Lunch is a battlefield

Wed 9/2/26 8:57pm

Comments: 0

Anime

The Melancholy of Tanya Degurechaff 2, episode 9

After ensuring the safe return of the Africa Corps last week, we get a peaceful tour of neutral not-Italy, with Our Over-Deployed Heroine driven insane by her first taste of good food in a long time.

Which makes their return home all the more jarring, especially when confronted with just how out-of-touch the high command is about the progress of the war.

Unrelated, I continue to wish that they hadn’t gone with the Innsmouth look for Visha. Everything about her is hot except for those fish-eyes.

Bumpkin & Harem 2, episode 9

The plot is thickening, and so are the harem antics. First Red siezes the chance to show him her favorite restaurants and wangle an invite to dinner, then White lets her mask slip far enough to blurt out her conflicted feelings with just a thin frame of “hypothetically…”, and then Our Noble Daughter delivers room service and states her intention to move to the capital to be near him, and to wait as long as necessary for him to answer her affections. Meanwhile, Our Junior Haremettes spontaneously hook up with Our Darling Daughter for some side training; it’s a valid tactic, and one I’m sure Noble Gal will adopt as soon as she’s in town.

On the serious side, they’re escorting the princess to her wedding, fully expecting another ambush by the people who kicked their asses last time. Side note: the death-flag-tripping guard who announced his plan to propose to his girlfriend did not die last week. They kind of left that ambiguous, but sure enough, he’s still breathing, although he may not be coming back to the guard.

Verdict: everything’s moving along, and I really, really hope that Beryl comes home to find that Mewi has been hosting all-night parties for her friends. I mean, why else would a teenage girl want to have the place to herself for two months?

Where do you plug in the joysticks?

Acer’s just-announced AI desktop PC looks like it should be sold at Sears in a bundle with three game cartridges, in 1978.

(fun fact: my Atari 2600 was the Sears-rebranded version; might still have it, down in the basement)

Speaking of retro video games…

This worked out better than expected. One of the many gaps in the vocabulary of image-generation models is the concept of playing spin-the-bottle, so I asked Claude to describe the game in a useful way. I had to tweak it a bit to get the speedlines, but at least now it’s recognizable, and combines nicely with the 16-bit LoRA.

(I still haven’t gotten an LLM to come up with a way to describe “knife switches” that works for image-gen)

More deterministic retrofying

Amusing as the LoRA is, it still randomly switches between grid and palette sizes, so I thought I might as well do it properly. By asking Claude and then cleaning up the results:

from PIL import Image
def retrofy(img, grid=5, palette=64):
    if not isinstance(palette, int) or palette < 2 or palette > 256:
        raise ValueError("palette must be an integer between 2 and 256")
    if not isinstance(grid, int) or grid <= 1:
        raise ValueError("grid must be a positive integer > 1")
    if img.mode == 'RGBA':
        img = img.convert('RGB')
    
    downsampled = img.resize((img.width // grid, img.height // grid),
        Image.Resampling.LANCZOS)
    quantized = downsampled.quantize(colors=palette, 
        method=Image.Quantize.MAXCOVERAGE)
    pixelated = quantized.resize((img.width, img.height),
        Image.Resampling.NEAREST)
    if pixelated.mode == 'P':
        pixelated = pixelated.convert('RGB')
    return pixelated

Well, it looks like a good idea…

Saw a Japanese Youtube short advertising this mandoline with collection bin and captive hand-protector/food-guide:

The pic comes from Amazon Japan, because I couldn’t find it imported anywhere. I’ve never liked the separate food guides that most mandolines come with (including mine); they tend to have trouble maintaining solid contact with what you’re slicing. The shipping and customs fees would be ruinous, so I was going to put it on my shopping list for the next time we’re in Japan, but then I noticed something.

There are no reviews. At least 20 different listings at a wide variety of price points, some with obviously Chinese branding, others just no-brand, but the one thing every listing had in common was a complete lack of reviews. Nobody’d even 1-starred one of them. So, yeah, the only way I’d buy one is if I found it in a store and had the chance to examine the parts in detail.

Dear Amazon,

What idiot decided that wishlists should be dynamically reordered by which one was most recently opened/edited, instead of the long-standing behavior of alphabetical-fucking-order? For more fun, you decided on an arbitrary limit of lists to show, with the rest hidden behind a “more” button, despite the fact that there’s plenty of room in the sidebar, because lists scroll.

(the icing on the cake is the presence of “Alexa List” in the first half of the list, which I’ve never, ever used)

Dear Apple,

I exited Safari. 14 GB of kernel memory was freed. Not application memory, kernel.

L’il Stinkers

An overnight feature by the creator of SwarmUI has done me two big favors: first, you can now turn off the legacy parsing that often interferes with use of JSON prompting; second, you can now do prompt-weighting in modern models.

To unpack that a bit, there was SDXL-era syntax for adding emphasis to certain words in the prompt, as “(boobs)” or “(boobs:1.5)”. This not only didn’t work with recent models that prefer complete sentences, there were some annoying edge cases where the presence of parens and colons in a JSON prompt caused it to try to interpret random strings as the emphasis weight, blowing chunks in the middle of a batch run.

So you can now uncheck the ParseAlternativePromptSyntaxes user setting, and use JSON-friendly weights such as <weight[2]: head-and-shoulders view, subject on left> and really push the model to do what you want.

How effectively? I removed all my hacky prompt overloading from Monday’s cowgirl and just emphasized the original breast phrasing with weight 2, and:

Another thing the new weighting is useful for is overriding the composition. Putting <weight[2]:subject on <random:left,right>> at the start neatly overrides the default tendency to center the main subject, and the same expression can also include things like “close-up of eye”, “head-and-shoulders, facing right”, etc.

And it turns out the simple front-loading even works for things like “super-deformed chibi, with skunk ears and skunk tail”: