It’s up into the nineties outside this week, something I only experience during the brief walk from my air-conditioned car to air-conditioned stores, before returning to my air-conditioned home.

Speaking of hot things, the image-analysis processing is over 103,000 images now, just finishing up a big batch of Hello!Project photobooks (the primary virtue of which is that their usual stylists were locked out of the room). After that is ~7,000 magazine pics to extract from ZIP and RAR archives, and then at some point I’ll do the ~35,000 pornsite pics.

Unrelated, my Macbook Air lost track of its storage again, requiring a hard power-cycle. Seems to happen when it’s been up about three weeks and I do a lot of SMB traffic to the NAS (like extracting all those RAR and ZIP archives…).

Number 1 in logistics!

Amazon: we think UPS lost your package, so we’re going to claim “delayed in transit” and make you wait another week before you can apply for a refund.

UPS: um, the truck arrived in Ohio 18 hours ago, at a depot about a two-hour drive from your house, and it’ll be delivered by noon.

Every Tuesday is Patch Tuesday

Have you noticed how frequently Microsoft is pushing Office updates recently? Like last month where they had to quickly fix basic functionality in Outlook, and then another today?

Jiraporn

Better than Bugzillaporn!

(the photo shoots are terrible, and the site is NSFW and ad-infested, but I’d still file her tickets)

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)

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.

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.

GenAI’s come a long way since it recommended using glue on your pizza toppings. You still can’t trust it not to write instructions that will kill us all and destroy the world, but you can at least ask nicely.

I happened to have on hand a bunch of Ghirardelli chocolate melting wafers, which I use in colder months to produce delicious hot chocolate in my Nespresso Barista machine (still one of the only milk/coffee/etc frothers to be able to handle chunks of chocolate and ice).

Unfortunately, I buy it in 5-pound bags, and it’s been “a touch warm” recently, and I don’t think I’ll be drinking hot chocolate for a while. So I fired up Claude and asked it for instructions to make a chocolate sauce topping suitable for ice cream and tiramisu, using melting wafers and espresso. I made a half-batch that turned out quite nice, and it works well in a squeeze bottle right out of the fridge:

(I had to clean up the instructions, because they were over-literal and repetitive, and I reordered the ingredient list to match the order they’re used; highlighting the ingredients in the instructions is a personal preference that I could tell Claude to do, if I were to build a recipe layout “skill” at some point)

A flaw in the Pexar

Out of the box, it turns off the display at 11 PM and stays off until 7 AM the next day. So, what happens if you’re loading a big batch of images from a USB stick when the clock strikes?

You wait until morning to finish the import. Seriously. It’s completely unresponsive until then.

(I’d say those aren’t Best Lizard Girl Theresia’s snacks, except they look delicious…)

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 you’re 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)