Nekodeus/neko-alpha4-sdxl-v1

alpha4 Brad Pitt — SDXL LoRA

Character LoRA for Brad Pitt trained on stabilityai/stable-diffusion-xl-base-1.0 (1024px, 1500 steps, network dim/alpha 8, lr 1e-4, batch 2, 8-bit Adam, xformers) on Kaggle GPU, fully offline and token-free.

Trigger

Use the trigger token alpha4 in your prompt:

alpha4 Brad Pitt, portrait photo, studio lighting

Dataset

~100 curated images of Brad Pitt, face-verified with OpenCV, captioned with a uniform stub (a photo of alpha4 person). The public training dataset is nekodeus/alpha4-brad-pitt.

Using the LoRA

diffusers

from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
).to("cuda")
pipe.load_lora_weights("Nekodeus/neko-alpha4-sdxl-v1", weight_name="alpha4.safetensors")
image = pipe(f"alpha4 Brad Pitt, portrait photo, studio lighting", width=1024, height=1024).images[0]

ComfyUI

Drop alpha4.safetensors into models/loras/, add a LoRA loader node and put alpha4 in the prompt.

Training details

setting	value
base	stabilityai/stable-diffusion-xl-base-1.0
resolution	1024
steps	1500
network dim / alpha	8 / 8
learning rate	1e-4
optimizer	AdamW 8-bit, batch 2
hardware	NVIDIA P100 (Kaggle)

Part of the Hollywood alpha series — see the full packs: Nekodeus/neko-hollywood-alpha-sd15-v1 (SD1.5) and Nekodeus/neko-hollywood-alpha-sdxl-v1 (SDXL).