import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("SG161222/RealVisXL_V5.0", torch_dtype=torch.bfloat16, device_map="cuda")
pipe.load_lora_weights("XaflocAI/BlondeJade-SDXL")
prompt = "blondejade, a woman in a white lace dress walking on a cobblestone street in Paris, golden hour, photorealistic, 85mm"
image = pipe(prompt).images[0]BlondeJade SDXL
An SDXL LoRA for the BlondeJade character — a consistent original AI persona. This is the SDXL successor to the earlier SD 1.5 version, retrained at 1024×1024 for sharper detail, better anatomy, and stronger prompt adherence.
Trigger word
Use the token blondejade in your prompt to invoke the character.
Usage
Load on top of any SDXL checkpoint. It was trained on RealVisXL V5.0, which is the recommended base for the most faithful results.
🧨 Diffusers
from diffusers import StableDiffusionXLPipeline
import torch
pipe = StableDiffusionXLPipeline.from_pretrained(
"SG161222/RealVisXL_V5.0",
torch_dtype=torch.float16,
).to("cuda")
pipe.load_lora_weights("XaflocAI/BlondeJade-SDXL", weight_name="blondejade_sdxl_v1.safetensors")
prompt = "blondejade, a woman in a white lace dress on a Paris street, golden hour, photorealistic, 85mm"
image = pipe(prompt, num_inference_steps=30, guidance_scale=5.0).images[0]
image.save("blondejade.png")
ComfyUI / Automatic1111 / Forge
Place blondejade_sdxl_v1.safetensors in your models/Lora folder and add
<lora:blondejade_sdxl_v1:0.8> (or the equivalent LoRA loader node) to your prompt
along with the trigger word blondejade. A weight of 0.7–0.9 works well.
Recommended settings
| Setting | Value |
|---|---|
| Base model | RealVisXL V5.0 (or any SDXL checkpoint) |
| Resolution | 1024×1024 (and SDXL-native buckets) |
| LoRA weight | 0.7 – 0.9 |
| Sampler / steps | DPM++ 2M Karras, 25–35 steps |
| CFG scale | 4 – 6 |
| Clip skip | 2 |
Training details
| Base model | RealVisXL V5.0 (SDXL 1.0) |
| Network | LoRA (networks.lora) |
| Rank / Alpha | 32 / 16 |
| Optimizer | Prodigy (LR 1.0, cosine, 100 warmup steps) |
| Resolution | 1024×1024, aspect-ratio bucketing (512–1536) |
| Dataset | 168 images (156 hi-res + 12 lo-res), DreamBooth-style |
| Steps | 4000 (epoch 5) |
| Precision | bf16, clip skip 2 |
| Trainer | kohya-ss sd-scripts |
License
Released under the CreativeML Open RAIL++-M license. You are responsible for the content you generate; do not use this model to create unlawful, harmful, or non-consensual content, or to depict or impersonate real individuals without their consent.
- Downloads last month
- 14
Model tree for XaflocAI/BlondeJade-SDXL
Base model
SG161222/RealVisXL_V5.0