Krea 2 DreamBooth LoRA - Ghaxalll/ghaxal-face

A face LoRA of Ghazal, trained on 21 original iPhone photos (7.2 to 12.2 MP, straight from the camera roll, never through a messenger).

Trigger word

Use ghaxal_woman in every prompt.

The auto-generated card first said TOK. That was a placeholder left by the training script, not the real trigger. TOK will not produce this face.

What the captions deliberately leave out

The hair (short straight bob) is never named in any training caption, so it is baked into the identity and arrives with the trigger word.

These were captioned, so they are optional and can be prompted away: red lanyard, sunglasses resting on the head, large black bead necklace, backpack straps, wireless earbud, and every location and lighting condition.

Checkpoints

Path	Steps
pytorch_lora_weights.safetensors	1000 (final)
checkpoint-1000/	1000
checkpoint-500/	500

If the final weights look over-baked, try checkpoint-500 before retraining.

Ignore the sample images

sample_0.png, sample_1.png and sample_2.png show a white tiger, a library interior and another generic scene. They came from the post-training showcase step, which generated style-type prompts that never contained ghaxal_woman. They say nothing about this model, good or bad.

Use it with the diffusers library

import torch
from diffusers import Krea2Pipeline

pipe = Krea2Pipeline.from_pretrained("krea/Krea-2-Turbo", torch_dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights("Ghaxalll/ghaxal-face")

prompt = "ghaxal_woman, sitting at a desk by a window, soft afternoon light, 85mm"
image = pipe(prompt, num_inference_steps=8, guidance_scale=0.0).images[0]
image.save("output.png")

Training details

Setting	Value
Base	krea/Krea-2-Raw
Method	DreamBooth LoRA
Images	21
Steps	1000
Resolution	1024
Dataset	Ghaxalll/ghaxal-face-dataset

Limitations

Strong on head-and-shoulders framing; weaker at full body and true profile, because the training set contained neither. All but one training image was a selfie, so camera distance barely varies.