(we now resume our regularly scheduled randomblogging, until the next slaughter of innocents by Left-wing domestic terrorists)

Boxxo Or Bust 2, episode 11

This is really not a good week for the show to go grimdark, revealing Our Shoe Fetishist’s bloody, revenge-fueled path. I’m just not going to watch it right now.

Here’s something wholesome instead:

Warner Bros wants to be the Broadcom of streaming

CEO insists HBO Max should be priced like diamonds. Six customers are so much easier to support than six million.

Not a fan

I bought a new unmanaged 10Gb switch, because my existing one only had 2 10Gb ports with the rest being 2.5Gb. Three years ago, the price of a switch with 8 10Gb RJ45 ports was a bit too spicy, but now it’s just annoying.

What really hurts is when a desktop switch has a fan loud enough to be heard from the next room. And I need it to be in the same room. For comparison, I barely notice the Synology NAS that’s on the same desk, even when it’s doing a RAID scrub. For the sake of my sanity and productivity, I guess I’ll be building the switch a padded crib…

New Monitor(s)

The new monitor is an “ASUS ZenScreen Duo OLED 360° Foldable Portable Monitor (MQ149CD)”, which is quite a mouthful, but the TL/DR is that it’s two 1920x1200 HDR panels that can be used separately or together, portrait or landscape. To use it as two separate monitors, you need to use two cables, but despite the one-inch gap between the panels, you can still get good use out of it as a single 1920x2400 display.

I’m using the 1/4”-by-20 tripod screw to mount it higher than the kickstand, and despite all claims of portability, yeah, you have to plug it in to drive it properly (1 of the 3 USB-C ports is power-only). Right now I’ve just got it hooked up with HDMI, because my current dock was low on free USB-C ports, but once I get things rearranged I’ll separate them, because Lightroom’s window layout is a bit inflexible, and the one-inch gap is bad for photo editing. Better to use the application’s “secondary display” support.

This does make me want a vertical 20+ inch 4K HDR portable monitor, though. I just don’t want it to be from a company I’ve never heard of before, which is what’s all over Amazon.

“Shared memory and fancy NPUs are no match for a good graphics card at your side, kid.”

While I’m discussing new toys, I’ll mention that I also picked up a refurbished M4 Pro Mac Mini, which in theory has much faster “AI” performance than my M2 MacBook Air. With 64 GB of RAM, it can run mid-sized offline LLMs at a decent pace, and even do quick image captioning if I downscale raw images to a sensible resolution.

What it can’t do is more than very basic image generation. It’s not the memory, it’s the cores: not enough of them, and not fast enough. This matches what I’ve heard about the insanely-pricy Mac Studio and the merely-way-overpriced systems built on the AMD Ryzen AI Max+ 395. Their performance comparisons are based entirely on running models that simply don’t fit into a consumer graphics card’s VRAM. “Our toaster is much faster at running this 48 GB model than a 24 GB Nvidia 4090!”

The Mini is a huge speedup for Lightroom and Photoshop, though, which made it worthwhile for me. Any “AI” playability is a bonus.

(I have never seen a human being hold a camera this way)

Related, “Dear Draw Things programmers”

Come back when you discover the concept of legibility. Dipshits.

Accidental Disney Princess

I was testing a few more LLM-enhanced dynamic prompts, and one in particular stood out: a series of very consistent images of a fresh-faced young gal who could make Bambi’s woodland friends sit up and beg.

4K resolution, crisp and highly detailed, create an illustration that exudes a high-budget aesthetic. Depict an average height, lovely ethnic Egyptian woman who is 18 years old with a petite figure. She has deep blue eyes, lobed ears, a straight nose, a wide chin, an U-shaped jaw, dimpled cheeks, and a flat forehead. Her heart-shaped face highlights her gentle features. Her skin is healthy and alabaster white, with festive holiday makeup that complements her almond-shaped eyes and full lips. Her hair is steel gray, styled in a cybernetic pixie cut with metallic edges and glowing circuit-like patterns. The woman has a happy expression as she stands with one arm raised, bathed in luminous waves of light. The background features a lush jungle filled with singing flora, glittering with raindrops under the radiant colors of a rainbow sun. Subtle dawn light filters through, creating an aerial perspective with open framing and pastel tones that evoke a gentle awakening. The composition highlights her graceful form against the vibrant greenery, capturing a dreamlike atmosphere.

Sadly, the prompt that worked so well in Qwen Image was… less successful with other models that don’t have an embedded LLM to parse complete sentences sensibly, and failed completely with a few that couldn’t handle the combination of the Beta scheduler and the Heun++ 2 sampler (I may redo those with their defaults).

It’s not that the others all failed, they just didn’t produce consistent characters who looked like my princess and/or lived in her enchanted forest.

more...

Everyone on the Left, raise your hand if you’re not a murderous savage. Now look around your peer group and count the lack of hands, then come over and rejoin civilization. There might still be time.

Those who didn’t raise your hands? Do not expect civilization to protect you any more.

(no cheeky anime picture, out of respect for the victims)

(…that’s not another season of Frieren or Apothecary Diaries)

The World’s Strongest Rearguard anime

I consider this to be one of the best cheat-power isekai light novel series. Our Ass-Guardian Hero reincarnates into a fantasy world that basically forces all the newcomers into dungeon crawls without any real choice or training, but while he’s being railroaded by a hot chick with big boobs, he accidentally chooses a rare class. Everyone else is trying to be things like “valkyrie”, “soldier”, “forager”, “tennis coach”, etc, but hearing that parties are always looking for someone to take the rearguard position, he just writes down “rearguard”.

And he gains the ability to buff any party member who is in front of him. It should come as no surprise that he ends up with a party of gorgeous gals, but it might be surprising that he is a complete gentleman who doesn’t so much as stare at the asses he’s guarding. He’s not oblivious to their (mostly-underage) charms, but he never, ever goes there.

What makes it work is that despite him being potentially and eventually overpowered, he and his party have to work for every fight, and their trust and teamwork keep them alive just as much as their skills. His power gives them an edge, but his decency and kindness have just as much of an impact on the world.

Add to that that every not-quite-haremette has a distinct personality and goal, and you have something that I’m begging the animators not to screw up.

The buried lede in this article is that book nine is finally coming out after a four-year wait. Book 8 wrapped up a major story arc, but left a lot of character arcs unresolved, so woo-hoo!

(Our Hero is in the background, with His Formerly Abusive Former Manager on the left, Our Terrifying Teen Swordsgal on the right, and Our Best Girl Rescue Lizardgal in the middle; and the butterflies are seriously bad news, with their presence suggesting two cours)

Just-Outta-Time Logistics

I bought an HDR monitor on Amazon, which was supposed to arrive Monday. At 4pm, they announced that for unknown reasons it never left an Ohio warehouse, and it would instead arrive Tuesday. Three hours later, that message disappeared, and the order status reverted to arriving-Monday-but-not-shipped-yet. It stayed like that until after 10pm, when it went to “delayed, not shipped” again, with no further explanation.

In the wee hours of the morning, Amazon’s entire order-history system apparently went offline, and all status was unavailable.

As of 8:30 AM, it is allegedly across town, waiting to get put on a truck. Unless they erase that status update.

Unrelated,

FedEx grudgingly coughed up an approximate delivery time for Tuesday’s signature-required package (not from Amazon), which is good, since I have to pick up my niece after school, and I’d hate to have to go chase down the package at their nearest office after the truck returned to base.

Amusingly, when I asked for details, they offered two confirmation methods: SMS to my (correct!) cellphone number, or email to a masked address that I didn’t immediately recognize. It would have reached me anyway, since I own the domain, but it took me a moment to figure out that they’d decided that the email address associated with my physical address was the one I’d created specifically for B&H Photo.

Because they shipped me a CompactFlash Express card (fastest storage for my 4k-video-capable Sony A7SIII) via FedEx a while back.

Mistral Small is gaaaaaaaay…

Given the following picture of Yuzuha Hongo:

And the relatively-capable vision-enabled LLM mistral-small-3.2, we get the following caption and tag set that contain a rather serious error:

A young woman with long brown hair, straight hair style, and a slim body type is standing in a bathroom. She has fair skin, small breasts, and is wearing a pink floral bikini top with thin straps. She has a slight smile on her face and is looking directly at the camera. The bathroom features white tiles, a shower head, and a towel rack with towels hanging on it. Water droplets are visible on her skin.

Tags: young woman, long brown hair, straight hair style, slim body type, fair skin, small breasts, pink floral bikini top, thin straps, slight smile, direct gaze, white tiles, shower head, towel rack, hanging towels, water droplets

(ran it a few more times, and got completely different results every time; sometimes it correctly handled her boobs (coughcough), other times it decided her nipples were visible, several times it hallucinated a towel rack with towels, occasionally it remembered to mention “Asian facial features”, etc, etc; honestly, I trust the output so little that I’m tempted to have my script loop over the same picture N times and keep only the tags that show up at least 50% of the time; I’m also going to borrow the styling of the Qwen prompt-enhancer below)

Enhanced dynamic prompting

With my new still-in-progress dynamic prompt collection, I’m turning this:

__style/photo__ of a __var/prettygal__ with __skin/_normal__ and __hair/_normal__, with a __mood/_normal__ expression. __pose/sexy__. Her location is __place/_normal__. __scene__.

into prompts like this:

4k, crisp, high budget, highly detailed, intricate, ultra textured. Create a photograph of a average height, beautiful, ethnic American, adolescent woman with supple figure, Steel gray eyes, softly-curled lobe ears, Angular Nose, V-Shaped Chin, Strong Jaw, Angular Cheeks, Angular Forehead, Triangle face shape, Natural “no-makeup” makeup with healthy Sun-Kissed-Tan skin and Natural Gray hair, Low twist with braids, creating texture and detail, with a warm expression. Standing with weight on one leg, other foot planted flat, arm crossed over chest, wrist aligned, chin up. Her location is Grand Baie, Mauritius. Rainy glow, close framing, layered reflection, muted palette, melancholic ambiance.

Then using LM Studio’s Python API to pass it through qwen2.5-7b-instruct, with this (taken from the demo code in their repo):

And getting enhanced prompts like this:

4K resolution, crisp and highly detailed, with a high budget aesthetic that emphasizes intricate texture. Create a photograph of an average-height, beautiful ethnic American adolescent woman with a supple figure. Her steel gray eyes sparkle softly, and her ears are gently curved. She has an angular nose, a V-shaped chin, a strong jaw, angular cheeks, and an angular forehead, forming a triangle face shape. Her natural “no-makeup” makeup highlights healthy sun-kissed tan skin, while her naturally gray hair is styled in a low twist with braids that add texture and detail. She stands with weight on one leg, the other foot planted firmly, and her arm is crossed over her chest with wrists aligned. Her chin is lifted slightly, giving a warm expression. The setting is Grand Baie, Mauritius. The photo captures a rainy glow, using close framing to create layered reflections in the background, maintaining a muted palette that evokes a melancholic ambiance.

Now, let’s see what sort of picture that produces…

more...

So it turns out that not only does conflict build character, but it works the other way around, too. The Power Of Friendship wins the trial, and in the process we get a pretty-well-incorporated flashback that fills in some more details on Our Mighty Tsuntail’s mom. Pity she wasn’t there for it, but at least she shows up for a minute or two, continuing the offscreen path of “I must get stronger”.

Verdict: if this weren’t so good, I’d really hate the trope abuse.

(“🎶 1. cut a hole in a box 🎶”; unrelated, of course)

Power sucks

My MacBook Air is running out of power while plugged in, despite not even being warm to the touch. The Anker dock, on the other hand, is blistering hot. The CPU is barely ticking over, and Apple’s battery monitor even claims “No Apps Using Significant Energy”.

The culprit, of course, is “AI”. I’ve got LM Studio running a prompt-enhancement prompt with qwen2.5-7b-instruct, taking the output of my dynamic-prompt-generation script and improving it (which it’s actually pretty good at). I had to find the much-higher-wattage original Apple charger, which I’ve never needed to use before. It turns out there’s an “energy impact” field reported in Activity Monitor, but it doesn’t charge GPU/NPU usage to the correct app.

Apple’s so excited about moving fast and breaking things with “AI” that they can’t spare the engineering talent to update their tools to even report how much power LLMs are using on their products.

In fact, when you sort Activity Monitor by “energy impact”, LM Studio is at the top despite having a “score” that’s less than 10% of what Activity Monitor claims for itself. The only field that seems to be honest is “12 hr power”, which puts LM Studio an order of magnitude above the rest. But only because it’s been running for several hours, sigh.

(like an iceberg (or “Apple Intelligence”), 90% of Froppy is underwater…)

More “stimulating” covers…

“We want YOU to join the Star League!”

more...

Our Great (Hot) Detective’s plan is revealed, as Our Wannabe-Sucker Hero confronts her and lays out the evidence. Unfortunately he didn’t plan for the aftermath.

Verdict: it was nice to see Nazuna go Maximum Effort and prove that Anko/Kyoko’s vampire-fighting tactics only really worked when her opponent wasn’t serious, neatly subverting the waif-fu girlboss stereotype that’s so common. She can’t beat a vampire in a straight-up fight, no matter how clever she is.

Text generation in Qwen Image

The majority of my failed attempts to generate text have been with relatively-rare words (“harem”, “Hornblower”, etc), so I was a bit surprised when it failed 90% of the time at the word Babies (as in Sugar), while Baby always worked. Then I found an article explaining that the way they got improved text generation in both Latin scripts and Hanzi/Kanji was to explicitly train it with tons of images of words rendered in specific fonts.

In other words, it will almost never produce correct results for words outside of the training set; it would need to integrate a text-rendering engine the way the big boys do. Given that they already bolted a full LLM into it to parse the prompts, that’s doable, but they’d probably have to open up their license to use something like Freetype or ImageMagick. This may be something where I end up creating automated vocabulary LoRAs with a Python script, or asking a high-end LLM to generate masking templates for the image generator and ImageMagick command-lines to composite the text onto the results.

Anyway, I was thinking about that social-media “boys don’t read” kerfluffle, and decided the thing to do was create some things that might stimulate their sense-a-wunda…

BTW, this is not what I had in mind when I asked for “glowing alien symbols”…

more...

A beauty pageant sounds like good clean fun, doesn’t it? Get all the gals fired up and onto a stage, with Our Vending Hero supplying fashion magazines and makeup to give His Best Girls an edge, and keep things light and fluffy.

Until they end on an ominous note by having Our Shoe Fetishist run off deep into the dungeon in the middle of the night after learning the secret origin of some Very Special Shoes. Really? You had to pull that shit?

Verdict: could we just have a fun episode? Is that too much to ask?

(my Pixiv archives really had nothing tagged with shoes or boots; the few things that turned up were about socks, which not only seems to be the more popular fetish in Japan, but the word for shoes, 靴, is part of the word for socks, 靴下, so they’re kind of inevitable in a search)

Dungeon Chibis Bluray is out!

This doesn’t uncensor anything, since it isn’t that kind of show, but there should be the usual quality improvements over the broadcast release. As a bonus, it won’t just disappear from my library like streamed shows eventually do.

(chibi Zelda is unrelated, but I’ve already used all the good fan-art from this show…)

Loner Harem Bluray soon

October 21, to be precise. This was a whole lot more wholesome than the horny-porny novels. In a good way.

(pony gals are unrelated)

Wishing for useful wish lists

Once upon a time, Amazon wish lists displayed release dates. This was handy for organizing my book purchases, so naturally it was removed. Now I have to open each book in a tab to find out its release date, then drag books into order using their flaky Javascript tiling code.

And then every time I visit the list, I have to remember to manually switch it back from “most recently added” to “custom” sort order. Every fucking time I load the page. Because their first-class H1B-or-AI developers don’t give a fuck.

(“the new South Pole remote development team is so much cheaper than hiring Americans!”)

Fuck Youtube

I’m really getting tired of clicking on a “video” by an established “channel” to find that it’s not only an animated clickbait listicle based on stolen content from a random website, but the “narrator” is just an AI-synthesized voice, with the inhuman affect that comes from not really understanding the sentences it’s converting into audio. And of course these “creators” never seem to get demonetized.

(and by “getting tired of”, I mean “these days I refuse to watch anything but official anime promo videos and Critical Drinker reviews”)

Jun Amaki jiggles your martini

Call Of The Night 2, episode 9

Ten years ago on Halloween, Our Vengeful Hot Detective’s life fell apart. Now it’s Halloween again, and the game is afoot! Except she’s not really a great runner. Some fan-service by Our Hot Undead Teacher and Our Still-Getting-Bigger Childhood Friend, but we’re rapidly moving into the endgame, so keep your eyes on the plot. Our Appropriately-Dressed Hero seems to be the only one who really cares what Anko/Kyoko’s goal is, but will he figure it out before it’s too late?

Verdict: so we’ve got a pattern now: more conventional storytelling = better animation.

(sadly, nobody dresses up like this for Halloween…)

Kaiju No. 8 2, episode 7

Conflict Builds Character. In this case, it finally establishes that Faux-Hawk Fang-Boy actually has a character, as he realizes that even a prima donna needs backup.

Verdict: someday there will be a shonen anime that does not use the “I must get stronger” trope. This isn’t going to be it. Still well-done, though.

(Magia Sulfer says, “suit up and show up!”)

Wait, “isekai potion loli” is a genre now?!?

Looking at next season, and not finding much of anything to watch. Folks who keep up with My Hero Academia are likely interested in the final season, and folks into Spy x Family have something to look forward to, but that doesn’t help me, unless I were to spend half the season catching up on them first.

Gate manga 22 & 23 out

You know what that means, right? A picture of Rory Mercury:

Dear Microsoft,

You won’t let me turn off the on-by-default sharing of data with third-party advertisers unless I first turn off my ad-blocker? Fucktards.

(note that this isn’t a browser setting, it’s system-wide, and it includes data hoovered up by Copilot, which I was in the process of turning off at the time…)

Why Boys Aren’t Reading

There’s been a bit of a fuss on social media where leftsplainers handwave away pretty much any negative thing about boys, and recently they’re claiming that boys aren’t reading because there’s just nothing being written for boys. At which point dozens of successful authors pile on and point out that publishing has been run by far-left women and female impersonators for so long that most of them had to go indie to be published at all, and that boys exposed to decent books will gobble them like candy.

I was thinking about this in the grocery store last night, because as I was hiking from the sporting goods department to the egg vault (Meijer stores are quite large), I passed a prominent, colorful display that boldly advertised “Reading For Children”.

Every book was for girls. Only.

“Change my mind” 1

The world was a better place when it was possible to buy ammo at the grocery store at 2 AM.

(thanks to looooooong Covid, now I can’t buy much of anything at 2 AM, except at gas-station convenience stores)

“Change my mind” 2

LLMs that show their “reasoning” are simply writing fan-fic in a parallel thread that has nothing to do with the process used to produce the actual output.

(last night’s “generate 200 unique sexy pin-up poses” test produced a list of 34 unique poses repeated an average of ~6 times, despite the extensive “reasoning” process it fabricated; and they weren’t even particularly interesting poses)

Prompts and Vision

Two things I’d never really played with are using LLMs to “improve” prompts, and using specialized LLMs that can analyze a picture and describe the contents.

Prompt enhancement is… hit or miss, but it’s good at turning the usual list-of-keywords prompting into real sentences and paragraphs, which makes a difference with Qwen Image’s LLM parser. I’ll do some before/after pics of that at some point.

Pity that they’re just as censored when “enhancing” as they are when handling other requests. openai-gpt-oss is just as big a nag as their online models, and the attempts people have made to “uncensor” it haven’t been fully successful yet; I’m not asking it to describe a gang-bang, but it still goes full wokescold when you ask for something simple like a list of “Fifties sexy pin-up poses”. Because it’s not respectful, according to some purple-haired pierced gender-ambiguous 23-year-old in San Francisco.

On to vision! For today (since this is a four-day weekend for me…), let’s try something simple: given a real photograph, ask two LLMs to “describe the contents of the image in detail”, then feed the resulting prompt to Qwen Image.

more...

Small praise to the writers for having Our Vending Hero actually spend some time using new hot-food vending machine styles to please his chilly customers, and having Our DFC Genius Gal suddenly get nervous about undressing in front of him while Our Mighty Girlfriend waves away her concerns because it’s okay if it’s him.

The rest of the episode is a lengthy and pointless quest that accidentally succeeds because a new disposable annoying villain shows up and solves their problem with a whole lot of collateral damage. There’s nothing wrong with Our OP New Legal Loli, but they dropped a plot coupon when she appeared that added to the pile created by Our Not-Really-Evil Betrayers, and they’re going to have to cash them all in with breathless exposition soon, and I simply don’t care about that.

Verdict: lame with brief reminders of what made the first season good.

Qwen Image caption competition

Qwen Image follows prompts quite well, but as with all diffusion models, if a concept isn’t clearly tagged in the training data, then the model can’t easily produce it, and you have to assemble an incantation that resembles the desired concept.

Qwen Image does not grok spanking, and I simply could not get it to swing a hand, paddle, or bat at a buttock. But it wasn’t a total loss, since it turned out to have quite decent understanding of clothing and hair styles from the 1950s…

So, what captions spring to mind for this image? I’ll save my ideas for later. (click the image to see it at full size)

NSFW

Qwen Image limitations

Around 90% of the “beautiful young women” I ask it to create are Statistically-Averaged White American Girls Who Could All Be Sisters. You have to push it away from the mean by adding descriptive words, which means I’ll be dusting off my wildcards to vary skin, hair, and eye color, ethnicity, figure, etc. Certain ethnic keywords produce significant changes, but the training data seems to be limited to only a handful of such tags.

There are a number of canned wildcard sets, and perhaps the most comprehensive are the “Billions of…” sets from DonMischo. As-is, they’re a bit too random, but the YAML has a decent structure, so I should be able to subset it and construct my own canned recipes for babe-making. Since SwarmUI doesn’t directly support YAML wildcard sets, I’ll use my wrapper script to generate large text files of prompts to use as standard wildcards.

That’s how I generated these, after filtering the file to remove the grannies and great-grannies:

more...