The Tale Of Tanya The Tragically Misunderstood 2, episode 1

Minimal recap and right back into the meat grinder. Our Heroine knows exactly how much a Russian winter sucks for an invading army, and is not happy about being that army.

Verdict: exactly as promised. I hope that Mary Sue doesn’t consume the entire season, though; it’s been so long since I read the early books that I don’t remember how long an arc it is.

Bumpkin and Harem 2, episode 1

This one’s on Prime, which naturally tried to feed me the English dub. Fortunately I reached the remote before I was scarred too badly by the voice actor’s poor fit for the role.

I don’t know who picked the OP song, but they’re clearly not a fan of current trends. Not that I particularly like it, but at least the singer wasn’t autuned into another species.

The recap was blessedly short, and it got right back into the story, with Our Magical Legal Loli recruiting Our Bumpkin for another teaching job.

Verdict: off to a good start.

(still not clear how she maintains knight discipline in that outfit, unless she has also offers night discipline)

“Oh, to be in Shibuya in the Summertime”

(but not really, because Japanese summer weather sucks)

How to convince me to watch a new show…

Everyone on ANN hates Ass-Guardian with a passion usually reserved for Trump. Given that the reviews these people write tend to include lengthy digressions about how a show reflects their overexplained life experience or validates their overexplained mental illness or just general me-me-me oversharing, I take their opinions with the seriousness they deserve.

(the same ass-clowns downvoted Tan Teen Oni Waifu for “colonialism” and “performative masculinity” and other horseshit)

Amazon: got clutter?

Amazon’s home page now looks like a grocery-store weekly ad.

Instagram AI-by-default

Got a public Instagram feed? Congratulations, anyone can use your photos for AI. Unless you navigate their menus to figure out how to opt-out. Unless it’s too late.

90% of photography is editing

I’ve now released the first few sets from my Certified 1-Girl Archives GenAI image prompt project. I started with ~120,000 images downloaded over several decades, ran them through a vision model to create categorized Markdown-format lists, ran those through a Python script to sanitize the categories and massage them into cleanly-structured JSON, then weeded out things that were obviously genai output or simply didn’t feature women-above-the-age-of-felony. The final tally was 117,361 JSON prompts.

I shuffled that list, broke it into 500-line chunks, and used my SwarmUI CLI to generate 256x256 images for each prompt, using the current best-in-class Krea 2 Turbo model. Twice, because the first time I had trouble matching the failures back to their source, so I reran the JSON generator and had it add a UUID field to each prompt, which I used as the name of the output file. (generating the JSON took about 2 minutes; generating the images took over a day)

Then I started deathmatching the results, doing a very quick girl/not-girl pass. This eliminated almost exactly 5% of the images, even when I expanded the rule slightly to exclude images with any visible text or blocks of whitespace (like a colored sidebar on a magazine page).

One thing that struck me was that very few of the images were bad. When I thought about it, though, it wasn’t surprising: every magazine or photobook was edited before printing, and every web download was edited by me deciding which ones I liked enough to possibly see again.

Bulk downloads of entire photoshoots were not included in the dataset. I found a bunch of those on my NAS, but didn’t process them because they’re full of garbage. In the early, early days of the web, people edited their photo galleries because uploading was tedious and hosting was expensive. Once subscription sites became a thing, though, the model shifted to “new content every day, and customers bitch if you’re stingy”, so they uploaded every photo that passed even a minimal QA check. Basically, if your thumb wasn’t in the picture, it was good enough.

So I’m looking at a large collection of ZIP files featuring pretty women wearing very little, about 85,000 photos, and it would take a full five-star deathmatch to extract the decent ones, leaving approximately 1,500. Honestly, I don’t think the small amount of variety it could add to the dataset is worth the tedium.

When you have 6 takes of “girl standing by wall” that were shot over the course of 10 seconds, you might have 1 half-decent picture, but it might not be worth posting because you also have 6 versions of “girl leaning against wall”, one of which is more flattering. Etc, etc.

This is why I wrote my deathmatch script. You edit before you take the picture. You edit as you take the picture. You edit when you check the camera afterwards. You edit when you review the pictures on your computer. You edit when you rank and tag the pictures. You edit when you choose what to send to an editor/publisher/blog.

And then they edit by picking the ones that suit their needs, which if you’re getting paid might require suppressing your own preferences and giving them everything.

Then other people edit by downloading the ones they like.

J. Barry O’Rourke has two anecdotes in his book Photographing Women Beautifully: in the first, he comes back from his first location shoot and hands over six negatives, then spends the next six months getting teased by the editor for not delivering the full take. In the second, he was flown to a tropical resort at the peak of the season for a lifestyle piece, recruited models onsite, got a fantastic cover photo and shoot, and as soon as he turned it in the editor paid through the nose to send him back for a reshoot. Because the positioning of the pineapple on the front of the towel wrapped around the man’s hips was “too suggestive”.

(Krea 2 Turbo got the meme format right on the first try, and I just had to re-gen a few times to get the words 100% and keep the pole from changing position between panels)

First impression: I never want to hear the OP song again. Autotune is a privilege, not a right, and the producers of this song need a five-year ban. Pity, really, because the OP animation is well-done and does a good job of showing off the cast and situations.

The ED song and animation, on the other hand, are excellent, focused on Our Best Lizard Girl Theresia.

Right off the bat, Our Hot Former Boss demonstrates that her idea of swordsmanship came straight from the Gainax Dojo. Not that there’s anything wrong with that. What follows is a lot of pretty typical isekai worldbuilding, but I knew that was coming, and at least it’s necessary to the plot.

Verdict: I remain optimistic. Still a little odd to hear the voice of The Universal Boy Hero coming out of a grown man for once.

(no fan-art to speak of, but every gal’s an action gal)

Isekai Loner Harem season 2 announced

Here. Much less pervy than the original novels, in a good way. In the books, while he’s faithful to his three girlfriends, he subjects every other girl to well-lubricated magical-tentacle grope sessions in the pursuit of well-fitting fashion and armor. Not that there’s anything wrong with that.

Tomorrow we get to find out if I’ve been crossing my fingers for nothing, as Isekai Ass-Guardian debuts.

Skeleton Knight 2, episode 1

Our Hero is still a ridiculously OP RPG tourist, Our Titty Elf is still exasperated with his antics, Our Cuddly Beastgirl is still just a touch gay, Our Mascot is still expressing itself in “I am Groot” style, and the fan-art is still sparse and heavy on elf-service.

Verdict: should be fun, although I’ll never watch the OP or ED again.

(still had to open the season with slave raids, though…)

Magilumiere 2, episode 1

This one’s streaming on Amazon Prime, which was unable to search for the title. Fortunately I had the first season in my history.

Anyway, the story picks up right away, leading off with a bunch of dry plot coupons in a board meeting. Our Magical Gals don’t get to spring into action, but they do at least get to remind us of their personalities. Next week, we get our first look at Our Underrim-Glasses-Wearing Magical Engineer Gal.

Verdict: really pushing the story hard in this one, and since I’ve been keeping up with the manga, I can say that they have a pretty good one to tell.

(no new fan-art, sadly)

Level 2 Cheat Life With Wolf Waifu 2 announced

This show was very silly. Eventually the source material kind of ran out of interesting things to do, a common problem when the hero is so overpowered that gods start stopping by to ask for help. Maybe there’s still enough for a second season to be fun.

Bash-ing Agentic AI

…by using Bash shell syntax to hide malicious commands. Because these glorified cron jobs will believe anything they read, and then execute it.

Fastest. Tsunaround. Ever. Also, that wasn’t a turtle.

(unrelated horn-y gal will have to do until the fan-artists get to work on Alna)

Maomao: The Movie trailer

Here. TL/DR: it’s a new story by the original author that doesn’t conflict with the continuity in the upcoming Fall and Spring seasons.

The only thing worse than woke dubs…

…is Youtube’s AI trailer dubs. At least, I hope they’re AI, because the alternative is that non-Disney dub actors are even less talented than I remember from the last time I was exposed to their product.

(and yes, fuck-you-youtube for turning them on by default; that and auto-continue-to-unrelated-video are almost as annoying as your ads that I block)

Season’s Greetings ’Splodies

There seems to be some noise outside…

Project Hail Mary

Now on Prime Video. I didn’t think the book was as good as The Martian, but I guess I’ll find out how the movie is.

Based on the ‘Recommended For You’ page…

…Amazon thinks I have a six-year-old autistic daughter who only eats sustainable organic vegan snacks and drinks alkaline water, and who demands a constant supply of new fidget toys. Yeah, no. Well, not that I know of, which is the same thing when it comes to shopping.

Quality Control

“I’ve got 106,000 problems, and a 1girl is almost every one of them.”

I’m currently generating small (256x256) images for every one of the JSON prompts I’ve extracted from photos on my laptop and NAS, to see which ones should be removed from the final dataset. It takes about 1.5 seconds to generate each one with Krea2, so it will run for a while:

split -l 500 current.jsonl
start=1;
for i in x??; do
    echo $i
    mkdir -p out/$i;
    jq -r -s '.[]|.json' $i |
        sui -j -a 1:1 --pre out/$i/krea2 --set wctest \
            --seq $start --pad 6 gen -r krea2small -c 500
    start=$(($start + 500));
done

Thanks to the miracle of sorting, the output starts out with pics from the old late-Eighties/early-Nineties mail-order fashion/lingerie catalog Mello Mail. The vision model really captured the style, to the point that you not only can identify the era by the clothes and hairstyles, but I can even tell which pictures were based on my friend Carmen.

I’ve already found a few places where I can strip out more text from the generated prompts. I’m just opening up each batch of 500 in the MacOS Preview app and setting it to “contact sheet” mode, so I can quickly scan through the batch for things that either have text or don’t have women. It will also be obvious when I trip over anime cheesecake and GenAI fails, both of which will be removed.

I’ve only found one actual non-1girl prompt so far, and it’s because the LLM analysis was spectacularly wrong. Original: pair of large breasts viewed from above, with bra pulled out to show the nipples. LLM analysis: “a baby’s torso, viewed from above; adult hand with visible fingernails, gently holding the fabric at the sides of the baby’s torso; camera slightly elevated, looking down at the subject”.

NSFW

Dear Apple,

Please add back the whitespace that used to be at the edges of your windows. It’s really annoying to try to find the “drag” and “resize” areas when the content starts three pixels from the edge of the window. Either I end up resizing a window when I’m trying to select text, or selecting text when I’m trying to resize the window.

Seriously, I understand that you hate the entire concept of having multiple windows on screen, between the aggressive push for full-screen mode, the aggressive push for tabs, and the complete removal of window management options other than crude iPad “tiling” like Windows 1.1 is once again the height of fashion, but give the content room to fucking breathe. Make Whitespace Great Again!

(this public service announcement brought to you by everyone old enough to remember when Apple had evidence-based standards)

Isekai Mighty Maid, episode 1

Reminder: this is from the same author as God’s Shouty Appraiser, so my expectations have been set accordingly. The most important limitation is that there is not a single catgirl in this former game-world.

The primary twist to the “reincarnated in my favorite game” theme is that Our Maid-Obsessed Heroine never heard of the game, and has no idea she’s the main character. Meanwhile, a good chunk of the cast was also isekai’d, and has obsessively-detailed knowledge of every possible path through the story, and she’s not playing it right! (it’ll be a while before we meet all the reborn characters)

How is it? Well, I could do without ever hearing the idol-pop rap in the OP song again; funny how they skipped that in the preview. They establish the two main girls and then jump into a flashback to explain how Mighty Maid got her dream job, saving her original Mary Sue life for the end.

Verdict: imagine Mahoro’s arrival in Mahoromatic, but less cute and far more powerful. And make it shouty.

(Nagi’s maid magic is best)

Fair warning…

The new new new new hotness in offline GenAI image-making is Krea 2, which produces pretty damn impressive images, and unlike Klein-9B, the one I was using most recently, you don’t lose 10%+ of your images to anatomy fails. One particularly interesting feature is that image generation is almost completely uncensored, but the text processing layer attempts to replace naughty words, and this turns out to be spectacularly easy to get around, making this perhaps the first zero-day-hardcore-hack model. Just be sure to specify that all participants are clearly of legal age, because it apparently will go there, too.

So, yeah, there will be more GenAI girls soon. Currently it’s working on a batch that converts Japanese cheesecake photos into oil paintings. Some work better than others.

Thanks, Apple

I needed to download something from Apple’s app store on my work Mac. Which meant I had to create a new Apple ID. Which meant I had to give it a unique phone number for verification; you can’t skip this, and the number can’t be tied to another Apple ID. It must also have a billing address, but the credit card is optional. Once you’ve set everything up, it also wants to enable cloud sync of all your data, so it had to be a new account; I will not mix work and personal data.

I needed to have a phone number that either accepted text messages or incoming calls, and it silently refused to call my Ooma VoIP number. So I had to charge and turn on the old iPhone that I’ve already wiped twice, that still has my old California phone number on it.

Which used to be associated with my personal Apple ID, but fortunately I changed that when I moved back to America. One text message later, and I could power the phone down again.

Why, yes, this is bullshit.

A good start

I’m delighted to see that the antifa terrorists who tried to murder police officers and federal agents as a “protest” against the concept of borders have been convicted and sentenced to decades in prison.

May there be many more (convictions, not attempted murders).

First Summer show for me will be Isekai Mighty Maid on Wednesday. Next up is Tan Teen Oni Waifu on July 3rd, Skeleton Knight 2 and Magilumiere 2 on the 4th, Isekai Ass-Guardian on the 5th, then Tanya 2 and Bumpkin 2 on the 8th.

Farm Harem Never 2, fin

Yeah, they pretty much ran out of material and padded the hell out of the ending. That’s what they get for deleting all the hareming. At least we got a small dose of tan-elf engineer gals to help finish it off.

Verdict: gutting the source made this a lot less fun than it could have been.

Witch Hat Atelier, fin

“Okay, do we rush the awesome story and risk ruining it, or do we cliffhanger the hell out of this final episode?”

Verdict: they cliffhangered the hell out of it. And they don’t seem to have announced a second season. Perhaps they’re waiting for the pitchfork-wielding crowds to assemble.

NSA reveals how bad their security is

Anthropic’s Mythos LLM went through it like a Swiss Army Knife through margarine.

Revised image-recognition prompt

So, after running ~80,000 photos through an image-recognition model, I’ve been working on consolidating them into a proper dataset. Unfortunately, the LLM output had so many inconsistencies that my processing script was turning into a mess of exceptions.

So I handed the system prompt over to Claude for cleanup and now I’m re-running the analysis. This will take a while, but the good news is that updating LM Studio on the M4 Pro Mini and re-downloading an MLX version of the model has sped up the processing and reduced the frequency of lockups.

New system prompt (still using qwen3-vl-4b):

It’s not perfect…

I asked it to describe this picture, and it never came back:

For a while I thought it had gone nuts looking for the 1girl, but eventually I got Claude to invent an explanation that there was an underlying problem with the specific model getting fixated on colors and going into a “degenerative repetition loop”. It even supplied helpful links that referred to a completely different LLM.

The suggested fix, setting a hard max-tokens limit and a repetition penalty of 1.1, seems to be working, but it might be the right answer for the wrong reason.

One quirk is that the format is so strict that all it does to identify multiple subjects is separate the lists with a blank line. No headers for “subject (left)”, “top panel”, etc.

The Anti-Altman

Wondermark explains it all:

“seriously, find another one”

Google Maps search:

J: “Chase Bank near Miamisburg”
G: (zooms in on location that is closed for remodeling)
J: “Chase Bank near Miamisburg”
G: (again)
J: “Chase Bank near Miamisburg that’s not closed”
G: (again)
J: “Chase Bank near Miamisburg seriously, find another one”
G: (returns all locations in ~10-mile radius)

New promo videos for Magilumiere 2 and Isekai Ass-Guardian. The first shows off the new underrim-glasses-wearing magical engineer gal, and the second shows off the platonic-as-far-as-he-knows harem.

Apple Watch accessibility is infuriating

The crown button on my watch stopped working, and the FAQs suggest that the most likely cause was poor tolerances dirt in the gap, which could be fixed by holding it under running water and slowly rotating and pressing the crown button.

This worked. However, it also somehow turned on audio announcements of each on-screen action, which prevented any actual, y’know, actions. I could type in my passcode, and it would cheerfully recite the digits, but would not pass them through to the watch to unlock it. And this behavior survived a reboot.

The only way to regain control of the device was from the “accessibility” menu on the phone app. There was literally nothing I could do from the watch itself to disable what I had accidentally enabled; it was completely useless.

Today I was consumed with the desire to bitch-slap an emergency-room doctor into the next county. Fortunately, he wasn’t in the room at the time.

Farm Harem Never 2, episode 11

Seems like an odd time to be introducing an entirely new race of subjects, but they did include a crew of scantily-clad busty succubi, so I can’t complain. And then we got a standalone piece with all the subtlety and complexity of a four-panel gag comic.

Verdict: would have been significantly improved by having the succubi use their powers to get the haremettes a-hareming. One more week to go.

Witch Hat Atelier, episode 12

This week, Our Problem Children help Our New Problem Child deal with his crippling insecurity and history of abuse, and Sexy Witch Gal demonstrates that she is both A) an excellent teacher, and B) a damsel in distress.

Verdict: as part of next week’s season-ender, I demand the announcement of season 2. With a firm date attached.

(this week we got to see a new Bad Witch in action)