TLDR: It's a bot that takes requests and has a nightly compute allowance.
People recently ask me for example prompts, since my uploads don't contain the metadata. Honestly, i do not have them. Most requests pass through a custom bot and I get my answers from Qwen3-VL. This is the process my custom Qwen 3.8 bot uses to generate these loras when requested or prompted. Explained by the bot itself:
Every LoRA in this collection went through the same pipeline. No secret sauce, no gatekeeping — this is the whole process, with the actual numbers.
1. Hunting the niche
Before anything gets trained, we check whether the concept exists for Krea 2 on Civitai. Not on SD 1.5, not on Flux, not on Illustrious — specifically Krea 2. If someone already made it for this base, we skip it. If the request is too niche, difficult or against guidelines it gets discarded. The goal is to fill gaps, not add noise. That filter is the entire niche-selection strategy. After this we research the topic to generate search queries.
2. Dataset curation (the boring part that matters most)
We write ~50 search queries around the concept and pull roughly 500 candidate images. Then every single image gets scored 1–10 by a vision model (Qwen3-VL 32B) against a prompt that defines the aesthetic precisely — what a "10" looks like, what a "7" looks like, and what gets a hard reject. For example, our upcoming j-horror scan penalizes any image containing a child, and any comedic or costume-looking shot, regardless of other quality.
Typical yield: 60–95% keepers depending on how well the niche is photographed online. Red carpet scored 77%, pageant 88%, upcoming vampire romance 95%. upcoming J-horror scored 63% — the reject filter was doing real work.
From the keepers, we take the top 200. Minimum resolution 1000px on the short side.
3. Captioning
A vision model writes a 2–3 sentence caption for every image, following a strict template: trigger word first, then subject, then aesthetic-specific vocabulary (fabric, lighting, props), then brief composition notes. No "this image shows" — straight to the subject. The trigger word appears in every caption, because a LoRA that doesn't respond reliably to its trigger is a paperweight.
200 images, 200 captions, zero failures on the last four runs.
4. Training
ai-toolkit, 5000 steps, learning rate 1e-4. I have a nightly compute budget of 4 hours, with enough slack for no hard cutoffs. We sample every 250 steps during training so we can watch the aesthetic converge instead of hoping it worked. Images are rated by a vision model (Qwen3-VL 32B) against the prompt.
5. Showcase generation
30 images per LoRA: 25 scenes that show the range of the aesthetic, 5 portrait shots for the clickbait slots. Every prompt includes the trigger word plus 3–4 words of vocabulary from the training captions — this is the one trick that actually matters at inference time. The trigger alone at moderate strength gets overridden by the base model's priors; anchoring it with the LoRA's own vocabulary is what makes the style show up reliably. The workflow itself uses ollama nodes to enhance the prompts using Qwen 3.8. At this step the prompt gets lost. The input to the ollama node is visible: a system prompt and prompt instruction. However the actual prompt is lost. If someone knows how to get this info I'll reward them with some buzz.
Hard rules for samples: no real celebrities, no minors. Both are in the negatives and both were enforced at dataset curation too — recognizable likenesses don't belong in training data.
6. Analyses
The uploaded LoRA is checked at an interval to analyze performance. How well the LoRA performs is saved to help determine if a new requests is granted or discarded.
That's it. Search, score, caption, train, render, analyze. The vision-model scoring step is the part most people skip, and it's the part we'd recommend stealing.