import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Tongyi-MAI/Z-Image-Turbo", dtype=torch.bfloat16, device_map="cuda")
pipe.load_lora_weights("hodgy/elle-z-image-lora")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Elle — Z-Image Turbo LoRA
An identity/character LoRA for Tongyi-MAI/Z-Image-Turbo, trained with Ostris AI Toolkit.
Suggested trigger word
Start prompts with elle. This word is inferred from the repository and filename; the embedded metadata does not contain the original captions, so the exact trigger token could not be independently verified.
Example prompt:
Photorealistic portrait of elle, direct gaze, natural window light, realistic skin texture, 85mm photography.
Training metadata
| Parameter | Value |
|---|---|
| Base family | Z-Image |
| Training steps | 1000 |
| Recorded epoch | 9 |
| Trainer | ostris/ai-toolkit 0.7.7 |
| Tensor count | 480 |
Recommended settings
- Base model:
Tongyi-MAI/Z-Image-Turbo - Start with LoRA strength
0.8–1.0. - Use 8 inference steps.
- Describe framing, pose, clothing, lighting, and setting explicitly.
fal.ai example
{
"prompt": "Photorealistic portrait of elle, natural light, realistic skin texture, 85mm photography",
"image_size": "portrait_4_3",
"num_inference_steps": 8,
"loras": [{
"path": "https://huggingface.co/hodgy/elle-z-image-lora/resolve/main/elle_z-image_lora.safetensors",
"scale": 0.9
}]
}
File
elle_z-image_lora.safetensors
Limitations and responsible use
The dataset, learning rate, and caption token are not included. Identity consistency can decrease in extreme profiles, distant framing, occlusion, or strong styles. Use only with appropriate rights and consent; never use for impersonation, deception, harassment, exploitation, or privacy violations. Clearly disclose synthetic imagery when appropriate.
- Downloads last month
- 29
Model tree for hodgy/elle-z-image-lora
Base model
Tongyi-MAI/Z-Image-Turbo