rakshazi / krea2-prompting

# krea2-prompting
I can run Krea 2 Turbo on my local hardware now, so I did the obvious thing: spent a stupid amount of time generating images and writing down what actually happened.
I ran it locally because paying a cloud provider by the hour to find out whether one adjective changes the anatomy is a bad business model.
This repo is the result. It is an agent skill for Krea 2 Turbo in ComfyUI: prompts, settings, pose blocks, style names, and a pile of "this wording works, this wording produces tasteful garbage" notes.
Prompting only. No models, no weights, no ComfyUI graph. Bring your own checkpoint.
To be clear: my day job is Go backend development. I don't have a Hugo Award in Prompt Engineering, and I'm not trying to win one. I just want the image to match the prompt.
## The take
Krea 2's text encoder is Qwen3-VL, so it reads sentences like a language model. Most prompt guides still hand you `masterpiece, best quality, 8k, ultra detailed`, then act surprised when the output looks like a stock photo from 2023.
This skill goes the other way: write what the camera sees, name the light, name the material, and stop stacking adjectives.
## About the AI thing
Yes, an AI wrote most of this. No, that doesn't make it another confident pile of prompt folklore. Fair question, though.
**Every "tested" claim comes from my own runs on my own hardware.** I ran Krea 2 Turbo locally, generated the images, compared seeds, and wrote down what broke. The skill documents actual behavior. A model's idea of what should happen doesn't make the cut. If something is untested, it says so. If you find a claim that smells like a hallucination, open an issue and bring a seed.
## How the tests worked
Same checkpoint, same seed, one variable per prompt. If a clause changed the image, it stayed. If the model ignored it, it went into the failure table with the exact wording that failed. That's the whole method. (Yes, it is slower than asking a chatbot. That's the point.) No vibes, no "I asked another model and it agreed with me".
## Examples
Ten topics, two passes by DeepSeek v4.1 Flash: one with only this README and the list of files as context, one with the entire skill loaded. The examples gallery has both renders side by side and both prompts for every topic. It is the fastest way to see what the skill actually changes.
## What's inside
Everything except `SKILL.md` lives in `references/`.
| File | What it does |
|---|---|
| `SKILL.md` | The entrypoint. Mental model, ComfyUI defaults, prompt order, core rules. Start here. |
| `photoreal.md` | Camera, lens, film, light, texture. Portraits, products, landscapes, architecture, heroes. |
| `krea2-core.md` | Deep architecture, negative families, poster text, composition modules, resolution table, reference reverse-engineering, character consistency. |
| `comfyui-setup.md` | Pipeline, the CFG 1.0 trap, `Krea2PromptWeight`, the enhancer, troubleshooting. |
| `poses.md` | The 11-field pose schema, usage, taxonomy, and the field fi