---
license: other
license_name: flux-kommunity-license
license_link: https://huggingface.co/black-forest-labs/flux-3-image-klein-base/blob/main/LICENSE.md
base_model: black-forest-labs/flux-3-image-klein-base
pipeline_tag: image-to-image
library_name: diffusers
tags:
- lora
- flux
- flux-3
- klein
- image-to-image
- skin
- texture
- high-pass
- restoration
- detail-enhancement
widget:
- text: "Case 01 · Close-up skin"
  output:
    url: assets/card_01_closeup.jpg
- text: "Case 02 · Portrait"
  output:
    url: assets/card_02_portrait.jpg
- text: "Case 03 · High-pass layer"
  output:
    url: assets/card_03_highpass.jpg
---

# Hires Skin Texture · FLUX 3 Klein LoRA

**A LoRA trained to restore high-pass skin detail lost to repeated VAE compression.**

It works on the high-frequency layer only — pores, fine lines, peach fuzz, micro-specular — so it brings back texture without touching the color, lighting or composition of your image.

![Case 01 · Close-up skin animation](assets/case.avif)


![Case 01 · Close-up skin](assets/card_01_closeup.jpg)

---

## Before / After

<table>
  <tr>
    <td width="50%"><img src="assets/card_01_closeup.jpg" alt="Case 01 close-up skin before and after"/></td>
    <td width="50%"><img src="assets/card_02_portrait.jpg" alt="Case 02 portrait before and after"/></td>
  </tr>
  <tr>
    <td align="center"><b>Close-up skin</b><br/>Pores and specular detail come back on the nose and under-eye area.</td>
    <td align="center"><b>Portrait</b><br/>Wrinkles, stubble and skin grain recovered on a mid-shot.</td>
  </tr>
</table>

![Case 03 · High-pass layer](assets/card_03_highpass.jpg)

*Case 03 shows what the model actually works on: the isolated high-pass layer, before and after.*

---

## Why this project?

Producing UGC-style renders means many back-and-forth rounds with clients — adding and removing elements from the frame. Every edit sends the image through another VAE encode/decode, whether you use local models or closed-source ones. Each pass destroys a little more of the fine detail, and after a few rounds the skin looks waxy and smoothed out.

This LoRA is the last step of that pipeline: it puts the texture back.

## Training method

The dataset is built from high-quality skin texture images, which are then degraded by passing them through **multiple VAEs**. This reproduces the real-world damage of an iterative editing workflow.

Since we only want to train texture, we **separate the high-pass detail** of both the input (degraded) and the target (clean) images. The model learns to map a low-quality high-pass skin layer to a high-quality one.

Instead of a classic img2img workflow, this abstraction:

- keeps the model's attention on fine detail rather than on global content;
- **avoids color shift** in the outputs, because the low-frequency layer (color, tone, lighting) never goes through the model.

| | |
|---|---|
| Base model | FLUX 3 Klein (`black-forest-labs/flux-3-image-klein-base`) |
| Type | LoRA |
| Input | High-pass layer of a VAE-degraded image |
| Target | High-pass layer of the clean source image |
| Degradation | Multiple VAE encode/decode passes |
| Rank / alpha | `32` |
| Steps / LR | `2000` |
| Resolution | `2048` |

## How to use

1. **Split** your image into a low-pass layer (Gaussian blur) and a high-pass layer
2. **Run** the high-pass layer through FLUX 3 Klein with this LoRA (image-to-image / edit). Trigger / prompt: `Restore Skin`, you can also use the given embeddings "restore_skin.safetensors".
3. **Recombine** the restored high-pass with the original low-pass.
4. 
> A ComfyUI workflow is included: `restore_skin_workflow.json`

## Why FLUX 3 Klein?

- **Pixel-perfect output.** The same training on FLUX 2 Klein (4B and 9B) produced pixel shift; FLUX 3 Klein stays aligned with the input, which is essential when the result is blended back onto the original.
- **Best VAE for high-pass detail.** The FLUX 2 VAE is, to date, the best available VAE for preserving high-frequency detail.

### What about Bria 2?

Training started before Bria 2 was released, so it hasn't been tested. It could be explored, but its licensing seems more restrictive.

## Limitations

- Trained on skin only — hair, fabric and backgrounds are not targeted and may receive limited or unwanted texture.
- Works best on images degraded by VAE round-trips; heavy JPEG blocking or strong motion blur are out of scope.

## License

This LoRA is a derivative of FLUX 3 Klein and is distributed under the [FLUX Kommunity License](https://huggingface.co/black-forest-labs/flux-3-image-klein-base/blob/main/LICENSE.md). Check the base model's terms before any commercial use.