X: The Algorithm

It’s been pretty bad in the past, filling my “for you” feed with assorted scams, engagement farmers, and hate-filled Leftist activists, but at the moment, it seems to be pretty well centered around the type of things posted by the 24 accounts I follow:

• 6 Japanese models/actresses
• 4 artists (2 Japanese)
• 4 novelists (1 who fled to Bluesky but never deleted his account)
• 2 contrarians
• 2 friends
• 2 anime series
• 2 waifu-posters (long idle)
• 1 cat-poster
• 1 app

Basically I get anime pics, RPG pics, cat pics, random snark and memes, and Japanese women in bikinis. This I can scroll for a while.

Truth in advertising, unfortunately

I was throwing a bunch of leftovers into a big pot of slumgullion, and decided to add a can of chili. The label read Stout Beef Chili, and despite it being waaaaay down the list of ingredients, the smell of beer was overwhelming. It cooked out somewhat, but it’s still pretty strong when you open the container. Enough that I can’t imagine using the other can that I bought, and will likely pitch it.

This is the quality of ‘thinking’ that AI provides…

(actual country of origin for “Ackers Science” and “Ackers BORO3.3” products? China, of course)

MasterCook, Cooked

After months of relentlessly pushing their new GenAI-enhanced version and their cloud subscription service, the current owner of the venerable MasterCook recipe-management software has sold the rights to some entity called Cook’n, which appears to be junking the software and only bought the customer list. They’re honoring subscriptions, but charging $10 to migrate you to their cloud.

Is their product comparable? No idea; I’ve seen so many products advertise MasterCook compatibility without actually implementing the full feature set that I gave up years ago.

Fulfillment Roulette

I ordered three things from Amazon. One is made of squishably soft silicone, the other two are made from borosilicate glass. Each one has its own tracking number, suggesting they did not combine them into one box, despite me checking the “take your time” button on shipping.

Now, which one will have the sturdiest package when it arrives tonight?

(I didn’t ask for Ricotta to serve dubious concoctions in champagne flutes, but I guess I didn’t not ask for it, either; this is what happens when you fall back to a model with less-capable parsing because it has the anime LoRAs you need for specific characters)

🎶 “One of these things is not like the other…” 🎶

The Oregon Highlander

I saw the name of this illegal alien convicted sex offender who was advising Oregon on healthcare, and my first thought was not “yeah, that tracks”, but rather:

I guess being chief metallurgist to King Charles V of Spain doesn’t pay what it used to.

(“there can be only one!”)

Somebody didn’t get the memo…

😁

Experimental girl-swapping

The Flux models have plasticky skin, fewer trained art styles, and better-than-SDXL-but-not-by-much prompting. Qwen Image has excellent prompting and posing, but a strong tendency to converge on a handful of styles, locations, and faces. One of the more recent Flux models is Krea, which is supposed to be heavily trained on photographic and art styles. The full version is also 22 GB, so I wasn’t sure how well it would perform on my 24 GB RTX 4090 at all.

It was surprisingly quick, and it did style the images more than Qwen or standard Flux, but it definitely didn’t have the kind of LLM-based prompting that makes Qwen stand out.

So I crossed my fingers and set things up so that Qwen generated a 36-step 576x1024 image and then handed it off to Krea for 24 steps of refining and upscaling to 4K. Performance was quite good, but the results were… rough. One gal had a second face growing out of her ankle, another had an eye for a nipple, another had blue feet and something hideous growing out of her mutated hand, etc, etc.

TL/DR: I have yet to find a refine-only model that works well with Qwen as its base; the ones that don’t produce awful images produce low-resolution ones. So that idea was a bust, and not even a bouncy one.

(I need models that are… rock solid)

Finally!

Windows 11 no longer reboots when you tell it to update-and-shutdown. The problem affected Windows 10 as well, so it’s been busted for ten years, neatly highlighting Microsoft’s QA priorities. Which somehow seem to involve cramming more ads, AI, and privacy violations into every product.

Oh, joy, more work for me!

My parents have a Windows 10 PC that they were worried about not being able to upgrade as the OS falls out of support. Last week, I got a call asking me to come over and fix its sudden inability to print, which involved deleting and reinstalling the printer driver. Odd, but hey, it’s Windows.

Except that there were more problems, like the fact that left-clicking the Start button did nothing, the taskbar Search field did nothing, and right-clicking the start button to launch Settings popped up an error dialog. I didn’t have a lot of time, so I screenshotted everything and went off to research fixes, which included some partially successful incantations.

Then I discovered what was really going on: it had silently upgraded itself to Windows 11. Mostly. So now I need to copy a recovery image to a USB stick and head back over there soon to repair or re-run the upgrade. Sigh.

(I’ve got half a dozen Windows PCs around the house that I would be happy to upgrade to Win11, but I can’t; several are old enough that the CPUs simply aren’t supported, so even the workarounds won’t work around, and one couldn’t run Linux or BSD for blood or money, due to proprietary drivers required for major components)

Python drag-and-drop wrapper for py2app

Turns out that Automator isn’t entirely scriptable, which seems like an obvious oversight, but Apple probably couldn’t figure out how to monetize those pixels. Instead, I updated the build script for my gallery-wall app (not yet uploaded to Github) to use Platypus, not to be confused with Python Platypus.

cat > _rungallery.sh <<'EOF'
#!/bin/zsh
exec open -a "gallerywall_backend" --args "$@"
EOF
chmod +x _rungallery.sh

platypus -a "Gallery Wall" \
    --interface-type None \
    --droppable \
    --quit-after-execution \
    --bundle-identifier "org.dotclue.gallerywall" \
    --author "J Greely" \
    --app-version "1.0.0" \
    --app-icon images/GalleryWall-wrapper.icns \
    --interpreter "/bin/zsh" \
    _rungallery.sh

(the downside is that it makes apps that won’t run on a nailed-down-by-default Mac, and in fact you can’t even install Platypus on most Macs, due to code-signing errors that are supposed to be fixed in the next release)

The “self-portraits” I’ve posted recently were all done with prompts starting with “slightly-cartoonish illustration” to set the style. I also used this phrasing for the Diablo 4 barbarian illustration below them, which isn’t cartoonish at all.

So why is it that the moment I start to describe them in detail, to add variety, my pinup gals go full-bore big-eye anime style? Either 2D or Frozen-style 3D?

TL/DR: mentioning eyes at all, even just their color, is enough to do it. The expressive LLM-generated mood descriptions I’ve been experimenting with were also contributing (and creating some contradictory pose instructions, which I’d already made a note to fix), but all it takes to turn “slightly-cartoonish illustration of a woman cooking” into pop-eyed anime is adding “with blue eyes”.

The following images were all done with the same settings (Qwen Image, CFG 6.5, 42 steps, seed 1019441477):

A slightly-cartoonish illustration of a woman cooking:

A slightly-cartoonish illustration of a woman with blue eyes cooking:

more...

…and setting his clock back an hour this weekend. I had a helluva time getting Qwen Image to even get near Brian Blessed’s Vultan costume from Flash Gordon to dress my not-entirely-accurate avatar up for Halloween.

Claude started out with a spectacularly bad attempt to describe the costume:

When called on it, it “researched” the correct costume and at least got into the general ballpark, but still without getting a single component correct. I couldn’t find a decent high-resolution still to feed in as input, so I just hacked at the prompt until it looked like it was kitbashed from Thor cosplay leftovers.

“Trunk or treat” is an abomination

(You can’t change my mind)

Number of kids who came to my door? Fifteen.

Bags of candy left over? Less than fifteen. I figure my niece’s high school is going to need some donations.

Oddest thing was that a third of the kids showed up without bags for candy (the pros had pillowcases, which I respected). They were in costume, but they expected to receive one or two pieces and hand them off to an adult waiting at the curb. This doesn’t work at my house, where a double handful with my hands isn’t going to fit in theirs. Fortunately I’d been to the grocery and had half a dozen plastic bags to give away.

(“this will look terrific on Arato-senpai!”)

Gals on the right

I decided to take my retro-sf wildcards and use them to generate wide-format wallpaper for my gaming PC, which has for several years been using a photo of a penguin appearing to operate a DSLR camera (I think it came from a Bing wallpaper rotation).

It wasn’t obvious when I was generating tall images, but Qwen Image has a strong bias toward putting the subject dead center. You can tell it to put her on the right side of the image, but explicit instructions to “place the main subject on the left side of the image” are almost always ignored, if not reversed. Compositionally speaking, this is kind of frustrating. It’s possible, just quite difficult to arrange in the prompt.

Of course, this is the same model that thinks freckles are the size and color of pennies, and “faint scars” should be rendered as deep gaping wounds. Seriously, what was Alibaba using for training data, medical-school cadavers?

Competition!

For the past few years, providers have been promising to have high-speed fiber Internet service in my area. Cincinnati-based AltaFiber seemed to be expanding rapidly, then went quiet, but for the past few weeks there’s been major digging going on along the nearest main street near my house, and yesterday I spotted little flags and paint markers in the utility easement at the edge of my back yard, and sure enough, Friday afternoon some big equipment arrived and spent the afternoon pulling cable from one end of the street to the other, accompanied by little door cards announcing the imminent arrival of AT&T fiber.

Since it’s not available yet, they won’t give me the details of the package, but that’s okay, I don’t actually want it. What I want, and had wanted from Alta, was leverage to use in a call to my current provider. They offer new customers more speed for less money than the package I’m paying for.

Done with Ruri’s adventures, I dug into my pre-Covid, pre-GenAI cheesecake archives, and deathmatched the gals I downloaded in May, 2019.

There is a lot of “Ai” in this set, but it’s autocompleted with “Shinozaki”, which is healthy and natural and good for growing boys.

more...

Best. Lecture. Ever. This episode is clearly the source of the “PG” promise of nudity, despite steam and chibification keeping things squeaky clean. And I’m confident the steam was not a buy-the-Bluray tease; it’s not that kind of show.

Anyway, along with limestone deposits, we get a discussion of future plans, and to the surprise of absolutely nobody, Nagi wants to teach, Shoko wants to become Yoko, and Ruri stumbles on her answer with a little help from her friends. The series closes with a montage of the near future, and a glimpse at Ruri’s suspiciously-familiar adult form.

Corn, popped

Literally for once. I hadn’t planned on making caramel corn Sunday night, but out of nowhere, my mother asked me how much unpopped corn you need to make 8 quarts of popcorn. She’d found a recipe for caramel corn somewhere, and it assumed you knew the conversion ratio. I ended up at popcorn.org, which says 2 TBSP of kernels for a quart of popped corn.

They also had a smaller, allegedly easier recipe that only required 5 quarts of the white stuff.

It somehow ended up being my job to run the air popper and follow the recipe, and the only difficulty was that the premium popcorn they’d bought produced significantly more volume than expected. 8 TBSP of kernels would have been more than enough.

I suspect I’ll also be drafted into making the next batch, this time with nuts.

Reminder: I am not actually this fat or this bald, I have much better trigger discipline, and I’m left-handed. Qwen is, shall we say, “not good” at guns, and definitely has a problem with the concept of holsters:

(I didn’t even ask to be holding the gun, just “holstered at his left hip”)

Diablo 4, season 10

I generally start new seasons with a Necromancer minion build, because being escorted by a pack of skellies does a good job of keeping you from getting overwhelmed at low levels as you acquire not-entirely-crappy gear and start to build up cash, materials, and abilities. Once you get some rare drops, it becomes an easy way to reach Torment 4 (Hand of Naz unique gloves, Nagu and Ceh runes, and Aspect of Occult Dominion cast on a helmet give you 14 skeletal mages and six spirit wolves to do your killing).

I’m currently farming in T3 because it’s faster, trying to get an Ophidian Iris for my incinerate/hydra sorcerer. I also thought I’d need some uniques to build up a whirlwind barbarian, but they’re letting you powerlevel alts through the seasonal content again, so I unlocked Deafening Chorus at level 30, and I already had an item enchanted with Aspect of Fierce Winds (DC = shouts are always active at +50%, AoFW = activating a shout creates 3 dust devils). The Neo and Ceh runes are another way of creating a pack of spirit wolves, so damage just kinda happens while you run around. And you’re also berserk and unstoppable at all times, with increased damage reduction and speed, making farming less of a chore.

(this is approximately 10% as chaotic as actually playing this build)

Reminder: X hallucinates your “interests”

If your feed seems skewed, it’s time to go in and uncheck the auto-generated horseshit “interests”. This week, mine was:

Not only have I never engaged with any tweet on most of these subjects, most aren’t even things that someone I follow would make fun of. And for the few that were memed by someone, that context should be taken into account. Pointing and laughing at something doesn’t mean you want to see more of it.

The only nice thing I can say about this list is that it’s just garbage, not explicitly hard-Left garbage.

Random randoms of randomness

Someone on the SwarmUI Discord posted a complex prompt that uses the app’s native randomizing syntax to create a wide variety of people portraits with diverse (both meanings) faces. It’s like a Perl one-liner had sex with a MadLibs book:

(analog photography.:2) in <random:an indoor|an outdoor|a studio> setting.
<random:<setvar[gender,false]:man><setvar[pronoun_n,false]:he><setvar[pronoun_p,false]:his>|<setvar[gender,false]:woman><setvar[pronoun_n,false]:she><setvar[pronoun_p,false]:her>|<setvar[gender,false]:person><setvar[pronoun_n,false]:they><setvar[pronoun_p,false]:their>>the <var[gender]> is in <var[pronoun_p]> <random:early|late|> (<random:teen years|twenties|thirties|forties|fifties|sixties|seventies|eighties>:2)..
<var[pronoun_n]> has a distinctive <random:oval|round|square|heart-shaped|long|oblong|diamond|triangular> face, looking <random:pleasant with a gentle smile|energetic with a broad smile and a cheerful grin|radiant, beaming with joy and sparkled eyes|amused, with a slight smirk and a twinkle in the eye|contended, in a peaceful and serene expression|playful, with a mischievous glint, hinting at fun|warm and welcoming|hopeful and optimistic|blissful, lost in happy thoughts|thoughful, deeply pensive|curious, eyes wide with interest and head tilted on a side|observant and scrutiny|introspective, lost in reflection|calm and peaceful|serene, untroubled|focused, intently concentrating|weary and exhausted|melancholic and wistful|pensive with a hint of sadness|resigned in acceptance|wistful, longing for the past|anxious with worry in the eyes|reserved, keeping it formal|overjoyed in exhuberant happiness|very sad in a genuine sorrow with tears|angry, bursting in rage|surprised, startled with wide-eyed wonder|fearful, apprehensive and scared|determined with strong will and resolve|intense, experiencing a deep feeling|smiling sadly|thoughtfully amused|weary but determined|curiously spektikal and doubtful|serenely hopeful>.
<var[pronoun_n]> has <random:porcelain|ivory|rosy-fair|pale golden|peach|golden|olive|light tan|tan|bronze|caramel|warm brown|cool brown|chocolat|deep bronze|ebony|rich dark brown|honeyed|copper|russet> <random:flawless|velvety|polished|dimpled|porous|weathered|sun-kissed|bumpy|freckled|densely freckled> <random:skin with <random:small,large, prominent> moles|skin  with <random:faint, noticeable, prominent> (scars:<random:1,2,3,4,5>).|skin>, <random:pixie cut|buzzcut|chin-length|jawline-length|shoulder-length|collarbone-length|mid-back length|waist-length|long|classic bob|layered bob> <random:jet black|raven black|soft black|chocalate brown|dark mahogany|chestnut brown|medium brown|light brown|ash brown|auburn|platinum blonde|golden blonde|honey blonde|strawberry blonde|ash blonde|dirty blonde|fiery red|ginger red|burgundy|silver gray|charcoal grey|snow white|salt-and-pepper|bronze|mahogany|russet|ombre|blue> hair with a <random:high hairline|low hairline| widow's peak|straight hairline>; <random:high|medium|low> and <random:wide|narrow|average> <random:sloping|straight|with prominent brow bone> forehead.
<var[pronoun_p]> <random:sparse|dense|regular|asymmetrical> eyebrows are <random:thin|medium|thick|bushy> and <random:arched|straight|angled upward|angled downward|rounded|curved>, <random:matching hair color|darker than hair|lighter than hair>, <random:almond-shaped|round|upturned|downturned|hooded|monolid|> <random:large|medium|small> eyes show a beautiful <random:sky blue|deep sapphire blue|grey-blue|turquoise|emerald green|olive green|hazel green|dark chocolate brown|light hazel brown|golden brown|mahogany brown|mixture of brown, green, and gold|silver grey|slate grey> hue<random:|, deep-set into <var[pronoun_p]> face|, protruding forward>, <random:short|medium|long> <random:sparse||thick> <random:straight|naturally curled|heavily curled with mascara> eyelashes.
<var[pronoun_n]> sports a <random:large|medium|small|pointed> <random:straight nose, with a classic and balanced profile|roman nose with a prominent bridge and a hump|greek nose, with a straight bridge and refined tip|snub nose, upturned and delicate|aquiline nose, hooked and curved downwards|button nose, nicely rounded|hawk-like nose, strong and prominent|wide nose, broad at the nostrils|narrow nose with a thin bridge|long nose, extended length from brow to tip>, <var[pronoun_p]> cheeks are <random:rosy, naturally flushed|dimpled with cute indentations|freckled|sun-kissed|sculpted sharp|softly curved>. <var[pronoun_n]> has <random:large|medium|small> <random:full, voluminous|thin, delicate|wide, spreading across <var[pronoun_p]> face|distinctive bow-shaped|softly curved round|down-turned drooping|> <random:lips.|lips with a defined cupid's bow.|lips with a prominent lower lips.|lips with a prominent upper lips.> <var[pronoun_p]> chin is <random:large|medium|small|strong|weak> and <random:rounded|square|pointed|cleft|receding|prominent|doubled, with fullness under the chin>, <var[pronoun_n]> has a <random:strong|soft|square|rounded|well defined|soft> and <random:wide|narrow> jawline, completing <var[pronoun_p]> face.
<var[pronoun_p]> build is <random:slender|lean|muscular|stocky|robust|petite|large-framed>, <var[pronoun_n]> stands in an <random:upright|slouching|stooped|relaxed|casual|tense|stiff> posture, showcasing <var[pronoun_p]> <random:broad|wide|narrow|sloping|rounded|square> shoulders and <random:full|flat|narrow> chest. <var[pronoun_p]> arms are <random:long|short|muscular|toned|slender|lean|strong|robust>, completed with <random:delicate|rough|small|large> hands.
<var[pronoun_n]> wears a <random:formal|casual|loose|fitted|tight> <random:cotton|silk|wool|denim|leather> <random:shirt|blouse|t-shirt|sweater|jacket> contrasting with <var[pronoun_p]> hair.

…and that’s why I converted it to the dynamicprompts YAML format, which revealed all sorts of typos and awkward phrasing. I added weighting to improve realism (completely random choices produce a lot of unusable crap), cleaned up the option lists, and added a new one to push Qwen Image away from its default faces.

My version is here, in my new Github repo for random genai-related stuff. The Python script I use as a wrapper for the dynamicprompts library is in there, too.

You know it’s time for a new theme when…

…you weed 500 images down to 162 and say, “forget the toes, maybe I’ll just use them all as wallpaper”. I’m not going to blog that many, though, so I deathmatched it down to 18 based on the newly revised-and-randomized sfnal locations.

I started with ~900 retro-sf locations generated by several different online and offline LLMs, fed them all to Claude to categorize and create sentence patterns, and then threw 500 imaginary pretty girls on top. Many of the results were visually incoherent, and quite a few had nothing about them that even hinted at SF, retro or otherwise, but most met the goal of being lively and colorful without drawing attention away from the girl.

No refine&upscale for this batch; it’ll run overnight for the full set of 162, so I probably won’t kick it off today.

more...