Bad Mood Salon

作者：privatewolf

Bad Mood Salon | Krea 2

Goth and punk women in thick, ridged paint, with faces pulled as far as they'll go. Snarls, wails, winks, sulks, eye-rolls and screams of laughter. The faces are caricatures: big eyes with heavy lids and dark lashes, small noses, full lips, and ears that stick out. Hair comes in teal, pink, scarlet, platinum and black, usually as a blunt bob, space buns or a mess. They wear spiked collars, O-ring chokers, lace, fur stoles and plenty of piercings. Backgrounds are mostly plain and warm (burnt orange, mustard, ochre), with slate blue, crimson, olive and plum for contrast. Now and then an elf, a cat-eared girl or a green-skinned monster turns up. Trigger word: 3xpr1x

How to use it

Start every prompt with 3xpr1x, and write the rest the way the training captions read. Each caption is one plain paragraph of content: the framing, the face feature by feature, the expression, the hair, the clothes and jewellery, then the background colour. None of them has style or medium words. Leave out "oil painting", "impasto", "textured" and "illustration". The trigger carries the paint, and those words pull Krea toward its own smooth version. Template:

3xpr1x, a [close-up / head-and-shoulders portrait / waist-up view] of a young woman in her twenties [DOING WHAT]. She has a [FACE SHAPE] face, [EYE COLOUR] eyes [LIDS OR LASHES], [BROWS], a [NOSE] and [LIPS] in [A STRONG EXPRESSION]. Her hair is [COLOUR AND CUT]. She wears [COLLAR OR JEWELLERY] and [TOP]. Plain [COLOUR] background.

Example:

3xpr1x, a close-up of a young woman in her twenties with large pointed black cat ears ringed with silver hoops, staring cross-eyed at a small red laser dot on the tip of her nose. She has a heart-shaped face, round green eyes with slit pupils, thin arched brows, a small upturned nose and a small open mouth. She raises both hands near her chin with her fingers curled like paws, ready to swat. Her hair is a choppy lilac bob with a short blunt fringe. She wears a black mesh top and a studded collar with a silver bell. Plain slate-blue background.

Settings

Krea 2 Turbo: weight 1.0, 8 steps, guidance 1. If a render comes out smooth, glossy or photographic, raise the weight to 1.2 on the same seed before changing the prompt. Krea 2 Base: start at 0.8 to 1.0. Most of the set is 4:5, so 1024 × 1280 suits single portraits. Group shots in the set are landscape, so use 1344 × 896 for two to four figures. 2:3 at 832 × 1248 works for waist-up figures with something tall in the frame.

Tips

• Push the expression. Every training face is pulling one, so "smiling" gives you less than the LoRA can do. Write what the face is doing: "brows furrowed, lips pulled back over clenched teeth", "eyes squeezed shut, mouth wide open laughing".
• Give each face its own quirk in the brows, nose or mouth, and pick the hair colour yourself. Several faces in the set share a black blunt bob with yellow or red eyes, and faces drift toward that look when the prompt doesn't say otherwise.
• Stay close in. The training pictures run from close-ups down to waist-up, and none of them is full length. Full figures and wide scenes are outside what it learned.
• For groups, press the figures together cheek to cheek or shoulder to shoulder, the way the set's group pictures do, and describe each one left to right.
• To get the cream paper border from some of the set, ask for it: "inside a cream paper border". It was captioned, so it won't show up on its own.
• Count the hands and give each one a job, like holding a cup, covering a mouth or resting under the chin, so none of them floats.
• Krea ignores negative prompts at guidance 1. To keep something out, describe what should be there instead.

Training

23 hand-picked images, each captioned by hand with content only: framing, faces, expressions, hair, clothes and background colour, with no style or medium words. 15 are single figures (one of them a three-pose sheet), three are pairs, three are trios and two are groups of four. Trained on Krea 2 Base with Civitai's trainer: 1,840 steps over 10 epochs (about 80 passes per image), dim 32, alpha 32, UNet LR 1e-4, constant scheduler, AdamW8bit, flip augmentation off. Released at epoch 10