Version: 150 Steps
Architecture: FLUX.2D
Rank: 1536
Parameters: 9,300,000,000 (9.3 billion trainable parameters)
Hardware Stack: NVIDIA RTX 4090 · Double Buffer Async CUDA Memory Manager · bf16 precision
Release Description
Introducing an extreme LoRA for FLUX.2D, trained at rank 1536 (9.3 billion parameters). At this level, the neural network reproduces the physics of classical oil painting: brushstroke density, multi-layered glazing, light reflection, and the texture of 19th-century academic landscape painting.
Important: All demonstration images in the gallery are pure raw generations, without any post-processing in Photoshop or upscaling.
Inference and Generation Parameters
To achieve the exact same results as the examples, use the following sampler settings:
Sampler: DPMPP_3M_SDE_GPU
Scheduler: Linear_Quadratic
CFG Scale: 3.5 (recommended value)
LoRA Weight:
1.0 (Default) for generation at standard resolutions (1024x1024, 1280x720, etc.).
Higher values (>1.0) when using our custom 3K workflow (details below).
Orakul Studio Custom Nodes (ComfyUI Workflow)
When importing the inference workflow from my PNG files into ComfyUI, you will see two custom nodes. The repository link for the nodes is available on GitHub: https://github.com/OrakulStudio
Orakul 3K Resolution
Purpose: An advanced aspect ratio control node that generates latent space directly at 3K resolution.
Requirements: Designed for users with high-end graphics cards (24GB+ VRAM / RTX 4090).
Note: Due to the 3K resolution, the generation density requires a higher LoRA strength. If your GPU struggles or you do not need 3K resolution, simply remove or disable this node and use the standard aspect ratio with a LoRA weight of 1.0.
Orakul SVP Engine
Purpose: An automated export module. When you click "Queue," the node simultaneously saves the result in three formats to a dedicated folder:
.EXR (for deep color grading without losing dynamic range)
.PNG (with embedded workflow metadata)
.TIFF (maximum quality for printing/stock sites)
Note: This node can be kept in your workflow; it operates independently of GPU power.
Additional Information
A distilled version of the model perfect for generation on 16GB VRAM GPUs will be released soon, along with details on the baked model version (Full Weight Merge).
Orakul Studio · Chernihiv, Ukraine 🇺🇦