readme.md

---
base_model: Tongyi-MAI/Z-Image-Turbo
tags:
  - lora
  - text-to-image
  - z-image-turbo
  - style
  - diffusion
license: other
---

# a-cold-wall — LoRA

A LoRA adapter trained for the concept/style **"a-cold-wall"**.

## Trigger word
Use this token in your prompt:
- **`a-cold-wall`**

## Base model
- **Tongyi-MAI/Z-Image-Turbo**

## Files
- `a-cold-wall.safetensors` — the LoRA weights
- `config.yaml`, `job_config.json` — training configuration (for reproducibility)

## How to use

### A) ComfyUI / AUTOMATIC1111
1. Put `a-cold-wall.safetensors` into your LoRA folder.
2. Use it in your prompt, e.g.:
   - `a-cold-wall, fashion outfits, editorial photo, high detail`

(Adjust LoRA strength to taste, e.g. 0.6–1.0.)

### B) Diffusers (generic example)
> Depending on your setup, you may need to use the correct pipeline class for Z-Image-Turbo.

```python
import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.bfloat16
).to("cuda")

pipe.load_lora_weights("thorjank/a-cold-wall-lora", weight_name="a-cold-wall.safetensors")

prompt = "a-cold-wall, fashion outfits, editorial photo, high detail"
image = pipe(prompt).images[0]
image.save("out.png")

Recommended prompting
	•	Start simple:
	•	a-cold-wall, fashion outfits
	•	Then add camera / lighting / composition as needed.

Training details (summary)
	•	Dataset: 35 images
	•	Steps: 3000
	•	Batch size: 1
	•	Learning rate: 1e-4
	•	Network: LoRA
	•	linear rank/alpha: 32/32
	•	conv rank/alpha: 16/16
	•	Trained modules: U-Net (text encoder not trained)
	•	Precision: bf16
	•	Noise scheduler: flowmatch
	•	Resolution buckets configured: 512 / 768 / 1024
	•	Default caption: fashion outfits

Notes / License

This repo contains only LoRA weights. Please ensure your use complies with the base model’s license and with the rights for any content you generate.