of the Pole-Dancer Misty pic I posted).

Anyway, I gave up after an hour. The above is not the best mech-and-both-protagonists scene, but was the best compromise for all three. After I added “pussy” to the negative prompting; like I said, a lot of models are overtrained for certain purposes.

DALL-E, on the other hand, immediately produced a striking rendering that correctly presented both protagonists and the mech in a gorgeous, well-composed image, but the mech was demonic and the engineer was casting a spell. In the second image, the mech was even more demonic, and the engineer was shooting a blaster. Also, after generating 12 batches, it put me in time-out for five minutes. On a paid account. I needed more batches because every time I said “just like that, only make the mech non-demonic” or “make the engineer use tools instead of shoot guns”, it broke at least one other element.

Final DALL-E image before I gave up:

TL/DR, they both give the illusion that you are providing editorial direction to your AI agent, but you’re really just shouting at the television set because it can’t find a channel showing Wheel Of Fortune and keeps pulling up Blue’s Clues instead, dubbed in Swedish.

Bork bork bork.

Okay, one last try after closing out the Friday shows…

One really cool thing about models based on SDXL that haven’t utterly borked the tagging is that the original model has a lot of artists and styles in its tagging, and calling them out will make your results look something like their art.

I regained a bit of enthusiasm for this project when I started off my prompt with:

Art by {Jack Kirby}.

Anyway, here’s a Liru chaser:

And no, the person who made the character LoRA did not manage to get her distinctive boots into the model, and I never managed to get the straps connecting her top to her shorts, either. I had enough trouble just getting her top to stay red; tiger stripes were very popular. Honestly, I think this model is intended for getting her out of her costume, not that there’s anything wrong with that.

It looks like some of the nine Liru LoRAs trained for other base models have managed to get the fringe at the top of her boots, but no one has tried to get the feet right. Because their users are busy lewding her. I’ll probably play with the highest-ranked Pony LoRA for her at some point, possibly to see if I can get her to make out with Wolf Bete from Miss Kuroitsu…

---

Comments via Isso

Markdown formatting and simple HTML accepted.

Sometimes you have to double-click to enter text in the form (interaction between Isso and Bootstrap?). Tab is more reliable.

---

Type Comment Here (at least 3 chars)