Z-image v20 (Experimental Merge)
Studio: Orakul Studio
About this model
Experimental test merge based on 200 high-quality images. The focus here is on high-fidelity representation of the artist's style and gesture.
Technical Specifications
- Resolution: 1024x1024
- Rank: 512
- Methodology: Full 8-bit stack, no quantization, bf16 precision.
- Hardware Stack: Custom async CUDA memory manager (RTX 4090, 128GB RAM).
Why Rank 512?
The core philosophy behind using Rank 512 is to move beyond simple pixel memorization. As we explored in Orakul Studio: Rank 32 learns style, Rank 128 learns technique, but Rank 512 learns the hand. This approach focuses on capturing the artist's gesture, ensuring the model learns the "hand" of the artwork rather than just copying pixels.
Showcase
π Installation & Usage Guide
This release is a ready-to-use baked monolith in pure BF16 format. It is a standalone checkpoint (Merge), NOT a LoRA. To ensure the model works correctly and achieves the specified quality, strictly follow the instructions below.
π Where to Place the File
- ComfyUI / Forge / reForge: Place the downloaded file strictly in the
models/diffusion_models/folder. - β οΈ CRITICAL: Do NOT place it in
models/checkpoints/ormodels/loras/.
βοΈ How to Use
- In your workflow, select this file via the UNet/DiT model loading node (e.g.,
Load Diffusion Modelin ComfyUI). - The model operates as a standalone, independent engine. No additional LoRA nodes are required to invoke this style.
π Mandatory Trigger Word
Even though the style is fully integrated into the checkpoint weights, you MUST use the trigger word at the very beginning of your prompt:
PenZlTr512
π‘ Technical Note: This token is the key that activates the custom rank 512 structure. Without this trigger, the model will output standard graphics and ignore the unique pencil lead texture, search lines, and custom Orakul Studio signature.
π¨ Inside the Tech: Dynamic Alpha Baking
This is the official merge of our high-ranking LoRA directly into the base checkpoint. Using a custom protocol with dynamic alpha, we were able to bake a live pencil texture without destroying the geometry.
This version is designed for those who want a refined, soft graphite stroke (HB/2B) with perfect airy shading, completely eliminating the heavy "marker" burnout and inky blackness.
β οΈ Strict Generation Rules
To ensure the model delivers the advertised analog quality and a soft gray tone, use only these settings:
| Parameter | Optimal Value | Description |
| :--- | :--- | :--- |
| **Sampling Steps** | `20` | *Strictly 20. No more, no less. This is the ideal diffusion point, where the dark fractions don't have time to turn into a dull ink.* |
| **CFG Scale** | `2.5` | *Critically important! Don't set it higherβit will burn the midtones.* |
| **Sampler** | `UniPC` | *Recommended by the author for perfect line accuracy.* |
| **Scheduler** | `G Uniform` | *Ensures a soft noise distribution.* |
| **Denoise** | `0.95` | *Optimal balance for revealing hidden graphics layers.* |
---
### π Official Repository
For updates and core development info, follow us on GitHub:
π **[Orakul Studio GitHub](https://github.com/OrakulStudio)**
---
## π Citation
```json
@article{team2025zimage,
title={Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer},
author={Z-Image Team},
journal={arXiv preprint arXiv:2511.22699},
year={2025}
}









