Trigger word : Make the subject's skin details more prominent and natural

The Qwen Image Edit 2509 version has gained immense popularity, even surpassing Flux on Civitai as the best recent open-source AI art model.

However, while the native Qwen Image model is great, it has a fatal flaw: the skin is overly smoothed and refined.

When it comes to AI painting, Chinese users and Western users pursue different directions—Chinese users prefer a maxed-out beauty filter effect, while Western users prefer heavy skin grain and texture.

To reduce the over-beautified effect of Qwen Image, someone developed a skin LoRA to achieve more realistic skin effects.

Fine-tuned LoRA: Enhancing Skin Realism in Qwen-Image-Edit-2509

This repository contains a fine-tuned Low-Rank Adaptation (LoRA) model designed to enhance the realism and detail of human skin in images. Trained on the powerful Qwen/Qwen-Image-Edit-2509 base model, this LoRA leverages its advanced image editing capabilities, focusing on generating more natural and detailed skin textures.

The model was trained for 5,000 steps on a local RTX 5090 using AI-Toolkit. The resulting LoRA is ideal for photographers, digital artists, and anyone looking to improve the quality of human subjects in generated or edited images.

Model Description The qwen-edit-skin LoRA is a specialized fine-tune of the Qwen/Qwen-Image-Edit-2509 base model. The base model is a versatile image editor with strong capabilities in multi-image editing and single-image consistency preservation, particularly in maintaining personal identity. Building upon this foundation, this LoRA focuses on the nuances of human skin, adding details and realism that might be missing in the original generations.

Training was conducted using a fork of AI-Toolkit, a comprehensive suite for fine-tuning diffusion models. The dataset curation process was as follows:

• Select real human portrait images with exposed skin.
• Label these images as "target" (THE AFTER) images—the desired final output in the standard Qwen editing workflow.
• Edit the images in Photoshop by adding Gaussian blur to smooth out skin tones and reduce the visibility of skin texture, tone variations, and pores.
• These blurred images serve as the "control" (THE BEFORE) images for the Qwen editing training.

Training Details The model was fine-tuned using the following key parameters, which can be found in the accompanying config.yaml file:

Hardware:

• GPU: NVIDIA RTX 5090

Training Configuration:

• Training steps: 5000
• Batch size: 1
• Gradient accumulation: 1
• Learning rate: 1.0e-04
• Optimizer: adamw8bit
• Noise scheduler: flowmatch
• Resolution: Trained on datasets at 512, 768, and 1024 pixel resolutions.
• Precision: bf16

Network Architecture:

• Type: LoRA
• Linear rank and alpha: 16
• Convolutional rank and alpha: 16

Selecting adamw8bit as the optimizer is significant because it reduces the memory footprint during training, enabling efficient fine-tuning on consumer-grade hardware without sacrificing performance. The flowmatch noise scheduler is a modern approach that achieves more efficient training and high-quality image generation.

A key characteristic of this LoRA architecture is that the alpha values for both linear and convolutional layers are equal to their respective ranks (16). This balanced approach is a common starting point in LoRA training, ensuring that learned adaptations are applied with a proportional scaling factor. This helps prevent overfitting while enabling the model to effectively learn the desired new features.

How to Use To use this LoRA, load the base model Qwen/Qwen-Image-Edit-2509, then apply the fine-tuned LoRA weights loaded as qwen-edit-skin.safetensors. Previous version weights have been uploaded to the repository for reference, but the final version is qwen-edit-skin.safetensors. You can also utilize the example ComfyUI workflow included in the repository to compare results across different weights.

The recommended weight is between 1 and 1.5. The provided examples show weights up to 2 merely to demonstrate the effect of an overly strong LoRA.

Intended Use This LoRA is intended for creative and artistic purposes to enhance the realism of human skin in digital images. It can be used by:

• Digital artists: Adding finer details and textures to character skin.
• Photographers: Retouching and enhancing portraits.
• AI art enthusiasts: Generating more photorealistic human images.

Limitations and Biases This model is a fine-tune of a large-scale pre-trained model and may carry some of its inherent biases. The training dataset for this LoRA focuses on improving skin details and may not equally represent the full diversity of human skin tones and types. Users should be aware of this and use the model responsibly. Model outputs are influenced by input prompts; users are encouraged to use descriptive and inclusive language to guide the generation process.

Disclaimer: This model is intended for artistic and creative purposes. Users are responsible for the content they create and should adhere to ethical guidelines, respecting individual privacy and dignity.

Trigger Word You should use "Make the subject's skin details more prominent and natural" to trigger image generation.

It can be observed that the higher the LoRA weight, the stronger the skin grain (realism) in the image.