---
base_model: Tongyi-MAI/Z-Image-Turbo
library_name: diffusers
pipeline_tag: text-to-image
tags:
  - lora
  - z-image-turbo
  - krea-2
  - comfyui
  - text-to-image
  - portrait
  - identity
  - diffusers
inference: false
---

# Kathimodel — Z-Image Turbo and Krea 2 LoRAs

Identity LoRAs for **[Tongyi-MAI/Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo)** and **Krea 2**. The files are base-model-specific and are not interchangeable.

## Trigger token

Use **`kathimodel`** in the prompt. Put it near the beginning of the subject description.

## Krea 2 — ComfyUI version

The new Krea 2 version was trained with a rebuilt 50-image dataset emphasizing detailed full-body and three-quarter-body views, supported by closer identity anchors. Backgrounds were blurred to reduce environment memorization.

Use this file only with a **Krea 2** workflow in ComfyUI:

- `krea2/kathimodel_krea2_fullbody_step_1200_comfyui.safetensors`

Recommended starting settings:

- Trigger: `kathimodel`
- LoRA strength: `0.8`
- Adjust within roughly `0.7–1.0`; higher values favor identity, while lower values allow more prompt flexibility.
- Describe the desired framing explicitly, such as `full-body fashion photograph`, `three-quarter portrait`, or `close portrait`.

Training details:

| Parameter | Value |
|---|---|
| Trainer | `fal-ai/krea-2-trainer` |
| Base model | `fal/krea-2` |
| Training steps | `1200` |
| Learning rate | `0.0003` |
| Resolution | `1024` |
| Training images | `50` |
| LoRA rank / alpha | `32 / 32` |
| Captioning | 50 paired captions; automatic captioning off |
| Output format | ComfyUI Krea 2 LoRA, `.safetensors` |

The ComfyUI conversion changed only tensor names. Its 528 learned tensors are byte-for-byte identical to the completed fal training artifact.

Krea 2 ComfyUI weight SHA-256:

```text
4490ba243b07e3ce40e85510b17d83196b70c805583ec22577bf62eb3d957743
```

The Krea 2 training manifest is available at `krea2/training_manifest.json`.

## Z-Image Turbo version

## Sample

![Sample generated with the final 1500-step LoRA](sample.png)

The sample above was generated from the final uploaded weights, without upscaling or face restoration.

| Setting | Value |
|---|---|
| Generator | `fal-ai/z-image/turbo/lora` |
| LoRA scale | `0.9` |
| Inference steps | `8` |
| Image size | `portrait_4_3` (768 × 1024) |
| Seed | `88417321` |
| Prompt expansion | Off |

**Sample prompt**

> Photorealistic close-up editorial portrait of kathimodel, an adult woman, head-and-shoulders framing, looking directly into the camera with a calm confident expression, long dark hair, simple black sleeveless top, soft natural window light, neutral warm gray studio background, realistic skin texture, sharp detailed eyes, 85mm portrait photography, shallow depth of field, natural proportions, no text, no watermark

## Recommended settings

- Start with a LoRA scale of **0.85–0.95**. Raise it toward `1.0` for stronger identity adherence; lower it toward `0.7` for more flexibility.
- Use **8 inference steps** with Z-Image Turbo.
- Close-up, head-and-shoulders, and waist-up compositions generally match the training distribution best.
- Mention `kathimodel` once and describe pose, wardrobe, lighting, and setting. Avoid over-specifying facial geometry, which can compete with the learned identity.
- For closer resemblance, use natural or soft studio lighting, direct or three-quarter gaze, and portrait focal-length language such as `85mm`.

## fal.ai example

Use the public Hugging Face weight URL as the LoRA `path`:

```json
{
  "prompt": "Photorealistic close-up portrait of kathimodel, direct gaze, long dark hair, soft natural light, detailed skin, 85mm photography",
  "image_size": "portrait_4_3",
  "num_inference_steps": 8,
  "seed": 88417321,
  "num_images": 1,
  "loras": [
    {
      "path": "https://huggingface.co/hodgy/kathimodel/resolve/main/kathimodel_z_image_turbo_closeup_step_1500.safetensors",
      "scale": 0.9
    }
  ],
  "output_format": "png",
  "enable_prompt_expansion": false
}
```

## Training details

| Parameter | Value |
|---|---|
| Trainer | `fal-ai/z-image-turbo-trainer-v2` |
| Base model | `Tongyi-MAI/Z-Image-Turbo` |
| Training steps | `1500` |
| Learning rate | `0.0005` |
| Training images | `50` |
| Captioning | One matching text caption per image |
| Trigger token | `kathimodel` |
| Training samples | Disabled |
| Output format | Diffusers LoRA, `.safetensors` |

The dataset was curated toward closer views of the adult subject. Images were cropped to increase facial and upper-body detail, and backgrounds were blurred to reduce environment memorization.

## Files

- `krea2/kathimodel_krea2_fullbody_step_1200_comfyui.safetensors` — Krea 2 LoRA with ComfyUI-compatible tensor keys
- `krea2/training_manifest.json` — Krea 2 training and conversion metadata with checksums
- `kathimodel_z_image_turbo_closeup_step_1500.safetensors` — final 1500-step LoRA weights
- `kathimodel_z_image_turbo_closeup_step_1500_config.json` — compact training configuration
- `sample.png` — generated example using the settings above

Weights SHA-256:

```text
a0135cfff0541fb969f3376f361acde52ee43488fcf65960a2da2bc035323275
```

## Limitations

- This is a small identity dataset, so results vary with framing, lighting, pose, and LoRA scale.
- The Z-Image version favors close portraits. The Krea 2 version includes more full-body coverage, but very distant subjects can still lose facial detail.
- Strong stylization, unusual angles, heavy occlusion, or prompts that impose conflicting facial features can reduce resemblance.
- The LoRA may reproduce biases or visual patterns present in the training photographs.
- A generated image is not a real photograph of the subject and should be represented as synthetic.

## Responsible use

Use only where you have the necessary rights and permission. Do not use this LoRA to impersonate, deceive, defame, harass, exploit, or violate the privacy or publicity rights of the depicted person. Clearly disclose AI-generated imagery when context could cause viewers to mistake it for a real event or endorsement.

## License

No separate license has been declared for this LoRA. The base model is released under Apache-2.0; users are responsible for complying with the base model terms and all applicable rights governing identity-based outputs.