🎨 Orakul Studio: Ultra-High Rank FLUX.2D LoRA Collection

Museum-Quality Art & Physical Texture Reproduction | Unquantized, No Compromise

Welcome to the official repository of Orakul Studio (Chernihiv, Ukraine 🇺🇦).

This collection contains extreme high-rank LoRA adapters (ranging from r1280 up to r2048) trained on top of 14B parameter FLUX models. Unlike standard low-rank adaptations, these models are engineered to capture not just the visual "style" or color palette of legendary master painters, but the true physical 3D texture: heavy impasto, direction of knife strokes, drying oil paint gloss, and raw canvas weave.

💡 The High-Rank Philosophy

Why train LoRAs at rank 1280–2048 with up to 12.48 Billion trainable parameters?

Rank 32–128: Learns general color palettes and surface style.
Rank 512: Learns artist technique and brush mechanics.
Rank 1280–2048: Learns the physical hand, pressure, and 3D gesture of the master.

At rank 2048, the adapter matrix covers nearly 50% of the base model's capacity, embedding ultra-deep spatial and physical information without quantization loss.

📂 Repository Structure & Available Models

🌊 Ivan Aivazovsky (Иван Айвазовский)

Focused on storm dynamics, light refraction through water, wave translucency, and maritime atmosphere.

Aivazovsky/r1280/ — Rank 1280 (~7.8B parameters). Fast, hyper-detailed wave physics.
Aivazovsky/r1536/ — Rank 1536 (~9.3B parameters). Deep atmospheric glow and water surface texture.

🌻 Vincent van Gogh (Винсент Ван Гог)

Focused on thick oil impasto, sculptural paint relief, raw burlap/canvas weave, and violent, expressive brushstrokes.

VanGogh/r1792/ — Rank 1792 (~10.92B parameters). Exceptional oil texture and vibrant color layering.
VanGogh/r2048/ — Rank 2048 (12.48 Billion parameters, ~24GB file). The ultimate museum-grade physical texture model.

⚙️ Technical Specifications & Hardware

All models in this repository are trained using a custom hardware-software pipeline on a single consumer GPU:

Hardware: NVIDIA RTX 4090 (24GB VRAM) + 128GB DDR5 System RAM.
Architecture: FLUX.2D (32B parameters).
Optimization: Custom Asynchronous CUDA Memory Manager, 8-bit AdamW optimizer, double-buffered offloading (0.91 ratio).
Precision: Full Precision / bf16 forcing (No Quantization).

🚀 Recommended Inference Settings (ComfyUI / Forge)

Due to the extreme size of these adapters (10GB–24GB), ensure your system has sufficient RAM to load the weights into memory.

LoRA Weight: 0.7 – 1.0 (Use 1.0 for maximum physical impasto/canvas texture).
Base Model: FLUX.2 Dev
Sampler: dpmpp_3m_sde_gpu.
scheduler: linear_quadratic.
Steps: 35–40 steps.
Guidance / CFG: 3.5.