A high-rank fine art LoRA focused on academic marine painting, atmospheric lighting, and dramatic sea physics in the style of Ivan Aivazovsky. Built for deep dynamic range, authentic brushwork, and dual-light source contrast (moonlight vs. fire/sunlight).Training Specs &amp; Dataset<ul><li>Dataset Size: 350+ curated high-quality images.</li><li>Dataset Resolution: Native 512 x 512 crop tagging to force the model to learn macro-lighting, sfumato, and color temperatures instead of micro-noise.</li><li>Training Duration: 2,000 steps.</li><li>LoRA Rank: 128 (provides exact gesture, light dispersion, and structural control without over-fitting).</li></ul>Showcase Integrity (Pure Raw Generations)<ul><li>No Post-Processing: All preview images are 100% raw outputs straight from the pipeline. No inpainting, no Photoshop, no external upscaling, and no seed cherry-picking tricks.</li><li>Prompt Consistency: All showcase images were generated using a unified single-prompt approach to demonstrate true model versatility.</li></ul>Recommended Parameters<ul><li>LoRA Weight / Strength: <code>1.0</code> (Default)</li><li>CLIP Weight: <code>1.0</code> (Default)</li><li>Aspect Ratio: 1024 x1536 (2:3 aspect ratio provides the sweet spot for spatial attention density and vertical composition depth).</li><li>Step Count Advice: 50 – 60 steps.(Note: Showcase renders were experimentally pushed up to 65 steps to test latent limits. If you notice faint vertical sampling lines in high-frequency sky areas, drop steps down to 55-60 for clean, smooth gradients).</li><li>Workflow: Full ComfyUI workflow is embedded directly into the post file metadata.</li></ul>

AI-Vazovsky - Marine Fine Art LoRA [Rank 128]

NEW_2026_HIdreem.json

r128hdremazovski.safetensors

v1.0

A high-precision character model trained for 2,000 steps on a large-scale computing system. It is designed to ensure the absolute preservation of the character's recognizable appearance across various angles, close-ups, and complex lighting conditions. The images are published without post-processing.Key Features:<ul><li>Identity Preservation: Holds the exact same facial geometry across all generations without face-swaps or ControlNet.</li><li>Close-Up Precision: Flawless skin micro-texture, iris details, and lip highlights without plastic smoothing.</li><li>Lighting Versatility: Handles aggressive neon split lighting, dark backgrounds, and soft studio ambient lights seamlessly.</li></ul>Recommended Settings:<ul><li>Model Weight: 1.0</li><li>CLIP Weight: 1.0</li><li>CFG Scale: 3.5 – 5.0</li></ul>

Files with SHA256: e2f38c06fe7a9531b309498b3ba1ff907485b82f345062e7f558bb268574228a