Upcycled cheesecake

Sun 6/28/26 11:16am

Comments: 0

Cheesecake

GenAI: elephants all the way down

Forget “vibe coding” and “prompt engineering”; the new hotness is “loop engineering”, in which you tell your Agent to write the prompts that write the prompts that write your code.

🎶🎶🎶🎶
AI write the prompts that make the GenAi sing,
AI write the prompts to exfiltrate all my things,
AI write the prompts that make IT guys cry,
AI write the prompts that write the prompts.
🎶🎶🎶🎶

Krea2 Gone Wild

Unlike most image-gen models, Krea2’s training included explicit sexual material, and just censored Those Words in the text-processing layers. Which means that people didn’t have to train LoRAs on their own collections of nekkid people, and just had to diddle a few matrices (Juan’s extension, Silvercoin’s LoRA; use the LoRA).

The upside of this is that enabling NSFW details doesn’t inject a lot of possibly-incompatible data into the model. Most NSFW LoRAs change faces and increase the frequency of broken anatomy, but Krea2 remains stable. The downside is that it doesn’t know all the words; you have to be a bit clinical in describing the naughty bits.

Related, my image-analysis script has been running in the background, generating pretty detailed descriptions of 75,000+ images from my blog, laptop, and NAS. The LLM wasn’t 100% consistent in naming the categories, but it was good enough that I was able to clean them up in a script and generate well-structured JSON.

The primary limitation to the dataset is that most of the women are Japanese, but another little script allows me to rewrite each entry so I can change the art style, setting, race, hair, etc. For instance, this recipe changes 18-year-old Japanese girls into grown-up Irish redheads and gives each one a different art style:

{
    "style": "shuf:styles",
    "age": "32 years old",
    "sex": "female",
    "ethnicity": "Irish",
    "skin_color": "pale and lightly freckled",
    "eye_color": "light-green",
    "looks": "extremely pretty face",
    "complexion": "lightly-freckled",
    "figure": "perky medium-sized breasts, small waist, slender hips, firm round buttocks, flat stomach",
    "expression": "friendly smile with visible teeth",
    "hair_color": "rich red"
}

Haven’t decided yet if the script should be called upcycle or reskin. 😁

Krea2 doesn’t always nail the “age” field, but I tried it with a variety of expressions starting at “18 or older”. 😁