The LoRA was trained using AI-Toolkit with a high rank of 128, resulting in a file size of approximately 1.3 GB. This unusually high rank was intentionally chosen to prioritize image sharpness, texture richness, and fine-detail retention, at the cost of increased model size.

Training Strategy

The dataset was built using a dual-resolution approach:

Full-frame images (1024×1024)
Original high-resolution images were resized entirely to 1024×1024 and tagged with the trigger h1r3s_image. These samples focus on global composition, proportions, and overall realism.

High-resolution tiles (1024×1024)
Original images larger than 1024×1024 were split into multiple 1024×1024 tiles and tagged with h1r3s_detail. These samples emphasize localized regions, micro-textures, and fine visual details.

This mixed-scale dataset allows the LoRA to learn both structural coherence and high-frequency detail.

How to Use

• Use h1r3s_image for full compositions and complete subjects.

• Use h1r3s_detail to emphasize texture, sharpness, and micro-detail.

• Recommended resolutions: 1024×1024 and 1536×1536

• Best suited for high-resolution outputs.

Notes / Observations

This LoRA is an experimental project exploring high-rank training and mixed-scale datasets.

In some cases, the LoRA may produce images that appear slightly warmer in color tone and more polished or post-processed compared to the base model. This does not represent a significant loss of realism at a high level, but rather a shift in visual character.

Due to the strong emphasis on sharpness and micro-detail, the output often contains very low noise, resulting in cleaner surfaces, smoother gradients, and higher pixel fidelity. While this can reduce the presence of natural noise that sometimes contributes to perceived realism, it also delivers exceptionally crisp details and minimal artifacts, especially at higher resolutions.

This trade-off is intentional and aligns with the goal of maximizing image clarity and texture quality. Users seeking additional realism may optionally reintroduce subtle grain or noise in post-processing.

Results may vary depending on prompts, samplers, and seeds. Feedback and experimentation are encouraged.