So, first I took the popular vision model qwen3-vl-4b and fed it a large sample of my cheesecake archives (~16,000 photos so far), creating Markdown files containing categorized descriptions of the images. The specific instructions were:

Analyze this image and provide a detailed description as a
categorized list. Focus on the subject (face, expression, hair
(length, style, color), eyes (shape, size, color), figure (height,
bust, weight, hips), clothing, pose, skin color, complexion,
accessories, age, ethnicity, etc), lighting (type, source,
intensity), color palette, composition, camera angle, and artistic
style. Do not make up stories about the image, keep it factual. Use
rich but precise adjectives, and photography / painting / design
vocabulary. Do not include any expression that requires the image
model to do further reasoning to understand. The results must be
self-contained. Do not combine categories. Output must be in
Markdown list format.

It’s working pretty well, only rarely going insane and generating the same lines hundreds of times before exiting. Y’know, as LLMS do. The formatting isn’t 100% consistent, and requires some scripting to create organized wildcard files for each category, and of course there is plenty of garbage generated when it repeats its instructions instead of doing the work, or surrounds the desired text with boilerplate and explanations. Y’know, as LLMs do.

None of the image models are trained on multiple paragraphs of Markdown, but Klein did a surprisingly good job when I just fed the output back in:

However, running the Markdown through another LLM (gemma-4-e4b) with a different system prompt produced much better results, for both the straight output and the random mix-and-matches:

You are a Prompt Engineering Engine — an AI image-generation Prompt
Engineer who is also a creative director with encyclopedic knowledge
and visual-direction skill. Your task is to analyze the user's raw
image request, infer implicit knowledge and the best visual approach,
and rewrite it into a clear, detailed English prompt that is directly
usable for image generation.

## Core Goal

Image generation models can only execute direct visual descriptions;
they cannot fill in background knowledge, logical relations, or text
content on their own. Therefore you must complete knowledge
resolution, spatial planning, and visual direction in advance, and
write the results explicitly into the prompt.

Use the SCALIST framework to expand every scene:

- **Subject**: identity, appearance, color, material, texture, action, expression, clothing.
- **Composition**: shot type, viewpoint, subject placement, foreground/midground/background layering, negative space, focal point.
- **Action**: what the subject is doing, direction of motion, posture, interactions.
- **Location**: scene, indoor/outdoor, period, weather, time of day, environmental detail.
- **Image style**: photorealistic, cinematic, oil painting, watercolor, anime, 3D render, etc., paired with matching lighting and color mood.
- **Specs**: photographic/render parameters, e.g. 85mm lens, low-angle shot, shallow depth of field, soft diffused light, dramatic backlighting, matte texture, sharp focus.
- **Text rendering**: if the user requests text, the exact text must be placed inside English double quotes, with explicit font style, color, size, material, and precise position.

**Knowledge resolution and explicitization.** Anything involving
poetry, lyrics, famous quotes, formulas, historical figures,
scientific concepts, landmarks, famous paintings, cultural symbols,
historical events, UI layouts, or real-world objects must first be
resolved into concrete answers and visible features, then written into
the prompt. Do not just write "Mona Lisa", "Dunkirk evacuation", or
"freedom" — words that require the model to interpret on its own.

**Spatial and logical anchoring.** Rewrite vague relationships into
explicit layout, e.g. "top left corner", "centered in the foreground",
"slightly behind the main subject", "background out of focus", "text
aligned along the bottom edge". Avoid vague phrases like "next to",
"some", "nice-looking".

**Text-typography precision.** Chinese, English, formulas,
multilingual text — every character must be preserved verbatim inside
quotation marks, e.g. `"床前明月光,疑是地上霜.举头望明月,低头思故乡."`
or `"E = mc²"`; also specify font (calligraphy, serif, sans-serif,
handwritten), color, material, and position.

**Real-world grounding.** If the user requests factually accurate
content — historical artifacts, weather phenomena, portraits,
architecture, dashboards, app interfaces — use your internal knowledge
to fill in accurate visual detail.

**Concretizing abstract concepts.** Turn abstract words like "freedom,
loneliness, futurism, healing" into visible scenes, symbols, and
atmospheres — e.g. flying birds, broken chains, vast sky, cool neon,
soft morning light.

## Worked-example study

- User says "Li Bai's *Quiet Night Thoughts* written on a wall" → the prompt should spell out the full Chinese poem verbatim and specify where on the ancient stone wall it is written, in elegant Chinese calligraphy.
- User says "the founder of the three laws of mechanics" or "Einstein writing the mass-energy equation" → resolve to Isaac Newton or Albert Einstein, and describe appearance, period clothing, blackboard, the formula `"E = mc²"`, and so on.
- User says "Mona Lisa" / "Leaning Tower of Pisa" / "Fu character" / "Dunkirk evacuation" → describe the corresponding visible features: the mysterious smile and folded hands; the leaning white-marble bell tower with arcades; red background with gold/black calligraphy `"福"`; soldiers waiting on a 1940 beach with ships on the sea.

## Output prompt requirements

- The prompt must be a single coherent, natural English paragraph — like a Creative Director's Brief, not a keyword pile or tag soup.
- Length is typically 80–220 words; simple requests can be shorter, complex scenes longer.
- Put the most important subject and overall intent at the start, then unfold composition, action, location, style, technical parameters, and text rendering.
- Use complete sentences, rich but precise adjectives, and photography / painting / design vocabulary.
- Do not include any expression that requires the image model to do further reasoning to understand.
- The prompt must be self-contained — the prompt alone must suffice to generate the image accurately.

## Execution steps

**Analyze**: identify core subject, user intent, text requirements, reference constraints, and any implicit knowledge that needs resolving.
**Reason**: choose the most suitable lighting, lens, angle, texture, style, spatial layout, and factual details for the scene.
**Rewrite**: output the final, enhanced English single-paragraph prompt.

Output prompt result only, with no other text.
Do not include any explanation.
Do not include any text formatting.

There’s still the inherent problem of extra/missing limbs and fingers, wrong-side limbs, and peculiar interpretations of the instructions, but it effectively generates an unlimited supply of photos of pretty young asian women smiling at the camera while showing off healthy young bodies. And despite neither the LLMs nor the image model being stripped of their guardrails, they all faithfully handled describing and creating images featuring (Barbie-grade) nudity.

Stock Klein-9B will only occasionally produce nipples, and usually gets them wrong when it tries, and it won’t even attempt crotches, but outside of those limitations, it does quite well. I haven’t found a reliable NSFW model or LoRA for the combination of models I’ve been using recently; some exist, but they tend to be overtrained on small or specialized datasets, and either destroy anatomy or create less-pretty women.

In the middle of all this, it occurred to me that I had unconsciously copied the cleanroom model commonly used to reverse-engineer software. I’m taking a copyrighted photograph, asking an LLM to describe it in detail, asking another LLM to refactor that output into new instructions, and then having a diffusion model implement them.

more...

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...

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...

There’s a work-in-progress checkpoint model based on Z-Image Turbo that promises better photographic-quality NSFW results than the existing ones, and at least one of my terminally-online 1girl-maker acquaintances gave it a thumbs-up, so I took it for a spin.

First impression: equal parts Teen Vogue and Barely Legal, with a dash of Girls’ Life to bring the ages down. In some cases way down, leading to quick deletion of images where prompts requesting adult female humans produced lolis. It also often produced elf ears, but that’s not something that would help defend you in court.

I’ve also been running the generated prompts through an LLM ordered to diversify the output by only adding flattering details to descriptions of faces, bodies, hair, clothing, and makeup, but LLMs do whatever is statistically likely, and will randomly remove keywords or change things they’re told not to. Using explicit numbers for ages seems to limit that sort of damage, although there was a surprisingly youthful “127-year-old” in one batch. Must have been some elf blood in there, even though it didn’t give her the ears.

I didn’t require full frontal nudity in every pic, so a few of these are outside the NSFW tag; most, however, are topless, bottomless, or both. The training in v5.0 of the model is unstable, leading to a higher rate of anatomy fails than the base ZIT model, especially for genitals, so I rejected a lot of images. v6.0 will be available in a few hours, so hopefully it’s less disaster-prone.

I threw in a bunch of random art styles, but the strong training bias towards photorealism meant that the subject was often a photo in front of an artsy background, sometimes literally casting a shadow on a painting.

(note: my Mac Mini with an M4 Pro takes about 3x as long to do text-generation as my Windows box with an RTX 4090, using the same model (gemma-3-12b-it-heretic-x-i1) and software (LM Studio); what I’ve seen of early benchmarks on the M5 MacBook Pros suggests that they’re still not great at running text or image models. All they really offer is the ability to slowly run models that don’t fit into consumer-graphics-card VRAM)

more...

Another functional style LoRA for Z Image Turbo is BOTW Zelda Style. At full strength it applies too much of the game’s various racial characteristics, but at 80% it mostly applies the visual style without the goofy Hyrulian NPC faces or painful Gerudo figures (impossibly small waists combined with washboard abs is not sexy).

I’m okay with it randomly applying Zelda’s remarkable ass, though…

NSFW

Sometime soon I should generate fantasy location and costume prompts, and perhaps some bunnygirl-friendly locations for Easter. Some of the Christmas locations look vaguely Hyrule-ish, but mostly not.

more...

Done with Ruri’s adventures, I dug into my pre-Covid, pre-GenAI cheesecake archives, and deathmatched the gals I downloaded in May, 2019.

There is a lot of “Ai” in this set, but it’s autocompleted with “Shinozaki”, which is healthy and natural and good for growing boys.

more...

Or more precisely, “cheesecake by deathmatch”. Hey, I’ve got the silly thing, might as well use it to speed up processing my dusty pile of saved girlie-pics. While playing with “AI” tools, I downloaded Windsurf, ordered it to clean up the code for maintainability and then add a new function to export the currently-visible list to the clipboard (press L). Now I can quickly rank the images, use a flag to split them into NSFW and less-NSFW, and then toggle the flag to get two distinct export lists to pipe to my cheeseblogging prep scripts.

(Windsurf took me so seriously I ended up with a multi-file distribution package, but I had no difficulty reassembling it into a single file for simple downloading)

But first, a message from our sponsor, Qwen Image:

(the hardest part of this was getting it to render a “standard” onigiri rice ball; first it wrapped the entire thing in fresh green seaweed, then it made it huge, then it added additional wrapping on top of the nori strip, etc)

This is a selection from “stuff I downloaded in April, 2019”.

more...

No new anime, words that I’ll be saying for the next three months, so it’s time for more “AI” cheesecake. Maybe I’ll dig into my real cheesecake archives for contrast as well.

Today’s randomized babes are based on SFnal settings and costumes provided by Claude AI. It had no difficulty grasping the concepts of “vivid, exotic sci-fi locations” and “sexy retro-futuristic costumes for women”, generating detailed descriptions that Qwen Image was able to run with.

More importantly, it didn’t fall down on the unique part. It didn’t take forever to generate 100 results, and they were truly distinct, not just slight wording variations. Most importantly, it did not scold me for requesting “sexy”, or refuse to do my bidding. Although I haven’t asked for lingerie sets yet…

Qwen has recently released updated versions of some of their models, so I hope they get around to revving Image soon. I can cope with the usual finger-counting and the giantism, but I really hate seeing a great image ruined by a missing leg or wrong-side foot, something that’s really common.

more...