Grid Girls

Up to this point, I’ve been more-or-less taking the advice of model creators and uploaded pictures on CivitAI when it comes to choosing the sampler and scheduler settings for Stable Diffusion models, but this produced problems when I tried to compare the same prompt and parameters across a large group of models, to see how they handled details like faces, finger counts, lighting, depth of field, and of course, “paying attention to the prompt”.

I was going to do a detailed comparison of the 13x31 grid of pictures I got from testing identical settings with all of the available schedulers and samplers, but as I worked my way through the results, I learned an important lesson: don’t choose a reference pic where the gal’s legs are crossed and her fingers are interlaced. This is pretty much the worst-case scenario for evaluating SD images of human beings…

TL/DR: over a third of the combinations produced garbage, and about half of the rest looked very similar in the foreground with some minor out-of-focus differences in the background, but there were quite a few small differences in her clothing’s shape, color, coverage, and material. Face and hair were pretty similar, with only a few looking like a completely different girl, and maybe a quarter having the hair parted on the other side. A fair number changed the pose in some way, although there were maybe six different poses total out of 403 images.

Next time, I’ll set the test up more carefully, so I can actually draw some conclusions beyond, “yeah, just don’t bother with most of the samplers and schedulers”. 😁

“Customers say”

Amazon’s “AI” comment-summarizer says this:

Customers find the story engaging and action-packed. They describe the book as a fun, intense read that is worth reading. The series is considered good to great by customers. Readers appreciate the complex characters and the author’s writing style. The pacing is described as fast and consistent. Overall, customers praise the author’s writing quality and consider it an excellent military adventure.

Human summarizer says, “OH JOHN RINGO NO!”. 😁

Artificial Incantations

How to get Flux.1-Dev to stab an orc: “…bleeding from a large chest wound. A sword grows vertically from the wound.” The official release seems a bit vague on what an “orc” looks like, but with some extra prompting will do the right thing:

side view, at night. photograph of a male ((orc)) warrior with green skin, pointed ears, and tusks, wearing armor, ((lying on back)) on a battlefield with his eyes closed, bleeding from a large chest wound. A sword grows vertically from the wound.

(new anime? not until Saturday!)

The relative ease of customizing Stable Diffusion models means that thousands of people are stirring the pot and training their own. This is good, since the official models are biased and censored, but it’s also bad, because the derivative models are biased in different directions, and often over-trained to the point that they simply snap when you find their edges.

Most people don’t do their custom training against the base SD models; they layer their collection of picture/keyword pairs on top of one that’s already been “uncensored” or augmented in some way, with the two major anime branches being Illustrious and Pony. What this means in practice is that feeding the same settings to related models will often produce very similar results.

So, just how similar do they get?

I’ve been using SwarmUI’s grid feature to evaluate different models by passing them all the same prompt, seed, and settings.

For each set, I used a character LoRA (small patch model that can be used to add character/style/location data onto other models with varying success depending on heredity), and generated multiple pictures in my go-to model for cute-and-occasionally-naughty material, CAT - Citron Anime Treasures (Illustrious-based), until I found something that looked like a decent starting point:

Setting aside the boilerplate and the character trigger words, the prompt was:

laughing, standing with arms spread, head back, grounded stance, freedom in motion, outdoors, at Santorini, Greece

more...

Once you’ve asked chatbots for information a few times, you start to spot patterns. Here’s a perfect example: I asked Google about putting vanilla extract in tiramisu (something traditional recipes don’t do). High on the list of results was Spoonable Recipes, and every line screams generative AI:

Mascarpone cheese is the most popular ingredient in tiramisu dishes. In fact, over 80% of tiramisu recipes contain mascarpone cheese.

Softened is a frequent preparation for mascarpone cheese in tiramisu dishes.

Mascarpone cheese is often included in tiramisu dishes in amounts of 8 ounces, 1 pound or 1 cup.

Another popular ingredient in tiramisu is white sugar. From the recipes we’ve sampled for tiramisu, over 70% have white sugar.

Tiramisu dishes often call for white sugar to be granulated.

White sugar is often included in tiramisu dishes in amounts of 1 tablespoon, three quarters of a cup or a quarter cup.

…

Another popular ingredient in tiramisu is vanilla extract. From the recipes we’ve sampled for tiramisu, over 40% have vanilla extract.

Vanilla extract is often included in tiramisu dishes in amounts of 1 teaspoon or half a teaspoon.

In tiramisu recipes that contain vanilla extract, it is on average, 0.7% by weight.

In recipes for tiramisu, vanilla extract is often used with mascarpone cheese, ladyfinger cookies, confectioners sugar, chocolate and white sugar.

Potential substitutions for vanilla extract in tiramisu:
    pumpkin pie spice

Also, vanilla extract is not often used with flour, white chocolate, pumpkin and lemon curd.

And remember, that cream won’t whip itself!

(as investor-hungry “AI” companies frantically scramble for fresh content to build their next-generation engines with, they’re hoovering up previous-generation output like this “recipe analysis” and spreading the contamination. A lot of people doing text-to-image generation rave about Flux over Stable Diffusion XL, but the first time I tried it, I got even more fingers per hand; one poor gal must have had a dozen, and that’s enough ladyfingers for three full servings of tiramisu!)

Merry Chris-mas to all…

… and a happy Kiwi!

Or, in the words of Henry The Red, “Thank you, generous hosts!”

(classical reference)

Happy Accidents

What I asked for (madlibs-style):

[boilerplate]. A pin-up photo of a pretty teen girl, [female-hairstyle], [sexy-pose], with a [positive-mood] expression, wearing from 2-5 [lingerie], at [famous-place]. [framing-light], [camera]

Negative: pregnant, frame, cropped, [negatives]

What that expanded to:

4k, breathtaking, crisp, gorgeous, high budget, highly detailed, intricate, professional, ultra textured. A pin-up photo of a pretty teen girl, Soft chignon, neatly tucked at the nape of the neck, Lying on side, elbow bent, hand supporting head, elegant silhouette formed, with a grateful expression, wearing compression shorts, bow-front bikini and camisole set, lace-up bodysuit, striped panties, at Scottish Highlands, Scotland. Harsh midday lighting, wide angle, scattered elements, vibrant contrast, dynamic shadows., Eye-level angle (from behind) view medium Close shot (focus on feet).

Negative: pregnant, frame, cropped. bad anatomy, bad proportions, banner, censored, collage, cropped, deformed, disconnected limbs, disfigured, duplicate, error, extra arms, extra digits, extra hands, extra limbs, fused fingers, grainy, gross proportions, logo, long neck, low contrast, low quality, low resolution, malformed limbs, missing arms, missing fingers, multiple panel, mutated, mutated hands, mutated limbs, out of focus, oversaturated, poorly drawn eyes, poorly drawn face, poorly drawn hands, signature, split frame, split screen, text, ugly, ugly, unreal, username, watermark, worst quality.

What I got:

“This doesn’t look like the Lincoln Tunnel, Sam.”

Also not a close-up foot shot in harsh midday lighting in the Scottish Highlands. And I didn’t even ask for an army; they must have been assembled from “silhouette”, “scattered elements” and “dynamic shadows”.

(classical reference)

How to think about generative AI

You are the pilot of a ship capable of traveling the multiverse. The cockpit contains thousands of unlabeled buttons, switches, dials, and sliders. Think Tardis, but taken to 11.

You may adjust any number of controls before hitting The Big Red Button, and then you will be transported to a completely different universe, where anything can happen.

With me so far? Good.

Order matters. As you adjust each control, your destination in the multiverse shifts, so that each additional control you adjust applies its effect from a different starting point. Turn a dial too far in one direction, and your destination could be so far from home that slugs are the dominant species on Earth, and you can’t get back by pressing three green buttons and toggling a switch.

Still here? Awesome.

Here comes the real fun: each control on your board is actually wired up to ten thousand completely independent engines, all of which impact your destination in some way, big or small. In the first engine, that dial setting puts slugs in charge, but in engine #751, it puts giant breasts on cats. Not catgirls, cats.

By your side is a quirky robot with tentacles. Its job is to convert your spoken orders into control adjustments, but it doesn’t fully understand human language, has a dim grasp of each control’s effect, and guesses to resolve ambiguity. Because it has received contradictory orders from its makers, at random intervals it goes insane and assaults you with the tentacles; it never remembers these episodes.

(I let DALL-E handle this one, because I’m busy cleaning, baking, and wrapping presents for Christmas dinner tomorrow)

After the jump, a not-so-Christmas NSFW Miracle!

more...

Forced cleaning

Thanksgiving will be at my house… next week?!? Time for some serious housework, at least in the public areas. I could call in a maid service, but I don’t think they have any Mysterious ones.

…and don’t call me Shirley

FYI, Anne of Green Gables has long been quite popular in Japan, one of the few things my friend Dan knew about the country (having spent time dealing with tour groups to Prince Edward in his misspent youth). So it shouldn’t be a surprise that there will be a new anime series next spring. From the brief description, it sounds like it will run for multiple cours.

(not planning to watch it; I’d rather have more Frieren)

Amazon Primer

The most useful announcement in this streaming roundup is that Amazon is folding Freevee into Prime, which means that I’ll finally be able to watch the newer Bosch seasons without ads. It never made sense to me that there was no way to watch Freevee shows without ads, no matter how much you paid the mothership each month.

Well below the fold is the announcement that Apple is losing a fortune on AppleTV+ and is trying to license their $20 billion in original productions to other services. Which means I might finally watch some of them. Now if only Disney and HBO would do the same…

(the only time I ever see television ads is when I’m in the barber chair, and it feels like Alex’s “treatment” scene in A Clockwork Orange; Amazon often tries to sneak one in when you start watching a show on Prime even if you pay the ad-free premium, but you can skip it)

Separate Rubbers

I mute every xTwitter account that displays an ad on my timeline, except for the truly obnoxious ones, which I block. They don’t provide any real management tools for these lists, so I can’t say how many accounts I’ve muted, but with an average of one ad for every 3 tweets viewed, it’s easily several thousand. Most are for products I’d never buy, sites I’d never visit, and people who wrongly think they have something to say that’s worth paying for, but many just seem like desperate cries for help.

Honestly, the one I came closest to engaging with was for an Etsy dealer advertising handmade bedside emergency condom boxes, and my only interest was in finding out if they were deliberately using the word “discrete” instead of “discreet” to name their product (and it was not a one-off typo).

…Yatta

(coincidentally, around midnight the clouds burst open and my neighborhood was inundated with liberal tears…)

Loser Food

I’ve never wanted turkey on a pizza, but it at least sounds plausible. Green beans and cranberries, however, are way over into hell-no territory. Might as well eat the bugs at that point.

(classical reference)

If we make it to January…

…we get more Maomao. With luck, we’ll still have Western Civilization as well.

We’ll also get Behemoth…

TL/DR: half-nekkid busty elf and friends who really like her kitten. This is definitely based on the manga, since the light novels collapsed after only two volumes.

(unrelated half-nekkid busty elf chick is half-nekkid, busty)

Age-old problem

Massad Ayoob’s advice for people having trouble getting crisp front-sight focus with pistols as they get older is to have reverse bifocals custom-made, where the reading prescription is on the top. My eye doctor’s advice on Sunday was to just go to Amazon and buy stick-on bifocal lenses and optionally trim them to cover just the part of your glasses that you use for sighting.

Ayoob talks about using your reading prescription, but the doc recommended using a computer prescription (diopter = NV-ADD / 2, rounded up to nearest 0.25; add to Sphere for dedicated computer glasses). Of course, Ayoob was talking about wearing them out in public, while I just want want them on my range glasses (which are safety glasses from Zenni).

Another idea comes from Good Glasses, which recommends splitting your prescription, putting your computer prescription in the dominant eye and distance prescription in the other.

My concern with the stick-ons is based on my historical inability to get window film to cling properly with no bubbles, and some reviews I’ve read have matched that experience.

(in addition to the new prescriptions and shooting advice, he also had some interesting reloading recommendations for light revolver loads for kids)

(related, his office is attached to a Costco, and while I gather that in some states their stores post no-gun signs, that is not the case in Ohio)

Bye-Bye Earth, the novels

Translation of the first one coming in April. I don’t know if I’ll be willing to take one for the team, but it would be nice to find out if the adaptation was bad, or if the source material was incomprehensible dreck, too.

How to sell an American SF novel…

Sexy catgirl with a blaster.

(this is not your daddy’s C’Mell)

Grid of thirds

JetPens has an interesting kanji practice notebook, with a useful feature I hadn’t seen before: thirds markings. It has the usual half-grid lines, but also dots at the thirds to help you balance characters that aren’t split evenly. Naturally I immediately cloned this feature into my Perl script and made a PDF to try it out.

Explaining The Lost Dangerous Visions

So, The Last Dangerous Visions is out, finally completed after Harlan Ellison’s death by executor J. Michael Straczynski. I’m working my way through the stories slowly, but I thought I’d comment on JMS’ intro and explainer, in which he gives a brief history of his close relationship with Ellison over the decades. What explanation does he give for all the delays, broken promises, and outright falsehoods that surrounded this project?

TL/DR: Ellison was manic-depressive for his entire adult life and not only refused to recognize it for most of that time, but held off on seeking treatment until way too late.

Yup, that’d do it. I don’t discuss family here much, but I am very familiar with what happens when a bipolar person goes off their meds, and it maps well to Ellison’s reported behaviors. I would never have tried to diagnose him myself, since I was only ever in the same (large) room with him once, and his behavior that hour was quiet enough that I only knew he was there because he loudly harrumphed upon learning that C.J. Cherryh had just written a Lois & Clark tie-in novel. The crowd’s oh-Harlan reaction was louder than he was.

How are the stories so far? Pretentious twaddle from a vanished world where the Left still thought their ideas were “edgy”. Some of them have the excuse of having been written decades ago by people who died waiting for Ellison to publish, but others do not.

My-HiME, revisited

Not by me, I haven’t lit one up in years; the original show just had its 20th anniversary. The busty and titular heroine Mai was easy on the eyes, and the spoiler-heavy episode previews they inserted for the DVD release led off with a memorable slow-pan of her in lingerie, but the story was a mess, with too many heroines and too many mechs and too many easily-killed keys/hostages. Much of this was due to the fact that the original manga was more of an action/harem story centered around Tate, who got some memorable “unlocking” offers from various girls.

The second and third series were better planned out.

Unrelated, “secondhand smoke” was always bullshit

…and the American Cancer Society finally agrees. Yes, cigarette smoke is foul stuff. Yes, constant exposure has an impact on people with respiratory conditions. Personally, I’ve always rated it right up there with “excessive hairspray” for interfering with my breathing, and second only to pot for leaving rooms and clothing reeking.

Pipe smoke, on the other hand, is potpourri.