The web is full of praise for Dolly Parton today, with the usual vicious bitter losers getting bitchslapped Left and Right for trying to soil her memory. Other people have explained her accomplishments far better than I ever could, and my only contribution to the bitchslapping was to speak bluntly to an asshole on one of the Discord forums I follow (who seemed genuinely shocked that no one praised him for his “contrarian” take).

A remarkable human being has passed.

The Lament Of Tanya the Practical 2, episode 8

In which Our Tiny-But-Fierce Heroine learns that you should never give ideas to a crazy man, and Our Long-Suffering Superior Officer learns that you should never give open-ended orders to a crazy little girl.

Bumpkin & Harem 2, episode 8

This week, two very different maidens gird their loins to do battle in another sort of arena, with Our White-Haired Haremette’s jealousy nearly going into overdrive as Our Noble’s Little Sister makes her intentions painfully clear. Just when we think it’s going to end in a glorious catfight, though, shit gets real.

Verdict: never mind the slow pans up from cleavage to face, it’s time for serious plotting!

(clearly I need to have the LLM design some more revealing costumes…)

Dear Apple,

Number of times I have taken a screenshot of my iPhone or iPad: ~753

Number of times I have wanted a screenshot of my iPhone or iPad: 0

(this is especially annoying on the iPad, where “swipe up from bottom” is overridden if you’re too close to the lower left corner, and replaced with a popup telling you that swipe-up-for-screenshot is disabled; YES, I KNOW; I’M THE ONE WHO DISABLED IT!)

AI Productivity!

Meta’s AI Facepalm:

On one hand, the company’s data showed that code changes to Meta’s AI software platforms and infrastructure were up 220 percent year over year. But that didn’t translate into equally dramatic productivity gains: new or improved features that actually reached users rose just 36 percent. And advancements in those areas were offset elsewhere, as the number of technical and security incidents rose 40 percent. The time employees spent on those problems grew by 70 percent.

WTF, Grok?

xTwitter’s automatic translation feature is generally useful, but far from reliable. It can’t reliably translate Japanese names, for instance, even if the correct reading for the name is in the username or profile.

But it also can make huge mistakes. Like this one:

😈🎉 Blu-ray Vol.3 On Sale Today 🎉😈

TV Anime “Miss Kuroitsu from the Monster Development Department Is a Total Pushover”
Blu-ray Vol.3 is finally hitting shelves today!!!

Featuring an original draw by nonco-sensei for the BOX design, this luxurious Blu-ray also includes a bonus CD <Serious ASMR by the Little Devil Sisters (CV. #Izumi Fuka & #Tono Hikaru)>! ✨

In episode 12’s beach volleyball scene, there is no white light for Lilim 💡

If you’re as baffled as I am by the mention of a beach volleyball scene in the finale, the “total pushover” added to the title, and the fact that Miss Kuroitsu came out four years ago, then you’ll understand why I switched it back into Japanese:

TVアニメ「カナン様はあくまでチョロい」
Blu-ray Vol.3 いよいよ本日発売開始！！！

nonco先生描き下ろしBOX仕様で、 特典CD〈小悪魔姉妹(CV.#和泉風花 & #遠野ひかる)による本気ASMR〉も収録された豪華なBlu-rayです！✨

第12話のビーチバレーシーンでは、 リリムの白い光はございません💡

▼詳細はこちら！

Yes, it translated “カナン様はあくまでチョロい” (Mistress Kanan is Devilishly Easy) as if it were “怪人開発部の黒井津さん” and added “is a total pushover”. Even without being able to read Japanese, you may notice that the original titles have exactly zero characters in common.

Krea 2 step count

With other recent image-gen models like Qwen Image, Z Image Turbo, and Flux.2-Klein-9B, the recommended step count is a minimum, and they continue improving and stabilizing for much longer than most people expect, with diminishing returns not setting in until 60, 90, or even 120 steps. Qwen Image, in fact, could keep tweaking small details at 300+ steps, which pretty much nobody has patience for.

With Krea 2 Turbo, just stop at 10. The quality doesn’t get worse at higher step counts, but it also doesn’t get visibly better. If you want to improve quality, add 2 refiner steps, and then feed it to SeedVR2 for final polish and any desired upscaling (there’s no reason to use any other method of upscaling; they’re all slower and/or lower-quality).

Fun with wildcards

All of my new fantasy wildcard prompt elements were generated with the following sysprompt, fed to gemma-4-26b-a4b-qat:

You are an orchestrator. Spin up 10 subagents to each create lists
of 10 unique, creative image-generation prompts describing $aesthetic
$things$class. Assign each subagent a different environment and
scenario. Do not use names of real places. Do not include people.
Do not include art style, medium, or depth of field. Describe only
the visual elements of each $thing in concrete terms, as a
self-contained paragraph of 25 to 75 words. Do not number the list
or include any text formatting. Once done, spin up a subagent to
review the total output and remove duplicates. Do not worry about
output length or truncated results; longer results are better than
shorter ones.

It’s not actually “orchestrating” anything or creating “agents”, it just pretends and generates 10 lists of 10, but that pretense makes it run 2-10x faster, complain less, and produce more creative results, to the point that I can run it multiple times with the same variables, and get maybe 0.25% duplicates.

The one instruction it doesn’t obey is the word length, usually stopping around 13-15. Enough for uniqueness, but not for flavor, so I run it through a different model with two different sysprompts.

---

First:

Update the supplied setting, adding and editing details to emphasize a
$aesthetic aesthetic. Do not use metaphor or emotional language. Do
not explain or describe the effect or impact of your creation. Do not
include art style, medium, camera angle, depth of field, bokeh, or
lighting. Describe only visible elements. Output prompt result only,
with no other text, as a single flowing paragraph. Do not include any
explanation. Do not include any text formatting or Markdown. Answer
only in English.

This expands it to somewhere around 80-120 words, at which point I use one of Juan’s prompts to bring it back down:

You are a precision text compressor. Your task is to condense detailed
descriptions to fit within a specified word limit while preserving the
core visual/conceptual essence.

1. Strip first: Remove adjective stacking, metaphors, similes, and
   flowery language
2. Prioritize: Keep subject, key visual elements, distinctive
   features, and actions
3. Merge: Combine related details into single efficient phrases
4. Be literal: Replace poetic descriptions with direct statements

- Hit the target word count (allow ±5 words flexibility) of 25-75 words
- Never invent new details
- Preserve proper nouns and specific terminology exactly
- Use commas and semicolons to pack information efficiently
- No filler words ("very," "really," "quite," "somewhat")
- Target word range is 25-75 words

You will receive a description to compress.

Return ONLY the compressed description. No commentary, no word count,
no explanations.

Over about 1,200 prompts, this is averaging 36 words per. For this latest batch of 4K vertical wallpapers, I added back some of my existing 1girl wildcards, varying hair/skin/eye color, ethnicity, etc. I left the pose constant as “wielding a weapon in a dynamic action/combat pose”, and still got decent variety.

I think out of about 800 gens, I got one extra limb, and maybe a dozen deformed swords, mostly with hilts at both ends that made them look like gaudy pizza cutters.

---

Comments via Isso

Markdown formatting and simple HTML accepted.

Sometimes you have to double-click to enter text in the form (interaction between Isso and Bootstrap?). Tab is more reliable.

---