I got a small-but-pleasant surprise when I did my taxes. The new exterior doors I bought last year qualified for an energy-saving tax reduction. Not huge, but worth the hassle of filling out the form.

Dear Amazon,

Can you please stop pretending that “top picks for you” is not just an excuse to shove paid promotions in our faces? It’s bad enough that you constantly fill it with seasonal promotions of things you know I won’t buy, but putting Lena Dunham’s new book at the top of the list was just mean.

After God’s Favored Appraiser comes Oblivious Saint Maid

The author of Appraiser must be feeling pretty good, with two adaptations back-to-back. The summer show will be Heroine? Saint? No, I’m an All-Works Maid (And Proud of It)!, in which a Japanese girl who was super-duper smart and accomplished in her first life finds herself reborn into a game world as the heroine, with the possibly-novel twist that she doesn’t know anything about the game or her role in it, so she indulges her long-held desire to become Super-Maid, totally screwing up the plot. Which was already going off the rails thanks to two other isekai’d teens who did play the game. More arrivals keep remembering their Earth pasts and trying to either take over as the main character or just change their fates, while Our Insanely OP Clueless Heroine just focuses on maidly perfection.

This season’s Foodie Maid has been doing a promotion with the Victorian-maid café I mentioned recently, who’s been getting a lot of visibility thanks to auto-translation on xTwitter. I wouldn’t be surprised if they hooked up with this show as well.

(fan-artists appreciate how Nagi rocks a maid costume)

Fuel for the fire

Or more precisely, lack of fuel. My co-workers in Belfast worked from home on Tuesday, to avoid the traffic-blocking protests. I liked the farmer who rode his bicycle to protest, because he couldn’t afford the fuel to bring his tractor.

(this LoRA is basically limited to drawing Sexy Grownup Misty, not that there’s anything wrong with that…)

More fun with image-gen previews

One of the lesser-used features of Flux.2 and its derivatives is that it was allegedly trained on structured JSON prompts. The examples make it seem like there’s a schema you should follow, but it turns out that’s not so. I took one of my recent ~500-word paragraph prompts and told the offline LLM Gemma 4 to analyze it and convert it to JSON, without specifying any particular structure.

I fed the resulting prompt to Flux.2-Klein-9b and got a quite surprising result: the animated WEBP preview showed the post and composition settling down much faster than with a text prompt. The pose was stable right away, and by 20 steps it was done adding background elements, and just steadily added details.

The minor downside is that running the 31B version of Gemma 4 on my Mac Mini took 3+ minutes per prompt, which does not scale. I’ll have to look for smaller, faster models that are still smart enough to do the analysis and generate valid JSON. In my experience, the various methods of uncensoring reduce the formatting accuracy, so I might have to tinker with the prompting to avoid frightening the horses.

The major downside is that SwarmUI insists on parsing your prompts to apply its own features, with no way to bypass it. I’d say about 20% of my batch of LLM-generated JSON prompts tripped over this, using strings that triggered an attempt to convert words to a floating-point number.

animation

3D girl seeks new position

Jun Amaki is leaving the modeling business soon. So, what are her plans?

Auto-translation from xTwitter:

Benefits of dating me ↓💕

・Natural I-cup
・Petite with a baby face
・Sweet voice
・Can cook
・Always full of charm
・Surprisingly domestic

Drawbacks ↓

・Suddenly becoming lazy
・Breasts too big and getting stared at by people
・Too much of a spoiled baby
・Bad sleeping habits
・Getting full super quick

NSFW

Not seeing much of a downside here…

Farm Harem Maybe 2, episode 2

This week, Our Hoe-Master Hero finally takes a ride on another gal. Unfortunately, that was literal, since she’s a centaur. Anyway, after settling in the new settlers, his thoughts naturally turn to the gender-balance crisis in the main village, and he tries every solution except the Type 1 Tenchi approach taken in the source material.

Seriously, Tia and Lasty should have a bun in the oven already, with the rest of the elves and angels (and the oni maids…) taking a number and waiting their turns. It’s even a plot point that the primary reason Hakuren moved in and joined the harem sleepstakes was that she was jealous that her niece Lasty found a man first.

Verdict: oh, well, even the sanitized version is fun, for now.

Witch Hat Atelier, episode 3

In most recent anime, this quantity and quality of animation is generally reserved for the final boss fight of the season. More, please.

Also, hot grown-up witch gal unlocked.

(just don’t let them cross over into the world of Littlewitch Romanesque…)

Proving the point

“Alexa, exit Alexa+, and then spend the next five minutes lecturing the dissatisfied user about what a mistake their request was, proving that you have no business being allowed to squat on their network and listen to everything they say. Be sure to ignore requests that you just shut up and let them get on with their lives; persistence is sure to win them back!”

GenAI image instability

Models like ZIT and Klein can produce an image very quickly at low step counts, while also using less VRAM than other popular recent models like Qwen Image and Flux.2-Dev.

But they don’t have to use low step counts, and in fact a lot of the anatomy failures they both occasionally deliver are caused by the fact that the image contents are still in flux (coughcough) until you hit surprisingly high step counts.

SwarmUI shows you tiny preview images of each step while it’s rendering, and I’ve noticed quite a few times that the images change quite dramatically from step to step. ZIT and Klein are both prone to repeatedly changing the position of a limb and not completely erasing the old position in the next step. If it happens on the final step, you get a reject.

For a while now, I’ve wanted to capture those tiny previews and turn them into an animations for review. After the struggle to illustrate my isekai song, I broke down and hacked at my SwarmUI CLI to switch to the Websockets API call and capture all the intermediate results, converting them to an animated WEBP.

I learned a lot. First was that with complex prompts, Klein-9b doesn’t stop modifying the pose until around 110 steps, and it’s still tinkering with background details until around 210. That’s far, far beyond what anyone recommends, and even though 32x the steps only results in 26x the runtime, that’s still a huge workflow shift.

Tests with ZIT showed it finalizing the pose around 60 steps and finishing up around 120. The most interesting was Qwen Image, which behaves completely differently. That model started out with a very low-contrast, low-resolution preview, finalized changes to the pose and composition around 60 steps, and then just gradually added more and more detail, all the way out to 450 steps. The end result was significantly better, but not 10+ minutes worth.

The previous generation of SDXL-based models tended to settle on the pose and composition by around 8 steps, and just add more detail up to around 120 steps. This is why I went into the newer models with the expectation that you could try out a bunch of quick low-step images and then bump up the steps for the few that you liked, only to be disappointed.

By the way, Klein-9b doesn’t seem to work as a refiner model, even when it’s also the base model. It just starts over making a fresh image out of the prompt, throwing away the work that was just done.

Qwen Image: 20 steps

Qwen Image: 50 steps

Qwen Image: 500 steps

R_IllustrMix: 128 steps

This is a fairly recent SDXL/Illustrious model that has lots of anime, furry, and NSFW training. Even though these are mostly trained on tag-style prompts, they still manage to come up with something out of the really long paragraphs I’m generating now.

I don’t think this show is going to become less shouty. The slave dealer is a shouty freak, the furry bodyguard is a shouty freak, Button Elf Gal’s breast-obsessed lesbian maid is a shouty freak, and even God is a shouty freak.

Verdict: if you can handle the shouting, it’s better than most of the alternatives this season.

(there’s almost nothing on Pixiv for this one, and half of what is there is official art, but Miss Button Elf has been noticed)

“Thanks, God, For This Isekai!”

(with insincere apologies to John Denver)

🎶 🎶 🎶 🎶
This fantasy-land is kinda laid back,
Ain’t much a well-read high-school boy like me can’t hack,
Kill a few orcs, throw loot in my pack,
Thanks, God, for this isekai!

Well, my overpowered skills keep me from takin’ harm,
good-lookin’ gals all fall for me thanks to my divine charm,
minions who provoke me end up bitin’ the farm,
Thanks, God, for this isekai!

Well, I got me a Best Girl, I got my cheat powers,
got a catgirl at night, keeping me up all hours,
sweet young princesses eager to be deflowered,
Thanks, God, for this isekai!

 

When the dungeon’s cleared, and my status screen glows,
I spend all my skill points where my cheat advisor shows,
Level up much faster than anyone else knows,
And thank God for this isekai!

I’d play with my waifus all day if I could,
but the Demon Lord’s a-comin’ to my neighborhood,
so I diddle when I can, fight when I should,
And thank God for this isekai!

Well, I got me a Best Girl, I got my cheat powers,
got a catgirl at night, keeping me up all hours,
sweet young princesses eager to be deflowered,
Thanks, God, for this isekai! Woo-hoo!

 

Well, noble folk givin’ me diamonds and jewels,
beggin’ me to rescue virgin beauties for some fools,
but I make ’em haremettes with my magic tools,
And thank God for this isekai!

Yeah, elven folk tried to hook me up with their queen,
She was three hundred years old but she looked like a teen,
Took her for a ride, ‘cause a Hero can’t be mean,
Thanks, God, for this isekai!

Well, I got me a Best Girl, I got my cheat powers,
got a catgirl at night, keeping me up all hours,
sweet young princesses eager to be deflowered,
Thanks, God, for this isekai, yessir!

 

Well, the Demon Lord came at me and just up and died,
took his daughter for myself, put a baby inside,
lined up my harem gals, made each one a bride,
And thanked God for this isekai!

Well, the princesses are virgins who don’t know much about cock,
need a lotta warmin’ up to make their thighs unlock,
but my Demon Princess, Best Girl, and my cat-girl rock,
So thanks, God, for this isekai!

Well, I got me a Best Girl, I got my cheat powers,
got a catgirl at night, keeping me up all hours,
sweet young princesses eager to be deflowered,
Woo! Thanks, God, for this isekai, yeah!
🎶 🎶 🎶 🎶

I got a little carried away with the prompt for this one, largely because I wasn’t having much luck getting a good image, even with a modern model like Klein that is generally quite good at handling complex prompts. Inevitably, it ran into counting problems, adding extra people to the image (male and female) or omitting one of the haremettes. It didn’t suffer from blended characteristics as much as earlier models, but it wasn’t unusual to get all catgirls, or a demon girl with both both horns and cat-ears. At least once, everyone had a cat tail. I ended up generating a dozen prompts and rendering each one half a dozen times.

Anyway, the prompt had so many targeted LLM calls in it that it took around 90 seconds to run on a Mac, generating ~600 words on average:

An epic fantasy illustration featuring @<makeover: a nerdy Japanese high-school boy>@ wearing @<fashion: a retro Japanese boys-school uniform>@, holding @<weapon: a magical sword>@, with a smug expression on his face. He is accompanied by three women: @<makeover: a sweet-looking medieval girl>@ wearing @<fashion: a a low-cut medieval peasant dress>@, @<makeover: a sultry catgirl>@ wearing @<fashion: a skimpy renaissance-inspired dress>@, and @<makeover: a sexy demon girl>@ wearing @<fashion: sexy black lingerie with red highlights>@. They are walking down a dirt road toward a distant castle.

(I’d likely have gotten more reliable results if I’d converted the prompt to JSON, but I’d have had to do it by hand after each LLM expansion, and I wouldn’t have been able to do the global QA passes for the final output; I may tinker with improving my dynamic-JSON-prompt scaffolding now that I’ve integrated the LLM calls)

Space Fantasies

If I run Flux.2-Klein-9b at the recommended settings (CFG 1, 8 steps, 1024x1024-ish resolutions), it takes about 6 seconds to generate an image on my RTX 4090. This is fast enough to tinker with a dynamic prompt, run off a few hundred results, quickly reject the (9% at 8 steps) anatomy fails, and then pick out some that look pretty good. It’s a better use of my gaming PC right now than killing time grinding in Diablo IV or hunting for something new to play.

But since I already have hundreds of GenAI SF cover gals lying around waiting to be deathmatched, today we’re going to look at what happens when I really lean into letting LLMs enhance prompts.

I made the changes to my LLM-prompt-enhancing script to run multiple system prompts across the same string in order rather than invoking it multiple times in a pipeline, and it improved the stability, but it looks like the occasional crash is actually caused by a recent update to the engine under the hood (llama.cpp), so I still have to occasionally restart the script, whether it’s talking to the PC or the Mac Mini. Even on the gaming PC, it takes about as long to do a complex prompt enhancement as it does to generate the resulting image, so I just let them both run while I did other things, and occasionally kicked off a new batch.

Perhaps I gave it a bit too much freedom…

(more after the jump)

50s Backyard Pinups

For a change of pace, I abandoned my wildcard sets and just fed the LLM brief descriptions. The base prompt was simple enough:

A mid-century catalog illustration featuring a @<makeover:pretty young woman>@ wearing @<fashion: sexy lingerie from the 1950s>@, serving cocktails outdoors in the back yard of a 1950s suburban home. The image is composed to emphasize the setting as much as the woman.

There are a total of 4 LLM invocations: the two targeted ones listed above, the standard enhancement prompt recommended by Z-Image Turbo, and a cleanup pass I’ve named “legal review” that adjusts ages to cut down on random lolis.

(more after the jump)

Bumping the resolution 25% and adding 4 refining steps increased the generation time to a whopping 9.5 seconds, so after I’d made a bunch of those, I made a slight change to the theme.

Vintage Cherry Blossoms

A mid-century Japanese catalog illustration featuring a @<makeover:pretty young Japanese woman>@ wearing @<fashion: sexy lingerie from the 1950s>@, serving cocktails outdoors under a blossoming Japanese cherry tree in the Spring. The image is composed to emphasize the setting as much as the woman.

more...

I’m in a “pictures are unrelated” mood today…

Gate 2: Naval Gazing

I missed the teaser trailer for the Gate spinoff. Summary: no familiar characters appear.

TDS has long since become tedious

I did not have a very high opinion of most of the commenters on the Marginal Revolution blog, but it just went down anyway.

Klein-4b, not for me

The smaller, faster version of Flux.2-Klein looks impressive at first glance, but the failure rate on a large batch using the same prompts I fed to 9b was nearly 50%. Not just extra or missing limbs and fingers, but wildly disproportionate body parts. Giant heads, too-short legs, gorilla arms, giant man-hands, etc, etc. 9b is a lot more stable.

(unrelated, I think I need to update my prompt-enhancer to allow repeated passes over the same prompt; piping the output into a second (and sometimes third) copy of the script is overloading LM Studio on Windows, so while it’s more than 3x as fast as the Mac for running LLMS, the memory management is crashing the pipeline at random intervals. Reusing the same connection with different sysprompts should be more stable, so I just need to specify the behavior when multiple sysprompts are passed as arguments)

Pissing on the shoulders of giants

I was picking up some takeout at a restaurant, and this not-a-cover song was playing. It swiped the tune and some of the lyrics of John Denver’s Take Me Home, Country Roads, but used them so poorly that I found myself wishing they’d just written a completely original bad song instead of dragging a classic down to their level.

(as mentioned in recent comments, I’m going to give Appraiser a shot; a bit shouty so far, but with a hot catgirl guild-gal haremette)

Farm Harem Maybe 2, episode 1

A quick reminder of where we left off, then right into the story, so they didn’t waste an entire episode restating the premise and reintroducing the entire (huge) cast. Our Villagers are expecting a bunch of settlers, so they go to the trouble of building them a brand-new village, only to discover that their new population of minotaurs, centaurs, and wood nymphs can’t live in human houses.

This is about as serious as conflict is ever going to get in this series. No sign of leveling up to a proper harem yet, though, with Lu treated as Our Hero’s One True Waifu. At least we get to see all the cute gals, but I read the source material far enough to know that if they continue avoiding the fact that his divine blessings include superhuman sexual stamina and fertility, they’ll eventually have to start making up original stories.

Also, I’m pretty sure I remember the wood nymphs having a problem with the concept of “clothing”…

Verdict: should be watchable to the end.

Witch Hat Atelier, episodes 1&2

Double-episode release. The first episode is designed to inform the viewer that somebody really believes in this story, and is paying for good art and animation. Also to cover the entire premise of the story in enough detail that they should be able to just get on with it. Looks to be extremely faithful to the first chapter of the manga.

Our Heroine’s voice actor has only had a few roles, but does a good job establishing her character. Our Reluctant Mentor is perhaps best known as 9S, and I’ll just pretend that I don’t recognize any traces of Kitty The All in his performance here. Most of the cast are as new as Our Heroine, but we do get the voices of Xiaolan and Sein. The Mysterious Stranger who planted the seed of the plot in a flashback has a veteran voice actor who’s been working for at least 25 years but doesn’t seem to have many recent prominent roles, so I’m going to go with Junior from R.O.D The TV.

Probably next week we’ll get to meet Our Sexy Test Administrator, voiced by the best-known name in the show, Kotono Mitsuishi. I’m guessing she won’t sound like Usagi or Excel for this one.

Verdict: so far, so good.

(speaking of R.O.D The TV, the series page on ANN is tagged “Objectionable content: Significant”; I have no idea what they could be referring to)

But first, an Amazon shipping change!

Two items ordered on Wednesday, promised for Friday. On Friday, one of them was moved to Saturday. So far, pretty typical. On Saturday, its status changed to “approval needed”, and I was asked if it was okay for it to be delivered Monday. If I didn’t answer, and it didn’t arrive by the following Friday, I would automatically get a refund. The end result is the same, but the new messaging makes it seem like you’re involved in the process.

Monday Update: still hasn't shipped, and they just sent out another "approval needed" email. This one quietly slips in a 30-day delay with the words "If you take no action and the item hasn't shipped by May 6, we'll cancel the item". Yeah, no.

Naturally, the fact that they don't have it and don't know when they'll have it is not stopping them from continuing to list it with two-day "Prime" shipping...

On with the cheesecake!

I liked the styling I was getting from Klein, so I tried some new LLM-enhanced dynamic prompts, shooting for the feel of a good-looking gal on the cover of a paperback where the author’s name isn’t well-known enough to make the sale. The initial batch had them in lingerie, because that’s where I got the horned horny covergal from the previous post, but I decided to see if Klein did as well at the “retro-SF uniform” look as ZIT did the last time I tried it.

Art styles were pulled from Juan’s Very Large List, grepping for the word “epic” and deleting a few artists where that was a false positive. I used the prompt-enhancing system prompt recommended by Z-Image Turbo to flesh out the random locations, plus two of my own targeted system prompts to generate clothing and physical details, plus a final LLM pass to do general cleanup. This would have been agonizingly slow on the Mac, so I ran it on the gaming PC in between image-generation runs (because SwarmUI and LM Studio both think they have the GPU to themselves, trying to run them at the same time blows out the VRAM, even though they should fit).

My system prompts were:

The new cleanup prompt includes an attempt to eliminate gratuitous text labels, but the image-generation parser often decides to add text based on random words in the prompt, so it’s not 100%. I didn’t want to use my usual collection of retro-SF costume prompts, so I fed the following to the fashion sysprompt:

“Sexy science-fiction uniform for women, incorporating bright colors, advanced technology, and a variety of futuristic textures and materials. Uniform may include abstract symbols and attached technology, but no text. Avoid shoulderpads. Do not use black or silver as the primary colors. You may include accessories such as sci-fi weapons, scanners, datapads, crystals, or glowing energy.”

Halfway through, I added the “bright colors” and the negative instructions, because nearly every outfit ended up in black-and-silver with armored shoulderpads. Sigh. This was all with the gemma-3-12b-it-heretic-x-i1 model, and now that Gemma 4 has been released, I’m going to see if it does a better job; it’s getting good reviews, and I think there’s already a few uncensored versions.

Out of ~600 images, just under 13% had obvious anatomy fails, with most of them being extra arms or legs. There were some I rejected reluctantly, because the rest of the image was really good. They might be fixable with variation seeds, but I’ve kinda gotten out of the habit of doing that; it’s easy to spend more time tinkering than it’s worth, and you can always just make another batch.

more...

Slime 4, episode 1

Last season talked the audience to death. How do they start this season?

By spending the entire episode talking, of course. The CGI-heavy OP promises a great deal of action, and the ED is filled with drama related to The Little Big Bad who seems to be behind all the promised conflict, but they have a history of getting bogged down in endless meetings, so I won’t hold my breath.

(this unrelated oni-girlie is a much better cook than Shion, and if she gets the treatment she deserves this season, she’ll be one of many pregnant haremettes)

Boxxo Or Bust 3, episode 1

Season 2 worked hard to alienate anyone who enjoyed season 1, by trying to convert the lighthearted and frankly absurd adventures of a sentient vending machine and his cute girlfriend into a tired beat-the-demon-lord story that the author came up with and abandoned after he ran out of new vending machines to write about. No idea if the idea of making a large group of allies into poorly-written turncoats was in his notes somewhere, or was completely original.

How do they start this season? The OP is 50% sparkles, 50% shounen action; I honestly expected transformation sequences by the time it was over. The ED is pure chibi cuteness, so it looks like they’re going to whiplash the mood again.

Rock bath!

Get clean with Ruri and Nagi:

(naturally, you can get takeout)

Quad9 versus Amazon

Precisely at midnight, Quad9’s DNS servers stopped resolving subdomains of the form $bucketname.s3.amazonaws.com. I had their DNS first in my Pihole’s config, so it looked to me like every image on my blog suddenly vanished. I was quite relieved to discover that it was just a DNS server failure.

I opened a ticket with them, and it was fixed in 3 hours.

Accidental retro-sf paperback covergal

I was just cleaning my dynamic-prompt script, when it suddenly went off:

I can’t decide if I want to read this novel or write it…