After training, I manually fine-tuned some of the LoRA layers to make images look more realistic, improve small details, and reduce safety refusals. In most cases, it works without any extra bypass nodes, but depending on your prompt, you may still see an occasional safety trigger.

Training Dataset

This LoRA was trained on 4,061 carefully selected images taken with DSLR cameras, mirrorless cameras, and smartphones.

The dataset includes many different styles and subjects, including:

*Realism

* Horror

* Cosplay

* Close-up skin details

* Dynamic and difficult poses

* Graphic posters

* Many other image styles

As part of this training, I also created a special dataset of around 100 RAW camera images. Each image was manually color graded using different cinematic 3D LUTs, then individually captioned to describe the specific color grading and look. This helped teach the LoRA different cinematic color styles and improved its understanding of color, mood, and lighting.

Recommended Settings

Trigger Word: None

Strength (used on its own): 0.8–1.0

Strength (mixed with character LoRAs): 0.3–0.5 to help keep character identity clean and reduce concept bleeding.

Workflow: https://civarchive.com/models/2608465/qwen-and-krea-2-sam3-face-inpainting-workflow

Training Information

Dataset: 4,061 images

Resolution: 1024

Network Rank: 128

Note: Sample images were generated using the full-precision FP32 version of the LoRA. A BF16 version is also included, some example prompts were taken from bananaprompts.xyz for testing.