lxe/holosomnia-zimage-turbo-lora

Holosomnia Z-Image Turbo LoRA

Style LoRA trained for Tongyi-MAI/Z-Image-Turbo on original Holosomnia artwork.

This upload contains the V10 "content moderate" checkpoint at step 800, which was selected because it showed strong style transfer while preserving more subject/detail structure than later, more overcooked checkpoints.

Strength Sweep Preview

The grid below compares the same five prompts/seeds across LoRA strengths 0.1, 0.2, 0.6, 0.9, and 1.1. In this sweep, 0.9 was the strongest general-purpose balance; 1.1 pushes the look harder but can simplify or over-saturate structure.

Files

• holosomnia_zimage_turbo_v10_step0800.safetensors - LoRA checkpoint
• training_config_v10_content_moderate.yaml - training configuration used for this run
• holosomnia_zimage_turbo_v10_step0800.sha256 - SHA256 checksum
• training_metadata.json - compact metadata summary
• images/holosomnia_lora_strength_sweep_contact_sheet.png - LoRA strength comparison sheet

Usage

Use with Z-Image Turbo as a LoRA.

Recommended starting points:

• LoRA strength: 0.6 to 1.0
• Lower strength preserves the base prompt/content more.
• Higher strength pushes the Holosomnia cloud/color/rendering style harder.
• Trigger keyword: zhlxart

Example prompts:

zhlxart, a lighthouse on a cliff above the ocean at sunset with dramatic clouds

zhlxart, a small white cottage on a rolling green hill beneath towering colorful clouds

zhlxart, an ultra detailed portrait of a woman with silver hair wearing embroidered clothing, soft skin texture, intricate jewelry, shallow depth of field

The trigger can be omitted for subtler use, but zhlxart gives the most direct activation.

Training Summary

Setting	Value
Base model	Tongyi-MAI/Z-Image-Turbo
Training adapter	ostris/zimage_turbo_training_adapter/zimage_turbo_training_adapter_v2.safetensors
Training tool	ai-toolkit 0.9.6
Checkpoint	V10 content moderate, step 800
Dataset	45 curated Holosomnia image/caption pairs
Captions	zhlxart trigger plus content-preserving descriptions
LoRA rank	32
LoRA alpha	32
Network target	Transformer only
Text encoder	Not trained
Optimizer	adamw8bit
Learning rate	8e-5
Batch size	1
Gradient accumulation	1
Training dtype	bf16
Save dtype	float16
Noise scheduler	flowmatch
Timestep type	sigmoid
Differential guidance	Enabled, scale 2.0
Resolution buckets	512, 768, 1024
Caption dropout	0.0

Checksum

08b6aa34bf46a8fce20c02726878a007637b576ca7072f93f9fd7c84f19b57a3  holosomnia_zimage_turbo_v10_step0800.safetensors

Rights

The training images are original artwork by Holosomnia and are not included in this repository.

This LoRA is provided for use with Z-Image Turbo. Users are responsible for following the license and terms of the base model and any other components used in their inference workflow. No license grant is made here for the original Holosomnia artwork or training images.

Limitations

This is not a standalone model. It requires a compatible Z-Image Turbo workflow. Like most style LoRAs, high LoRA strength may shift composition, color balance, or facial/detail fidelity; reduce strength when content preservation matters.